From 7602f9161218afa4e24fe6eee711e92060ae2cdd Mon Sep 17 00:00:00 2001
From: dapan1121 <wpan@taosdata.com>
Date: Fri, 1 Apr 2022 19:45:45 +0800
Subject: [PATCH 01/25] feature/qnode

---
 include/common/tmsg.h                         |   7 +
 include/common/tmsgdef.h                      |   1 +
 include/libs/command/command.h                |  12 ++
 include/libs/executor/executor.h              |   2 +
 include/libs/nodes/nodes.h                    |   1 +
 include/libs/nodes/plannodes.h                |   1 +
 include/libs/nodes/querynodes.h               |   1 +
 include/libs/scheduler/scheduler.h            |   2 +-
 source/client/src/clientImpl.c                |   2 +-
 source/libs/command/inc/commandInt.h          |  13 +-
 source/libs/command/src/explain.c             | 154 +++++++-------
 source/libs/executor/src/executorMain.c       |   9 +
 source/libs/executor/src/executorimpl.c       |  15 ++
 source/libs/nodes/inc/nodesUtil.h             |   1 +
 source/libs/nodes/src/nodesUtilFuncs.c        |  30 +++
 source/libs/qworker/inc/qworkerInt.h          |   6 +-
 source/libs/qworker/src/qworker.c             | 100 +++++----
 source/libs/qworker/src/qworkerMsg.c          |   2 +-
 source/libs/scheduler/inc/schedulerInt.h      |   5 +
 source/libs/scheduler/src/scheduler.c         | 194 ++++++++++++------
 source/libs/scheduler/test/schedulerTests.cpp |   2 +-
 21 files changed, 376 insertions(+), 184 deletions(-)

diff --git a/include/common/tmsg.h b/include/common/tmsg.h
index cd16bbf862..4cfcf0e888 100644
--- a/include/common/tmsg.h
+++ b/include/common/tmsg.h
@@ -915,6 +915,12 @@ typedef struct {
   char    data[];
 } SRetrieveMetaTableRsp;
 
+typedef struct {
+  int32_t numOfItems;
+  char    data[];
+} SExplainRsp;
+
+
 typedef struct {
   char    fqdn[TSDB_FQDN_LEN];  // end point, hostname:port
   int32_t port;
@@ -1052,6 +1058,7 @@ typedef struct SSubQueryMsg {
   uint64_t taskId;
   int64_t  refId;
   int8_t   taskType;
+  int8_t   explain;
   uint32_t sqlLen;  // the query sql,
   uint32_t phyLen;
   char     msg[];
diff --git a/include/common/tmsgdef.h b/include/common/tmsgdef.h
index 36a489eb59..1d4667cda0 100644
--- a/include/common/tmsgdef.h
+++ b/include/common/tmsgdef.h
@@ -188,6 +188,7 @@ enum {
   TD_DEF_MSG_TYPE(TDMT_VND_SHOW_TABLES_FETCH, "vnode-show-tables-fetch", SVShowTablesFetchReq, SVShowTablesFetchRsp)
   TD_DEF_MSG_TYPE(TDMT_VND_QUERY_CONTINUE, "vnode-query-continue", NULL, NULL)
   TD_DEF_MSG_TYPE(TDMT_VND_QUERY_HEARTBEAT, "vnode-query-heartbeat", NULL, NULL)
+  TD_DEF_MSG_TYPE(TDMT_VND_EXPLAIN, "vnode-explain", NULL, NULL)
 
   TD_DEF_MSG_TYPE(TDMT_VND_SUBSCRIBE, "vnode-subscribe", SMVSubscribeReq, SMVSubscribeRsp)
   TD_DEF_MSG_TYPE(TDMT_VND_CONSUME, "vnode-consume", SMqCVConsumeReq, SMqCVConsumeRsp)
diff --git a/include/libs/command/command.h b/include/libs/command/command.h
index 7e58d39692..31e370ac5c 100644
--- a/include/libs/command/command.h
+++ b/include/libs/command/command.h
@@ -17,8 +17,20 @@
 #include "tmsg.h"
 #include "plannodes.h"
 
+typedef struct SExplainCtx SExplainCtx;
+
+typedef struct SExplainExecInfo {
+  uint64_t startupCost;
+  uint64_t totalCost;
+  uint64_t numOfRows;
+} SExplainExecInfo;
+
 int32_t qExecCommand(SNode* pStmt, SRetrieveTableRsp** pRsp);
 
 int32_t qExecStaticExplain(SQueryPlan *pDag, SRetrieveTableRsp **pRsp);
+int32_t qExecExplainBegin(SQueryPlan *pDag, SExplainCtx **pCtx, int32_t startTs);
+int32_t qExecExplainEnd(SExplainCtx *pCtx);
+int32_t qExplainUpdateExecInfo(SExplainCtx        *pCtx, SExplainRsp *pRspMsg, int32_t groupId, SRetrieveTableRsp **pRsp);
+void    qExplainFreeCtx(SExplainCtx *pCtx);
 
 
diff --git a/include/libs/executor/executor.h b/include/libs/executor/executor.h
index d3cd828cf5..4d289147d0 100644
--- a/include/libs/executor/executor.h
+++ b/include/libs/executor/executor.h
@@ -174,6 +174,8 @@ void** qReleaseTask(void* pMgmt, void* pQInfo, bool freeHandle);
 
 void qProcessFetchRsp(void* parent, struct SRpcMsg* pMsg, struct SEpSet* pEpSet);
 
+int32_t qGetExplainExecInfo(qTaskInfo_t tinfo, int32_t *resNum, SExplainExecInfo **pRes);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/include/libs/nodes/nodes.h b/include/libs/nodes/nodes.h
index 9849dfca39..2979040c07 100644
--- a/include/libs/nodes/nodes.h
+++ b/include/libs/nodes/nodes.h
@@ -216,6 +216,7 @@ int32_t nodesListToString(const SNodeList* pList, bool format, char** pStr, int3
 int32_t nodesStringToList(const char* pStr, SNodeList** pList);
 
 int32_t nodesNodeToSQL(SNode *pNode, char *buf, int32_t bufSize, int32_t *len);
+char *nodesGetNameFromColumnNode(SNode *pNode);
 
 #ifdef __cplusplus
 }
diff --git a/include/libs/nodes/plannodes.h b/include/libs/nodes/plannodes.h
index 37163f60dd..eda7882e2c 100644
--- a/include/libs/nodes/plannodes.h
+++ b/include/libs/nodes/plannodes.h
@@ -308,6 +308,7 @@ typedef enum EExplainMode {
 typedef struct SExplainInfo {
   EExplainMode mode;
   bool verbose;
+  double ratio;
 } SExplainInfo;
 
 typedef struct SQueryPlan {
diff --git a/include/libs/nodes/querynodes.h b/include/libs/nodes/querynodes.h
index ddcbaa0bee..04fa8f21df 100644
--- a/include/libs/nodes/querynodes.h
+++ b/include/libs/nodes/querynodes.h
@@ -315,6 +315,7 @@ bool nodesIsTimelineQuery(const SNode* pQuery);
 
 void* nodesGetValueFromNode(SValueNode *pNode);
 char* nodesGetStrValueFromNode(SValueNode *pNode);
+char *getFillModeString(EFillMode mode);
 
 #ifdef __cplusplus
 }
diff --git a/include/libs/scheduler/scheduler.h b/include/libs/scheduler/scheduler.h
index 16a6ae32cf..5ab4ead89c 100644
--- a/include/libs/scheduler/scheduler.h
+++ b/include/libs/scheduler/scheduler.h
@@ -71,7 +71,7 @@ int32_t schedulerInit(SSchedulerCfg *cfg);
  * @param nodeList  Qnode/Vnode address list, element is SQueryNodeAddr
  * @return
  */
-int32_t schedulerExecJob(void *transport, SArray *nodeList, SQueryPlan* pDag, int64_t* pJob, const char* sql, SQueryResult *pRes);
+int32_t schedulerExecJob(void *transport, SArray *nodeList, SQueryPlan *pDag, int64_t *pJob, const char *sql, int64_t startTs, SQueryResult *pRes);
 
 /**
  * Process the query job, generated according to the query physical plan.
diff --git a/source/client/src/clientImpl.c b/source/client/src/clientImpl.c
index 8cd53f18b0..1512322f1a 100644
--- a/source/client/src/clientImpl.c
+++ b/source/client/src/clientImpl.c
@@ -241,7 +241,7 @@ int32_t scheduleQuery(SRequestObj* pRequest, SQueryPlan* pDag, SArray* pNodeList
   void* pTransporter = pRequest->pTscObj->pAppInfo->pTransporter;
 
   SQueryResult res = {.code = 0, .numOfRows = 0, .msgSize = ERROR_MSG_BUF_DEFAULT_SIZE, .msg = pRequest->msgBuf};
-  int32_t      code = schedulerExecJob(pTransporter, pNodeList, pDag, &pRequest->body.queryJob, pRequest->sqlstr, &res);
+  int32_t      code = schedulerExecJob(pTransporter, pNodeList, pDag, &pRequest->body.queryJob, pRequest->sqlstr, pRequest->metric.start, &res);
   if (code != TSDB_CODE_SUCCESS) {
     if (pRequest->body.queryJob != 0) {
       schedulerFreeJob(pRequest->body.queryJob);
diff --git a/source/libs/command/inc/commandInt.h b/source/libs/command/inc/commandInt.h
index 771baca2ab..de0a7c433a 100644
--- a/source/libs/command/inc/commandInt.h
+++ b/source/libs/command/inc/commandInt.h
@@ -50,8 +50,9 @@ extern "C" {
 
 typedef struct SExplainGroup {
   int32_t   nodeNum;
+  SRWLatch  lock;
   SSubplan *plan;
-  void     *execInfo;  //TODO
+  SArray   *execInfo;
 } SExplainGroup;
 
 typedef struct SExplainResNode {
@@ -67,10 +68,18 @@ typedef struct SQueryExplainRowInfo {
 } SQueryExplainRowInfo;
 
 typedef struct SExplainCtx {
-  int32_t   totalSize;
+  double    ratio;
   bool      verbose;
+
+  int32_t   rootGroupId;
+  int32_t   dataSize;
+  bool      execDone;
+  int64_t   reqStartTs;
+  int64_t   jobStartTs;
+  int64_t   jobDoneTs;
   char     *tbuf;
   SArray   *rows;
+  int32_t   groupDoneNum;
   SHashObj *groupHash;
 } SExplainCtx;
 
diff --git a/source/libs/command/src/explain.c b/source/libs/command/src/explain.c
index 847a863e76..4bf354dee5 100644
--- a/source/libs/command/src/explain.c
+++ b/source/libs/command/src/explain.c
@@ -17,32 +17,31 @@
 #include "plannodes.h"
 #include "commandInt.h"
 
-int32_t qGenerateExplainResNode(SPhysiNode *pNode, void *pExecInfo, SExplainResNode **pRes);
-int32_t qAppendTaskExplainResRows(void *pCtx, int32_t groupId, int32_t level);
+int32_t qExplainGenerateResNode(SPhysiNode *pNode, void *pExecInfo, SExplainResNode **pRes);
+int32_t qExplainAppendGroupResRows(void *pCtx, int32_t groupId, int32_t level);
 
 
-void qFreeExplainResTree(SExplainResNode *res) {
-  if (NULL == res) {
+void qExplainFreeResNode(SExplainResNode *resNode) {
+  if (NULL == resNode) {
     return;
   }
 
-  taosMemoryFreeClear(res->pExecInfo);
+  taosMemoryFreeClear(resNode->pExecInfo);
 
   SNode* node = NULL;
-  FOREACH(node, res->pChildren) {
-    qFreeExplainResTree((SExplainResNode *)node);
+  FOREACH(node, resNode->pChildren) {
+    qExplainFreeResNode((SExplainResNode *)node);
   }  
-  nodesClearList(res->pChildren);
+  nodesClearList(resNode->pChildren);
   
-  taosMemoryFreeClear(res);
+  taosMemoryFreeClear(resNode);
 }
 
-void qFreeExplainCtx(void *ctx) {
-  if (NULL == ctx) {
+void qExplainFreeCtx(SExplainCtx *pCtx) {
+  if (NULL == pCtx) {
     return;
   }
 
-  SExplainCtx *pCtx = (SExplainCtx *)ctx;
   int32_t rowSize = taosArrayGetSize(pCtx->rows);
   for (int32_t i = 0; i < rowSize; ++i) {
     SQueryExplainRowInfo *row = taosArrayGet(pCtx->rows, i);
@@ -54,7 +53,7 @@ void qFreeExplainCtx(void *ctx) {
   taosMemoryFree(pCtx);
 }
 
-int32_t qInitExplainCtx(void **pCtx, SHashObj *groupHash, bool verbose) {
+int32_t qExplainInitCtx(SExplainCtx **pCtx, SHashObj *groupHash, bool verbose, double ratio) {
   int32_t code = 0;
   SExplainCtx *ctx = taosMemoryCalloc(1, sizeof(SExplainCtx));
   if (NULL == ctx) {
@@ -75,6 +74,7 @@ int32_t qInitExplainCtx(void **pCtx, SHashObj *groupHash, bool verbose) {
   }
 
   ctx->verbose = verbose;
+  ctx->ratio = ratio;
   ctx->tbuf = tbuf;
   ctx->rows = rows;
   ctx->groupHash = groupHash;
@@ -92,35 +92,7 @@ _return:
   QRY_RET(code);
 }
 
-
-char *qFillModeString(EFillMode mode) {
-  switch (mode) {
-    case FILL_MODE_NONE:
-      return "none";
-    case FILL_MODE_VALUE:
-      return "value";
-    case FILL_MODE_PREV:
-      return "prev";
-    case FILL_MODE_NULL:
-      return "null";
-    case FILL_MODE_LINEAR:
-      return "linear";
-    case FILL_MODE_NEXT:
-      return "next";
-    default:
-      return "unknown";
-  }
-}
-
-char *qGetNameFromColumnNode(SNode *pNode) {
-  if (NULL == pNode || QUERY_NODE_COLUMN != pNode->type) {
-    return "NULL";
-  }
-  
-  return ((SColumnNode *)pNode)->colName;
-}
-
-int32_t qGenerateExplainResChildren(SPhysiNode *pNode, void *pExecInfo, SNodeList **pChildren) {
+int32_t qExplainGenerateResChildren(SPhysiNode *pNode, void *pExecInfo, SNodeList **pChildren) {
   int32_t tlen = 0;
   SNodeList *pPhysiChildren = NULL;
   
@@ -192,38 +164,38 @@ int32_t qGenerateExplainResChildren(SPhysiNode *pNode, void *pExecInfo, SNodeLis
   SNode* node = NULL;
   SExplainResNode *pResNode = NULL;
   FOREACH(node, pPhysiChildren) {
-    QRY_ERR_RET(qGenerateExplainResNode((SPhysiNode *)node, pExecInfo, &pResNode));
+    QRY_ERR_RET(qExplainGenerateResNode((SPhysiNode *)node, pExecInfo, &pResNode));
     QRY_ERR_RET(nodesListAppend(*pChildren, pResNode));
   }
 
   return TSDB_CODE_SUCCESS;
 }
 
-int32_t qGenerateExplainResNode(SPhysiNode *pNode, void *pExecInfo, SExplainResNode **pRes) {
+int32_t qExplainGenerateResNode(SPhysiNode *pNode, void *pExecInfo, SExplainResNode **pResNode) {
   if (NULL == pNode) {
-    *pRes = NULL;
+    *pResNode = NULL;
     qError("physical node is NULL");
     return TSDB_CODE_QRY_APP_ERROR;
   }
 
-  SExplainResNode *res = taosMemoryCalloc(1, sizeof(SExplainResNode));
-  if (NULL == res) {
+  SExplainResNode *resNode = taosMemoryCalloc(1, sizeof(SExplainResNode));
+  if (NULL == resNode) {
     qError("calloc SPhysiNodeExplainRes failed");
     return TSDB_CODE_QRY_OUT_OF_MEMORY;
   }
 
   int32_t code = 0;
-  res->pNode = pNode;
-  res->pExecInfo = pExecInfo;
-  QRY_ERR_JRET(qGenerateExplainResChildren(pNode, pExecInfo, &res->pChildren));
+  resNode->pNode = pNode;
+  resNode->pExecInfo = pExecInfo;
+  QRY_ERR_JRET(qExplainGenerateResChildren(pNode, pExecInfo, &resNode->pChildren));
   
-  *pRes = res;
+  *pResNode = resNode;
 
   return TSDB_CODE_SUCCESS;
 
 _return:
 
-  qFreeExplainResTree(res);
+  qExplainFreeResNode(resNode);
   
   QRY_RET(code);
 }
@@ -249,7 +221,7 @@ int32_t qExplainResAppendRow(SExplainCtx *ctx, char *tbuf, int32_t len, int32_t
   memcpy(row.buf, tbuf, len);
   row.level = level;
   row.len = len;
-  ctx->totalSize += len;
+  ctx->dataSize += len;
 
   if (NULL == taosArrayPush(ctx->rows, &row)) {
     qError("taosArrayPush row to explain res rows failed");
@@ -444,7 +416,7 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
         }
       }
 
-      QRY_ERR_RET(qAppendTaskExplainResRows(ctx, pExchNode->srcGroupId, level + 1));
+      QRY_ERR_RET(qExplainAppendGroupResRows(ctx, pExchNode->srcGroupId, level + 1));
       break;
     }
     case QUERY_NODE_PHYSICAL_PLAN_SORT:{
@@ -468,7 +440,7 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
     }
     case QUERY_NODE_PHYSICAL_PLAN_INTERVAL:{
       SIntervalPhysiNode *pIntNode = (SIntervalPhysiNode *)pNode;
-      EXPLAIN_ROW_NEW(level, EXPLAIN_INTERVAL_FORMAT, qGetNameFromColumnNode(pIntNode->pTspk), pIntNode->window.pFuncs->length, 
+      EXPLAIN_ROW_NEW(level, EXPLAIN_INTERVAL_FORMAT, nodesGetNameFromColumnNode(pIntNode->pTspk), pIntNode->window.pFuncs->length, 
         INVERAL_TIME_FROM_PRECISION_TO_UNIT(pIntNode->interval, pIntNode->intervalUnit, pIntNode->precision), pIntNode->intervalUnit, 
         pIntNode->offset, getPrecisionUnit(pIntNode->precision), 
         INVERAL_TIME_FROM_PRECISION_TO_UNIT(pIntNode->sliding, pIntNode->slidingUnit, pIntNode->precision), pIntNode->slidingUnit, 
@@ -481,7 +453,7 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
 
       if (verbose) {
         if (pIntNode->pFill) {
-          EXPLAIN_ROW_NEW(level + 1, EXPLAIN_FILL_FORMAT, qFillModeString(pIntNode->pFill->mode));      
+          EXPLAIN_ROW_NEW(level + 1, EXPLAIN_FILL_FORMAT, getFillModeString(pIntNode->pFill->mode));      
           EXPLAIN_ROW_END();
           QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
         }
@@ -540,7 +512,7 @@ int32_t qExplainResNodeToRows(SExplainResNode *pResNode, SExplainCtx *ctx, int32
   return TSDB_CODE_SUCCESS;
 }
 
-int32_t qAppendTaskExplainResRows(void *pCtx, int32_t groupId, int32_t level) {
+int32_t qExplainAppendGroupResRows(void *pCtx, int32_t groupId, int32_t level) {
   SExplainResNode *node = NULL;
   int32_t code = 0;
   SExplainCtx *ctx = (SExplainCtx *)pCtx;
@@ -551,19 +523,19 @@ int32_t qAppendTaskExplainResRows(void *pCtx, int32_t groupId, int32_t level) {
     QRY_ERR_RET(TSDB_CODE_QRY_APP_ERROR);
   }
   
-  QRY_ERR_RET(qGenerateExplainResNode(group->plan->pNode, group->execInfo, &node));
+  QRY_ERR_RET(qExplainGenerateResNode(group->plan->pNode, group->execInfo, &node));
 
   QRY_ERR_JRET(qExplainResNodeToRows(node, ctx, level));
 
 _return:
 
-  qFreeExplainResTree(node);
+  qExplainFreeResNode(node);
   
   QRY_RET(code);
 }
 
 
-int32_t qGetExplainRspFromCtx(void *ctx, SRetrieveTableRsp **pRsp) {
+int32_t qExplainGetRspFromCtx(void *ctx, SRetrieveTableRsp **pRsp) {
   SExplainCtx *pCtx = (SExplainCtx *)ctx;
   int32_t rowNum = taosArrayGetSize(pCtx->rows);
   if (rowNum <= 0) {
@@ -572,7 +544,7 @@ int32_t qGetExplainRspFromCtx(void *ctx, SRetrieveTableRsp **pRsp) {
   }
   
   int32_t colNum = 1;
-  int32_t rspSize = sizeof(SRetrieveTableRsp) + sizeof(int32_t) * colNum + sizeof(int32_t) * rowNum + pCtx->totalSize;
+  int32_t rspSize = sizeof(SRetrieveTableRsp) + sizeof(int32_t) * colNum + sizeof(int32_t) * rowNum + pCtx->dataSize;
   SRetrieveTableRsp *rsp = (SRetrieveTableRsp *)taosMemoryCalloc(1, rspSize);
   if (NULL == rsp) {
     qError("malloc SRetrieveTableRsp failed, size:%d", rspSize);
@@ -582,7 +554,7 @@ int32_t qGetExplainRspFromCtx(void *ctx, SRetrieveTableRsp **pRsp) {
   rsp->completed = 1;
   rsp->numOfRows = htonl(rowNum);
 
-  *(int32_t *)rsp->data = htonl(pCtx->totalSize);
+  *(int32_t *)rsp->data = htonl(pCtx->dataSize);
 
   int32_t *offset = (int32_t *)((char *)rsp->data + sizeof(int32_t));
   char *data = (char *)(offset + rowNum);
@@ -604,13 +576,13 @@ int32_t qGetExplainRspFromCtx(void *ctx, SRetrieveTableRsp **pRsp) {
   return TSDB_CODE_SUCCESS;
 }
 
-int32_t qExecStaticExplain(SQueryPlan *pDag, SRetrieveTableRsp **pRsp) {
+
+int32_t qExplainPrepareCtx(SQueryPlan *pDag, SExplainCtx **pCtx) {
   int32_t code = 0;
   SNodeListNode *plans = NULL;
   int32_t        taskNum = 0;
   SExplainGroup *pGroup = NULL;
-  void *pCtx = NULL;
-  int32_t rootGroupId = 0;
+  SExplainCtx *ctx = NULL;
 
   if (pDag->numOfSubplans <= 0) {
     qError("invalid subplan num:%d", pDag->numOfSubplans);
@@ -629,7 +601,7 @@ int32_t qExecStaticExplain(SQueryPlan *pDag, SRetrieveTableRsp **pRsp) {
     QRY_ERR_RET(TSDB_CODE_QRY_OUT_OF_MEMORY);
   }
 
-  QRY_ERR_JRET(qInitExplainCtx(&pCtx, groupHash, pDag->explainInfo.verbose));
+  QRY_ERR_JRET(qExplainInitCtx(&ctx, groupHash, pDag->explainInfo.verbose, pDag->explainInfo.ratio));
 
   for (int32_t i = 0; i < levelNum; ++i) {
     plans = (SNodeListNode *)nodesListGetNode(pDag->pSubplans, i);
@@ -666,22 +638,62 @@ int32_t qExecStaticExplain(SQueryPlan *pDag, SRetrieveTableRsp **pRsp) {
         QRY_ERR_JRET(TSDB_CODE_QRY_INVALID_INPUT);
       }
 
-      rootGroupId = plan->id.groupId;
+      ctx->rootGroupId = plan->id.groupId;
     }
 
     qDebug("level %d group handled, taskNum:%d", i, taskNum);
   }
 
-  QRY_ERR_JRET(qAppendTaskExplainResRows(pCtx, rootGroupId, 0));
-  
-  QRY_ERR_JRET(qGetExplainRspFromCtx(pCtx, pRsp));
+  *pCtx = ctx;
+
+  return TSDB_CODE_SUCCESS;
   
 _return:
 
-  qFreeExplainCtx(pCtx);
+  qExplainFreeCtx(ctx);
 
   QRY_RET(code);
 }
 
 
+int32_t qExplainUpdateExecInfo(SExplainCtx        *pCtx, SExplainRsp *pRspMsg, int32_t groupId, SRetrieveTableRsp **pRsp) {
+
+}
+
+
+int32_t qExecStaticExplain(SQueryPlan *pDag, SRetrieveTableRsp **pRsp) {
+  int32_t code = 0;
+  SExplainCtx *pCtx = NULL;
+
+  QRY_ERR_RET(qExplainPrepareCtx(pDag, &pCtx));
+
+  QRY_ERR_JRET(qExplainAppendGroupResRows(pCtx, pCtx->rootGroupId, 0));
+  
+  QRY_ERR_JRET(qExplainGetRspFromCtx(pCtx, pRsp));
+  
+_return:
+
+  qExplainFreeCtx(pCtx);
+
+  QRY_RET(code);
+}
+
+int32_t qExecExplainBegin(SQueryPlan *pDag, SExplainCtx **pCtx, int32_t startTs) {
+  QRY_ERR_RET(qExplainPrepareCtx(pDag, pCtx));
+  
+  (*pCtx)->reqStartTs = startTs;
+  (*pCtx)->jobStartTs = taosGetTimestampMs();
+
+  return TSDB_CODE_SUCCESS;
+}
+
+int32_t qExecExplainEnd(SExplainCtx *pCtx) {
+  pCtx->jobDoneTs = taosGetTimestampMs();
+  
+  atomic_store_8((int8_t *)&pCtx->execDone, true);
+
+  return TSDB_CODE_SUCCESS;
+}
+
+
 
diff --git a/source/libs/executor/src/executorMain.c b/source/libs/executor/src/executorMain.c
index 6d78fa835a..d7c99d54df 100644
--- a/source/libs/executor/src/executorMain.c
+++ b/source/libs/executor/src/executorMain.c
@@ -229,3 +229,12 @@ void qDestroyTask(qTaskInfo_t qTaskHandle) {
   queryCostStatis(pTaskInfo);   // print the query cost summary
   doDestroyTask(pTaskInfo);
 }
+
+int32_t qGetExplainExecInfo(qTaskInfo_t tinfo, int32_t *resNum, SExplainExecInfo **pRes) {
+  SExecTaskInfo *pTaskInfo = (SExecTaskInfo *)tinfo;
+  int32_t capacity = 0;
+
+  return getOperatorExplainExecInfo(pTaskInfo->pRoot, pRes, &capacity, resNum);  
+}
+
+
diff --git a/source/libs/executor/src/executorimpl.c b/source/libs/executor/src/executorimpl.c
index 169e47eb33..e423690740 100644
--- a/source/libs/executor/src/executorimpl.c
+++ b/source/libs/executor/src/executorimpl.c
@@ -9516,3 +9516,18 @@ void releaseQueryBuf(size_t numOfTables) {
   // restore value is not enough buffer available
   atomic_add_fetch_64(&tsQueryBufferSizeBytes, t);
 }
+
+int32_t getOperatorExplainExecInfo(SOperatorInfo *operator, SExplainExecInfo **pRes, int32_t *capacity, int32_t *resNum) {
+  if (*resNum >= *capacity) {
+    *capacity += 10;
+    
+    *pRes = taosMemoryRealloc(*pRes, (*capacity) * sizeof(SExplainExecInfo));
+    if (NULL == *pRes) {
+      qError("malloc %d failed", capacity * sizeof(SExplainExecInfo));
+      return TSDB_CODE_QRY_OUT_OF_MEMORY;
+    }
+  }
+
+}
+
+
diff --git a/source/libs/nodes/inc/nodesUtil.h b/source/libs/nodes/inc/nodesUtil.h
index 976044c16f..c6233ba980 100644
--- a/source/libs/nodes/inc/nodesUtil.h
+++ b/source/libs/nodes/inc/nodesUtil.h
@@ -31,6 +31,7 @@ extern "C" {
 #define NODES_RET(c) do { int32_t _code = c; if (_code != TSDB_CODE_SUCCESS) { terrno = _code; } return _code; } while (0)
 #define NODES_ERR_JRET(c) do { code = c; if (code != TSDB_CODE_SUCCESS) { terrno = code; goto _return; } } while (0)
 
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/source/libs/nodes/src/nodesUtilFuncs.c b/source/libs/nodes/src/nodesUtilFuncs.c
index 732d9acfbe..812fa333b6 100644
--- a/source/libs/nodes/src/nodesUtilFuncs.c
+++ b/source/libs/nodes/src/nodesUtilFuncs.c
@@ -1023,3 +1023,33 @@ int32_t nodesCollectFuncs(SSelectStmt* pSelect, FFuncClassifier classifier, SNod
   
   return TSDB_CODE_SUCCESS;
 }
+
+
+char *getFillModeString(EFillMode mode) {
+  switch (mode) {
+    case FILL_MODE_NONE:
+      return "none";
+    case FILL_MODE_VALUE:
+      return "value";
+    case FILL_MODE_PREV:
+      return "prev";
+    case FILL_MODE_NULL:
+      return "null";
+    case FILL_MODE_LINEAR:
+      return "linear";
+    case FILL_MODE_NEXT:
+      return "next";
+    default:
+      return "unknown";
+  }
+}
+
+char *nodesGetNameFromColumnNode(SNode *pNode) {
+  if (NULL == pNode || QUERY_NODE_COLUMN != pNode->type) {
+    return "NULL";
+  }
+  
+  return ((SColumnNode *)pNode)->colName;
+}
+
+
diff --git a/source/libs/qworker/inc/qworkerInt.h b/source/libs/qworker/inc/qworkerInt.h
index 573eaed2e6..cfd4a3ec7b 100644
--- a/source/libs/qworker/inc/qworkerInt.h
+++ b/source/libs/qworker/inc/qworkerInt.h
@@ -107,15 +107,17 @@ typedef struct SQWTaskCtx {
   SRWLatch        lock;
   int8_t          phase;
   int8_t          taskType;
+  int8_t          explain;
   
-  bool            emptyRes;
   bool            queryFetched;
   bool            queryEnd;
   bool            queryContinue;
   bool            queryInQueue;
   int32_t         rspCode; 
 
-  SQWConnInfo     connInfo;
+  SQWConnInfo     ctrlConnInfo;
+  SQWConnInfo     dataConnInfo;
+
   int8_t          events[QW_EVENT_MAX];
   
   qTaskInfo_t     taskHandle;
diff --git a/source/libs/qworker/src/qworker.c b/source/libs/qworker/src/qworker.c
index 42d10a2361..726d0ebf79 100644
--- a/source/libs/qworker/src/qworker.c
+++ b/source/libs/qworker/src/qworker.c
@@ -432,8 +432,10 @@ int32_t qwKillTaskHandle(QW_FPARAMS_DEF, SQWTaskCtx *ctx) {
 
 
 void qwFreeTask(QW_FPARAMS_DEF, SQWTaskCtx *ctx) {
-  tmsgReleaseHandle(ctx->connInfo.handle, TAOS_CONN_SERVER);
-  ctx->connInfo.handle = NULL;
+  tmsgReleaseHandle(ctx->ctrlConnInfo.handle, TAOS_CONN_SERVER);
+  ctx->ctrlConnInfo.handle = NULL;
+
+  // NO need to release dataConnInfo
 
   qwFreeTaskHandle(QW_FPARAMS(), &ctx->taskHandle);
   
@@ -537,6 +539,26 @@ int32_t qwDropTask(QW_FPARAMS_DEF) {
   return TSDB_CODE_SUCCESS;
 }
 
+
+int32_t qwHandleTaskComplete(QW_FPARAMS_DEF, SQWTaskCtx *ctx) {
+  qTaskInfo_t *taskHandle = &ctx->taskHandle; 
+
+  if (TASK_TYPE_TEMP == ctx->taskType) {
+    if (ctx->explain) {
+      SExplainExecInfo *execInfo = NULL;
+      int32_t resNum = 0;
+      QW_ERR_RET(qGetExplainExecInfo(ctx->taskHandle, &resNum, &execInfo));
+
+      QW_ERR_RET(qwBuildAndSendExplainRsp(&ctx->ctrlConnInfo, &rsp));
+    }
+    
+    qwFreeTaskHandle(QW_FPARAMS(), taskHandle);
+  }
+
+  return TSDB_CODE_SUCCESS;
+}
+
+
 int32_t qwExecTask(QW_FPARAMS_DEF, SQWTaskCtx *ctx, bool *queryEnd) {
   int32_t code = 0;
   bool  qcontinue = true;
@@ -562,10 +584,8 @@ int32_t qwExecTask(QW_FPARAMS_DEF, SQWTaskCtx *ctx, bool *queryEnd) {
       QW_TASK_DLOG("qExecTask end with empty res, useconds:%"PRIu64, useconds);
 
       dsEndPut(sinkHandle, useconds);
-      
-      if (TASK_TYPE_TEMP == ctx->taskType) {
-        qwFreeTaskHandle(QW_FPARAMS(), taskHandle);
-      }
+
+      QW_ERR_RET(qwHandleTaskComplete(QW_FPARAMS(), ctx));
 
       if (queryEnd) {
         *queryEnd = true;
@@ -658,19 +678,6 @@ int32_t qwGetResFromSink(QW_FPARAMS_DEF, SQWTaskCtx *ctx, int32_t *dataLen, void
   bool queryEnd = false;
   int32_t code = 0;
 
-  if (ctx->emptyRes) {
-    QW_TASK_DLOG_E("query end with empty result");
-    
-    qwUpdateTaskStatus(QW_FPARAMS(), JOB_TASK_STATUS_SUCCEED);
-    QW_ERR_RET(qwMallocFetchRsp(len, &rsp));      
-    
-    *rspMsg = rsp;
-    *dataLen = 0;
-    pOutput->queryEnd = true;
-    
-    return TSDB_CODE_SUCCESS;
-  }
-
   dsGetDataLength(ctx->sinkHandle, &len, &queryEnd);
 
   if (len < 0) {
@@ -760,12 +767,12 @@ int32_t qwHandlePrePhaseEvents(QW_FPARAMS_DEF, int8_t phase, SQWPhaseInput *inpu
       }
 
       if (QW_IS_EVENT_RECEIVED(ctx, QW_EVENT_DROP)) {
-        dropConnection = &ctx->connInfo;
+        dropConnection = &ctx->ctrlConnInfo;
         QW_ERR_JRET(qwDropTask(QW_FPARAMS()));
         dropConnection = NULL;
 
-        qwBuildAndSendDropRsp(&ctx->connInfo, code);
-        QW_TASK_DLOG("drop rsp send, handle:%p, code:%x - %s", ctx->connInfo.handle, code, tstrerror(code));
+        qwBuildAndSendDropRsp(&ctx->ctrlConnInfo, code);
+        QW_TASK_DLOG("drop rsp send, handle:%p, code:%x - %s", ctx->ctrlConnInfo.handle, code, tstrerror(code));
 
         QW_ERR_JRET(TSDB_CODE_QRY_TASK_DROPPED);
         break;
@@ -798,12 +805,12 @@ int32_t qwHandlePrePhaseEvents(QW_FPARAMS_DEF, int8_t phase, SQWPhaseInput *inpu
       }
 
       if (QW_IS_EVENT_RECEIVED(ctx, QW_EVENT_DROP)) {
-        dropConnection = &ctx->connInfo;
+        dropConnection = &ctx->ctrlConnInfo;
         QW_ERR_JRET(qwDropTask(QW_FPARAMS()));
         dropConnection = NULL;
 
-        qwBuildAndSendDropRsp(&ctx->connInfo, code);
-        QW_TASK_DLOG("drop rsp send, handle:%p, code:%x - %s", ctx->connInfo.handle, code, tstrerror(code));
+        qwBuildAndSendDropRsp(&ctx->ctrlConnInfo, code);
+        QW_TASK_DLOG("drop rsp send, handle:%p, code:%x - %s", ctx->ctrlConnInfo.handle, code, tstrerror(code));
 
         QW_ERR_JRET(TSDB_CODE_QRY_TASK_DROPPED);
       }
@@ -863,17 +870,13 @@ int32_t qwHandlePostPhaseEvents(QW_FPARAMS_DEF, int8_t phase, SQWPhaseInput *inp
   }
 
   if (QW_PHASE_POST_QUERY == phase) {
-    if (NULL == ctx->taskHandle && NULL == ctx->sinkHandle) {
-      ctx->emptyRes = true;
-    }
-
 #if 0    
     if (QW_IS_EVENT_RECEIVED(ctx, QW_EVENT_READY)) {
       readyConnection = &ctx->connInfo;
       QW_SET_EVENT_PROCESSED(ctx, QW_EVENT_READY);
     }
 #else
-    connInfo.handle = ctx->connInfo.handle;
+    connInfo.handle = ctx->ctrlConnInfo.handle;
     readyConnection = &connInfo;
     
     QW_SET_EVENT_PROCESSED(ctx, QW_EVENT_READY);
@@ -886,8 +889,8 @@ int32_t qwHandlePostPhaseEvents(QW_FPARAMS_DEF, int8_t phase, SQWPhaseInput *inp
       QW_ERR_JRET(TSDB_CODE_QRY_APP_ERROR);
     }
 
-    qwBuildAndSendDropRsp(&ctx->connInfo, code);
-    QW_TASK_DLOG("drop rsp send, handle:%p, code:%x - %s", ctx->connInfo.handle, code, tstrerror(code));
+    qwBuildAndSendDropRsp(&ctx->ctrlConnInfo, code);
+    QW_TASK_DLOG("drop rsp send, handle:%p, code:%x - %s", ctx->ctrlConnInfo.handle, code, tstrerror(code));
 
     QW_ERR_JRET(qwDropTask(QW_FPARAMS()));
     QW_ERR_JRET(TSDB_CODE_QRY_TASK_DROPPED);
@@ -931,7 +934,7 @@ _return:
   QW_RET(code);
 }
 
-int32_t qwProcessQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg, int8_t taskType) {
+int32_t qwProcessQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg, int8_t taskType, int8_t explain) {
   int32_t code = 0;
   bool queryRsped = false;
   struct SSubplan *plan = NULL;
@@ -947,9 +950,10 @@ int32_t qwProcessQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg, int8_t taskType) {
   QW_ERR_JRET(qwGetTaskCtx(QW_FPARAMS(), &ctx));
   
   atomic_store_8(&ctx->taskType, taskType);
+  atomic_store_8(&ctx->explain, explain);
 
-  atomic_store_ptr(&ctx->connInfo.handle, qwMsg->connInfo.handle);
-  atomic_store_ptr(&ctx->connInfo.ahandle, qwMsg->connInfo.ahandle);
+  atomic_store_ptr(&ctx->ctrlConnInfo.handle, qwMsg->connInfo.handle);
+  atomic_store_ptr(&ctx->ctrlConnInfo.ahandle, qwMsg->connInfo.ahandle);
 
   QW_TASK_DLOGL("subplan json string, len:%d, %s", qwMsg->msgLen, qwMsg->msg);
 
@@ -1011,8 +1015,8 @@ int32_t qwProcessReady(QW_FPARAMS_DEF, SQWMsg *qwMsg) {
   }
   
   if (ctx->phase == QW_PHASE_PRE_QUERY) {
-    ctx->connInfo.handle == qwMsg->connInfo.handle;
-    ctx->connInfo.ahandle = qwMsg->connInfo.ahandle;
+    ctx->ctrlConnInfo.handle == qwMsg->connInfo.handle;
+    ctx->ctrlConnInfo.ahandle = qwMsg->connInfo.ahandle;
     QW_SET_EVENT_RECEIVED(ctx, QW_EVENT_READY);
     needRsp = false;
     QW_TASK_DLOG_E("ready msg will not rsp now");
@@ -1089,10 +1093,13 @@ int32_t qwProcessCQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg) {
       
       if (rsp) {
         bool qComplete = (DS_BUF_EMPTY == sOutput.bufStatus && sOutput.queryEnd);
+        
         qwBuildFetchRsp(rsp, &sOutput, dataLen, qComplete);
-        atomic_store_8((int8_t*)&ctx->queryEnd, qComplete);
+        if (qComplete) {
+          atomic_store_8((int8_t*)&ctx->queryEnd, true);
+        }
 
-        qwMsg->connInfo = ctx->connInfo;
+        qwMsg->connInfo = ctx->dataConnInfo;
         QW_SET_EVENT_PROCESSED(ctx, QW_EVENT_FETCH);            
         
         qwBuildAndSendFetchRsp(&qwMsg->connInfo, rsp, dataLen, code);                
@@ -1113,7 +1120,7 @@ _return:
       qwFreeFetchRsp(rsp);
       rsp = NULL;
       
-      qwMsg->connInfo = ctx->connInfo;
+      qwMsg->connInfo = ctx->dataConnInfo;
       qwBuildAndSendFetchRsp(&qwMsg->connInfo, rsp, 0, code);
       QW_TASK_DLOG("fetch rsp send, handle:%p, code:%x - %s, dataLen:%d", qwMsg->connInfo.handle, code, tstrerror(code), 0);
     }
@@ -1151,14 +1158,17 @@ int32_t qwProcessFetch(QW_FPARAMS_DEF, SQWMsg *qwMsg) {
   QW_ERR_JRET(qwGetResFromSink(QW_FPARAMS(), ctx, &dataLen, &rsp, &sOutput));
 
   if (NULL == rsp) {
-    atomic_store_ptr(&ctx->connInfo.handle, qwMsg->connInfo.handle);
-    atomic_store_ptr(&ctx->connInfo.ahandle, qwMsg->connInfo.ahandle);
+    atomic_store_ptr(&ctx->dataConnInfo.handle, qwMsg->connInfo.handle);
+    atomic_store_ptr(&ctx->dataConnInfo.ahandle, qwMsg->connInfo.ahandle);
     
     QW_SET_EVENT_RECEIVED(ctx, QW_EVENT_FETCH);
   } else {
     bool qComplete = (DS_BUF_EMPTY == sOutput.bufStatus && sOutput.queryEnd);
+    
     qwBuildFetchRsp(rsp, &sOutput, dataLen, qComplete);
-    atomic_store_8((int8_t*)&ctx->queryEnd, qComplete);
+    if (qComplete) {
+      atomic_store_8((int8_t*)&ctx->queryEnd, true);
+    }
   }
 
   if ((!sOutput.queryEnd) && (DS_BUF_LOW == sOutput.bufStatus || DS_BUF_EMPTY == sOutput.bufStatus)) {    
@@ -1236,8 +1246,8 @@ int32_t qwProcessDrop(QW_FPARAMS_DEF, SQWMsg *qwMsg) {
   }
 
   if (!rsped) {
-    ctx->connInfo.handle = qwMsg->connInfo.handle;
-    ctx->connInfo.ahandle = qwMsg->connInfo.ahandle;
+    ctx->ctrlConnInfo.handle = qwMsg->connInfo.handle;
+    ctx->ctrlConnInfo.ahandle = qwMsg->connInfo.ahandle;
     
     QW_SET_EVENT_RECEIVED(ctx, QW_EVENT_DROP);
   }
diff --git a/source/libs/qworker/src/qworkerMsg.c b/source/libs/qworker/src/qworkerMsg.c
index 80ae013653..c9779f9252 100644
--- a/source/libs/qworker/src/qworkerMsg.c
+++ b/source/libs/qworker/src/qworkerMsg.c
@@ -327,7 +327,7 @@ int32_t qWorkerProcessQueryMsg(void *node, void *qWorkerMgmt, SRpcMsg *pMsg) {
   QW_SCH_TASK_DLOG("processQuery start, node:%p, handle:%p, sql:%s", node, pMsg->handle, sql);
   taosMemoryFreeClear(sql);
 
-  QW_ERR_RET(qwProcessQuery(QW_FPARAMS(), &qwMsg, msg->taskType));
+  QW_ERR_RET(qwProcessQuery(QW_FPARAMS(), &qwMsg, msg->taskType, msg->explain));
 
   QW_SCH_TASK_DLOG("processQuery end, node:%p", node);
 
diff --git a/source/libs/scheduler/inc/schedulerInt.h b/source/libs/scheduler/inc/schedulerInt.h
index 22bd039219..62a96b6438 100644
--- a/source/libs/scheduler/inc/schedulerInt.h
+++ b/source/libs/scheduler/inc/schedulerInt.h
@@ -26,6 +26,7 @@ extern "C" {
 #include "scheduler.h"
 #include "thash.h"
 #include "trpc.h"
+#include "command.h"
 
 #define SCHEDULE_DEFAULT_MAX_JOB_NUM 1000
 #define SCHEDULE_DEFAULT_MAX_TASK_NUM 1000
@@ -165,6 +166,7 @@ typedef struct SSchJob {
   SHashObj        *succTasks; // succeed tasks, key:taskid, value:SQueryTask*
   SHashObj        *failTasks; // failed tasks, key:taskid, value:SQueryTask*
 
+  SExplainCtx     *explainCtx;
   int8_t           status;  
   SQueryNodeAddr   resNode;
   tsem_t           rspSem;
@@ -211,6 +213,7 @@ extern SSchedulerMgmt schMgmt;
 #define SCH_JOB_NEED_FETCH(_job) SCH_IS_QUERY_JOB(_job)
 #define SCH_IS_WAIT_ALL_JOB(_job) (!SCH_IS_QUERY_JOB(_job))
 #define SCH_IS_NEED_DROP_JOB(_job) (SCH_IS_QUERY_JOB(_job))
+#define SCH_IS_EXPLAIN_JOB(_job) (EXPLAIN_MODE_ANALYZE == (_job)->attr.explainMode)
 
 #define SCH_IS_LEVEL_UNFINISHED(_level) ((_level)->taskLaunchedNum < (_level)->taskNum)
 #define SCH_GET_CUR_EP(_addr) (&(_addr)->epSet.eps[(_addr)->epSet.inUse])
@@ -251,6 +254,8 @@ int32_t schFetchFromRemote(SSchJob *pJob);
 int32_t schProcessOnTaskFailure(SSchJob *pJob, SSchTask *pTask, int32_t errCode);
 int32_t schBuildAndSendHbMsg(SQueryNodeEpId *nodeEpId);
 int32_t schCloneSMsgSendInfo(void *src, void **dst);
+int32_t schValidateAndBuildJob(SQueryPlan *pDag, SSchJob *pJob);
+void schFreeJobImpl(void *job);
 
 
 #ifdef __cplusplus
diff --git a/source/libs/scheduler/src/scheduler.c b/source/libs/scheduler/src/scheduler.c
index 9c5019aedc..658d924f38 100644
--- a/source/libs/scheduler/src/scheduler.c
+++ b/source/libs/scheduler/src/scheduler.c
@@ -67,6 +67,81 @@ int32_t schInitTask(SSchJob *pJob, SSchTask *pTask, SSubplan *pPlan, SSchLevel *
   return TSDB_CODE_SUCCESS;
 }
 
+int32_t schInitJob(SSchJob **pSchJob, SQueryPlan *pDag, void *transport, SArray *pNodeList, const char *sql,
+                              int64_t startTs, bool syncSchedule) {
+  int32_t code = 0;
+  SSchJob *pJob = taosMemoryCalloc(1, sizeof(SSchJob));
+  if (NULL == pJob) {
+    qError("QID:%" PRIx64 " calloc %d failed", pDag->queryId, (int32_t)sizeof(SSchJob));
+    SCH_ERR_RET(TSDB_CODE_QRY_OUT_OF_MEMORY);
+  }
+
+  pJob->attr.explainMode = pDag->explainInfo.mode;
+  pJob->attr.syncSchedule = syncSchedule;
+  pJob->transport = transport;
+  pJob->sql = sql;
+
+  if (pNodeList != NULL) {
+    pJob->nodeList = taosArrayDup(pNodeList);
+  }
+
+  SCH_ERR_JRET(schValidateAndBuildJob(pDag, pJob));
+
+  if (SCH_IS_EXPLAIN_JOB(pJob)) {
+    SCH_ERR_JRET(qExecExplainBegin(pDag, &pJob->explainCtx, startTs));
+  }
+
+  pJob->execTasks =
+      taosHashInit(pDag->numOfSubplans, taosGetDefaultHashFunction(TSDB_DATA_TYPE_UBIGINT), false, HASH_ENTRY_LOCK);
+  if (NULL == pJob->execTasks) {
+    SCH_JOB_ELOG("taosHashInit %d execTasks failed", pDag->numOfSubplans);
+    SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
+  }
+
+  pJob->succTasks =
+      taosHashInit(pDag->numOfSubplans, taosGetDefaultHashFunction(TSDB_DATA_TYPE_UBIGINT), false, HASH_ENTRY_LOCK);
+  if (NULL == pJob->succTasks) {
+    SCH_JOB_ELOG("taosHashInit %d succTasks failed", pDag->numOfSubplans);
+    SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
+  }
+
+  pJob->failTasks =
+      taosHashInit(pDag->numOfSubplans, taosGetDefaultHashFunction(TSDB_DATA_TYPE_UBIGINT), false, HASH_ENTRY_LOCK);
+  if (NULL == pJob->failTasks) {
+    SCH_JOB_ELOG("taosHashInit %d failTasks failed", pDag->numOfSubplans);
+    SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
+  }
+
+  tsem_init(&pJob->rspSem, 0, 0);
+
+  int64_t refId = taosAddRef(schMgmt.jobRef, pJob);
+  if (refId < 0) {
+    SCH_JOB_ELOG("taosAddRef job failed, error:%s", tstrerror(terrno));
+    SCH_ERR_JRET(terrno);
+  }
+
+  if (NULL == schAcquireJob(refId)) {
+    SCH_JOB_ELOG("schAcquireJob job failed, refId:%" PRIx64, refId);
+    SCH_RET(TSDB_CODE_SCH_STATUS_ERROR);
+  }
+
+  pJob->refId = refId;
+
+  SCH_JOB_DLOG("job refId:%" PRIx64, pJob->refId);
+
+  pJob->status = JOB_TASK_STATUS_NOT_START;
+
+  *pSchJob = pJob;
+
+  return TSDB_CODE_SUCCESS;
+
+_return:
+
+  schFreeJobImpl(pJob);
+  SCH_RET(code);
+}
+
+
 void schFreeRpcCtx(SRpcCtx *pCtx) {
   if (NULL == pCtx) {
     return;
@@ -1050,6 +1125,40 @@ int32_t schHandleResponseMsg(SSchJob *pJob, SSchTask *pTask, int32_t msgType, ch
 
       break;
     }
+    case TDMT_VND_EXPLAIN_RSP: {
+      SExplainRsp *taskRsp = (SExplainRsp *)msg;
+
+      SCH_ERR_JRET(rspCode);
+      if (NULL == msg) {
+        SCH_ERR_JRET(TSDB_CODE_QRY_INVALID_INPUT);
+      }
+
+      if (!SCH_IS_EXPLAIN_JOB(pJob)) {
+        SCH_TASK_ELOG("invalid msg received for none explain query, msg type:%s", TMSG_INFO(msgType));
+        SCH_ERR_JRET(TSDB_CODE_QRY_INVALID_INPUT);
+      }
+
+      if (pJob->resData) {
+        SCH_TASK_ELOG("explain result is already generated, res:%p", pJob->resData);
+        taosMemoryFreeClear(taskRsp);
+        SCH_ERR_JRET(TSDB_CODE_SCH_STATUS_ERROR);
+      }
+
+      SRetrieveTableRsp *pRsp = NULL;
+      SCH_ERR_JRET(qExplainUpdateExecInfo(pJob->explainCtx, taskRsp, pTask->plan->id.groupId, &pRsp));
+
+      if (pRsp) {
+        atomic_store_ptr(&pJob->resData, pRsp);
+        atomic_store_32(&pJob->resNumOfRows, htonl(pRsp->numOfRows));
+
+        SCH_SET_TASK_STATUS(pTask, JOB_TASK_STATUS_SUCCEED);
+
+        SCH_TASK_DLOG("got explain rsp, rows:%d, complete:%d", htonl(pRsp->numOfRows), pRsp->completed);
+
+        schProcessOnDataFetched(pJob);
+      }
+      break;
+    }
     case TDMT_VND_FETCH_RSP: {
       SRetrieveTableRsp *rsp = (SRetrieveTableRsp *)msg;
 
@@ -1058,6 +1167,17 @@ int32_t schHandleResponseMsg(SSchJob *pJob, SSchTask *pTask, int32_t msgType, ch
         SCH_ERR_JRET(TSDB_CODE_QRY_INVALID_INPUT);
       }
 
+      if (SCH_IS_EXPLAIN_JOB(pJob)) {
+        if (rsp->completed) {
+          SCH_ERR_JRET(qExecExplainEnd(pJob->explainCtx));
+          return TSDB_CODE_SUCCESS;
+        }
+
+        SCH_ERR_JRET(schFetchFromRemote(pJob));
+
+        return TSDB_CODE_SUCCESS;
+      }
+
       if (pJob->resData) {
         SCH_TASK_ELOG("got fetch rsp while res already exists, res:%p", pJob->resData);
         taosMemoryFreeClear(rsp);
@@ -1767,6 +1887,7 @@ int32_t schBuildAndSendMsg(SSchJob *pJob, SSchTask *pTask, SQueryNodeAddr *addr,
       pMsg->taskId = htobe64(pTask->taskId);
       pMsg->refId = htobe64(pJob->refId);
       pMsg->taskType = TASK_TYPE_TEMP;
+      pMsg->explain = SCH_IS_EXPLAIN_JOB(pJob);
       pMsg->phyLen = htonl(pTask->msgLen);
       pMsg->sqlLen = htonl(len);
 
@@ -2083,6 +2204,8 @@ void schFreeJobImpl(void *job) {
   taosArrayDestroy(pJob->levels);
   taosArrayDestroy(pJob->nodeList);
 
+  qExplainFreeCtx(pJob->explainCtx);
+
   taosMemoryFreeClear(pJob->resData);
   taosMemoryFreeClear(pJob);
 
@@ -2090,70 +2213,17 @@ void schFreeJobImpl(void *job) {
 }
 
 static int32_t schExecJobImpl(void *transport, SArray *pNodeList, SQueryPlan *pDag, int64_t *job, const char *sql,
-                              bool syncSchedule) {
+                              int64_t startTs, bool syncSchedule) {
   qDebug("QID:0x%" PRIx64 " job started", pDag->queryId);
 
-  if (pNodeList == NULL || (pNodeList && taosArrayGetSize(pNodeList) <= 0)) {
+  if (pNodeList == NULL || taosArrayGetSize(pNodeList) <= 0) {
     qDebug("QID:0x%" PRIx64 " input exec nodeList is empty", pDag->queryId);
   }
 
   int32_t  code = 0;
-  SSchJob *pJob = taosMemoryCalloc(1, sizeof(SSchJob));
-  if (NULL == pJob) {
-    qError("QID:%" PRIx64 " calloc %d failed", pDag->queryId, (int32_t)sizeof(SSchJob));
-    SCH_ERR_RET(TSDB_CODE_QRY_OUT_OF_MEMORY);
-  }
+  SSchJob *pJob = NULL;
+  SCH_ERR_JRET(schInitJob(&pJob, pDag, transport, pNodeList, sql, startTs, syncSchedule));
 
-  pJob->attr.explainMode = pDag->explainInfo.mode;
-  pJob->attr.syncSchedule = syncSchedule;
-  pJob->transport = transport;
-  pJob->sql = sql;
-
-  if (pNodeList != NULL) {
-    pJob->nodeList = taosArrayDup(pNodeList);
-  }
-
-  SCH_ERR_JRET(schValidateAndBuildJob(pDag, pJob));
-
-  pJob->execTasks =
-      taosHashInit(pDag->numOfSubplans, taosGetDefaultHashFunction(TSDB_DATA_TYPE_UBIGINT), false, HASH_ENTRY_LOCK);
-  if (NULL == pJob->execTasks) {
-    SCH_JOB_ELOG("taosHashInit %d execTasks failed", pDag->numOfSubplans);
-    SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
-  }
-
-  pJob->succTasks =
-      taosHashInit(pDag->numOfSubplans, taosGetDefaultHashFunction(TSDB_DATA_TYPE_UBIGINT), false, HASH_ENTRY_LOCK);
-  if (NULL == pJob->succTasks) {
-    SCH_JOB_ELOG("taosHashInit %d succTasks failed", pDag->numOfSubplans);
-    SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
-  }
-
-  pJob->failTasks =
-      taosHashInit(pDag->numOfSubplans, taosGetDefaultHashFunction(TSDB_DATA_TYPE_UBIGINT), false, HASH_ENTRY_LOCK);
-  if (NULL == pJob->failTasks) {
-    SCH_JOB_ELOG("taosHashInit %d failTasks failed", pDag->numOfSubplans);
-    SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
-  }
-
-  tsem_init(&pJob->rspSem, 0, 0);
-
-  int64_t refId = taosAddRef(schMgmt.jobRef, pJob);
-  if (refId < 0) {
-    SCH_JOB_ELOG("taosAddRef job failed, error:%s", tstrerror(terrno));
-    SCH_ERR_JRET(terrno);
-  }
-
-  if (NULL == schAcquireJob(refId)) {
-    SCH_JOB_ELOG("schAcquireJob job failed, refId:%" PRIx64, refId);
-    SCH_RET(TSDB_CODE_SCH_STATUS_ERROR);
-  }
-
-  pJob->refId = refId;
-
-  SCH_JOB_DLOG("job refId:%" PRIx64, pJob->refId);
-
-  pJob->status = JOB_TASK_STATUS_NOT_START;
   SCH_ERR_JRET(schLaunchJob(pJob));
 
   *job = pJob->refId;
@@ -2266,7 +2336,7 @@ int32_t schedulerInit(SSchedulerCfg *cfg) {
 }
 
 int32_t schedulerExecJob(void *transport, SArray *nodeList, SQueryPlan *pDag, int64_t *pJob, const char *sql,
-                         SQueryResult *pRes) {
+                         int64_t startTs, SQueryResult *pRes) {
   if (NULL == transport || NULL == pDag || NULL == pDag->pSubplans || NULL == pJob || NULL == pRes) {
     SCH_ERR_RET(TSDB_CODE_QRY_INVALID_INPUT);
   }
@@ -2274,7 +2344,7 @@ int32_t schedulerExecJob(void *transport, SArray *nodeList, SQueryPlan *pDag, in
   if (EXPLAIN_MODE_STATIC == pDag->explainInfo.mode) {
     SCH_ERR_RET(schExecStaticExplain(transport, nodeList, pDag, pJob, sql, true));
   } else {
-    SCH_ERR_RET(schExecJobImpl(transport, nodeList, pDag, pJob, sql, true));
+    SCH_ERR_RET(schExecJobImpl(transport, nodeList, pDag, pJob, sql, startTs, true));
   }
 
   SSchJob *job = schAcquireJob(*pJob);
@@ -2292,7 +2362,11 @@ int32_t schedulerAsyncExecJob(void *transport, SArray *pNodeList, SQueryPlan *pD
     SCH_ERR_RET(TSDB_CODE_QRY_INVALID_INPUT);
   }
 
-  SCH_ERR_RET(schExecJobImpl(transport, pNodeList, pDag, pJob, sql, false));
+  if (EXPLAIN_MODE_STATIC == pDag->explainInfo.mode) {
+    SCH_ERR_RET(schExecStaticExplain(transport, pNodeList, pDag, pJob, sql, false));
+  } else {
+    SCH_ERR_RET(schExecJobImpl(transport, pNodeList, pDag, pJob, sql, 0, false));
+  }
 
   return TSDB_CODE_SUCCESS;
 }
diff --git a/source/libs/scheduler/test/schedulerTests.cpp b/source/libs/scheduler/test/schedulerTests.cpp
index cf04b06579..5d9322a575 100644
--- a/source/libs/scheduler/test/schedulerTests.cpp
+++ b/source/libs/scheduler/test/schedulerTests.cpp
@@ -951,7 +951,7 @@ TEST(insertTest, normalCase) {
   taosThreadCreate(&(thread1), &thattr, schtSendRsp, &insertJobRefId);
 
   SQueryResult res = {0};
-  code = schedulerExecJob(mockPointer, qnodeList, &dag, &insertJobRefId, "insert into tb values(now,1)", &res);
+  code = schedulerExecJob(mockPointer, qnodeList, &dag, &insertJobRefId, "insert into tb values(now,1)", 0, &res);
   ASSERT_EQ(code, 0);
   ASSERT_EQ(res.numOfRows, 20);
 

From 08ed99b0692cef6d791887dd2351f6bb8f1d1523 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Sat, 2 Apr 2022 02:56:09 +0000
Subject: [PATCH 02/25] more TDB

---
 source/libs/tdb/src/btree/tdbBtreeBalance.c | 14 ------
 source/libs/tdb/src/btree/tdbBtreeCommon.c  | 14 ------
 source/libs/tdb/src/btree/tdbBtreeDelete.c  | 14 ------
 source/libs/tdb/src/btree/tdbBtreeInsert.c  | 14 ------
 source/libs/tdb/src/btree/tdbBtreeInt.h     | 27 ------------
 source/libs/tdb/src/btree/tdbBtreeOpen.c    | 14 ------
 source/libs/tdb/src/db/tdbPCache.c          | 47 ++++++++++++---------
 source/libs/tdb/src/inc/tdbPCache.h         |  1 +
 source/libs/tdb/src/inc/tdbTxn.h            |  1 +
 9 files changed, 29 insertions(+), 117 deletions(-)
 delete mode 100644 source/libs/tdb/src/btree/tdbBtreeBalance.c
 delete mode 100644 source/libs/tdb/src/btree/tdbBtreeCommon.c
 delete mode 100644 source/libs/tdb/src/btree/tdbBtreeDelete.c
 delete mode 100644 source/libs/tdb/src/btree/tdbBtreeInsert.c
 delete mode 100644 source/libs/tdb/src/btree/tdbBtreeInt.h
 delete mode 100644 source/libs/tdb/src/btree/tdbBtreeOpen.c

diff --git a/source/libs/tdb/src/btree/tdbBtreeBalance.c b/source/libs/tdb/src/btree/tdbBtreeBalance.c
deleted file mode 100644
index 6dea4a4e57..0000000000
--- a/source/libs/tdb/src/btree/tdbBtreeBalance.c
+++ /dev/null
@@ -1,14 +0,0 @@
-/*
- * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
- *
- * This program is free software: you can use, redistribute, and/or modify
- * it under the terms of the GNU Affero General Public License, version 3
- * or later ("AGPL"), as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but WITHOUT
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
- * FITNESS FOR A PARTICULAR PURPOSE.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program. If not, see <http://www.gnu.org/licenses/>.
- */
\ No newline at end of file
diff --git a/source/libs/tdb/src/btree/tdbBtreeCommon.c b/source/libs/tdb/src/btree/tdbBtreeCommon.c
deleted file mode 100644
index 6dea4a4e57..0000000000
--- a/source/libs/tdb/src/btree/tdbBtreeCommon.c
+++ /dev/null
@@ -1,14 +0,0 @@
-/*
- * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
- *
- * This program is free software: you can use, redistribute, and/or modify
- * it under the terms of the GNU Affero General Public License, version 3
- * or later ("AGPL"), as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but WITHOUT
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
- * FITNESS FOR A PARTICULAR PURPOSE.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program. If not, see <http://www.gnu.org/licenses/>.
- */
\ No newline at end of file
diff --git a/source/libs/tdb/src/btree/tdbBtreeDelete.c b/source/libs/tdb/src/btree/tdbBtreeDelete.c
deleted file mode 100644
index 6dea4a4e57..0000000000
--- a/source/libs/tdb/src/btree/tdbBtreeDelete.c
+++ /dev/null
@@ -1,14 +0,0 @@
-/*
- * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
- *
- * This program is free software: you can use, redistribute, and/or modify
- * it under the terms of the GNU Affero General Public License, version 3
- * or later ("AGPL"), as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but WITHOUT
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
- * FITNESS FOR A PARTICULAR PURPOSE.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program. If not, see <http://www.gnu.org/licenses/>.
- */
\ No newline at end of file
diff --git a/source/libs/tdb/src/btree/tdbBtreeInsert.c b/source/libs/tdb/src/btree/tdbBtreeInsert.c
deleted file mode 100644
index 6dea4a4e57..0000000000
--- a/source/libs/tdb/src/btree/tdbBtreeInsert.c
+++ /dev/null
@@ -1,14 +0,0 @@
-/*
- * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
- *
- * This program is free software: you can use, redistribute, and/or modify
- * it under the terms of the GNU Affero General Public License, version 3
- * or later ("AGPL"), as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but WITHOUT
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
- * FITNESS FOR A PARTICULAR PURPOSE.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program. If not, see <http://www.gnu.org/licenses/>.
- */
\ No newline at end of file
diff --git a/source/libs/tdb/src/btree/tdbBtreeInt.h b/source/libs/tdb/src/btree/tdbBtreeInt.h
deleted file mode 100644
index b8a935a614..0000000000
--- a/source/libs/tdb/src/btree/tdbBtreeInt.h
+++ /dev/null
@@ -1,27 +0,0 @@
-/*
- * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
- *
- * This program is free software: you can use, redistribute, and/or modify
- * it under the terms of the GNU Affero General Public License, version 3
- * or later ("AGPL"), as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but WITHOUT
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
- * FITNESS FOR A PARTICULAR PURPOSE.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program. If not, see <http://www.gnu.org/licenses/>.
- */
-
-#ifndef _TDB_BTREE_INT_H_
-#define _TDB_BTREE_INT_H_
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif /*_TDB_BTREE_INT_H_*/
\ No newline at end of file
diff --git a/source/libs/tdb/src/btree/tdbBtreeOpen.c b/source/libs/tdb/src/btree/tdbBtreeOpen.c
deleted file mode 100644
index 6dea4a4e57..0000000000
--- a/source/libs/tdb/src/btree/tdbBtreeOpen.c
+++ /dev/null
@@ -1,14 +0,0 @@
-/*
- * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
- *
- * This program is free software: you can use, redistribute, and/or modify
- * it under the terms of the GNU Affero General Public License, version 3
- * or later ("AGPL"), as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but WITHOUT
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
- * FITNESS FOR A PARTICULAR PURPOSE.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program. If not, see <http://www.gnu.org/licenses/>.
- */
\ No newline at end of file
diff --git a/source/libs/tdb/src/db/tdbPCache.c b/source/libs/tdb/src/db/tdbPCache.c
index d886cfd889..024143c8b2 100644
--- a/source/libs/tdb/src/db/tdbPCache.c
+++ b/source/libs/tdb/src/db/tdbPCache.c
@@ -18,6 +18,7 @@ struct SPCache {
   int         pageSize;
   int         cacheSize;
   tdb_mutex_t mutex;
+  SPage      *pList;
   int         nFree;
   SPage      *pFree;
   int         nPage;
@@ -35,16 +36,17 @@ struct SPCache {
 #define PAGE_IS_PINNED(pPage) ((pPage)->pLruNext == NULL)
 
 static int    tdbPCacheOpenImpl(SPCache *pCache);
-static void   tdbPCacheInitLock(SPCache *pCache);
-static void   tdbPCacheClearLock(SPCache *pCache);
-static void   tdbPCacheLock(SPCache *pCache);
-static void   tdbPCacheUnlock(SPCache *pCache);
-static bool   tdbPCacheLocked(SPCache *pCache);
 static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid, bool alcNewPage);
 static void   tdbPCachePinPage(SPCache *pCache, SPage *pPage);
 static void   tdbPCacheRemovePageFromHash(SPCache *pCache, SPage *pPage);
 static void   tdbPCacheAddPageToHash(SPCache *pCache, SPage *pPage);
 static void   tdbPCacheUnpinPage(SPCache *pCache, SPage *pPage);
+static int    tdbPCacheCloseImpl(SPCache *pCache);
+
+static void tdbPCacheInitLock(SPCache *pCache) { tdbMutexInit(&(pCache->mutex), NULL); }
+static void tdbPCacheDestroyLock(SPCache *pCache) { tdbMutexDestroy(&(pCache->mutex)); }
+static void tdbPCacheLock(SPCache *pCache) { tdbMutexLock(&(pCache->mutex)); }
+static void tdbPCacheUnlock(SPCache *pCache) { tdbMutexUnlock(&(pCache->mutex)); }
 
 int tdbPCacheOpen(int pageSize, int cacheSize, SPCache **ppCache) {
   SPCache *pCache;
@@ -69,7 +71,10 @@ int tdbPCacheOpen(int pageSize, int cacheSize, SPCache **ppCache) {
 }
 
 int tdbPCacheClose(SPCache *pCache) {
-  /* TODO */
+  if (pCache) {
+    tdbPCacheCloseImpl(pCache);
+    tdbOsFree(pCache);
+  }
   return 0;
 }
 
@@ -99,19 +104,7 @@ void tdbPCacheRelease(SPCache *pCache, SPage *pPage) {
   }
 }
 
-static void tdbPCacheInitLock(SPCache *pCache) { tdbMutexInit(&(pCache->mutex), NULL); }
-
-static void tdbPCacheClearLock(SPCache *pCache) { tdbMutexDestroy(&(pCache->mutex)); }
-
-static void tdbPCacheLock(SPCache *pCache) { tdbMutexLock(&(pCache->mutex)); }
-
-static void tdbPCacheUnlock(SPCache *pCache) { tdbMutexUnlock(&(pCache->mutex)); }
-
-static bool tdbPCacheLocked(SPCache *pCache) {
-  assert(0);
-  // TODO
-  return true;
-}
+int tdbPCacheGetPageSize(SPCache *pCache) { return pCache->pageSize; }
 
 static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid, bool alcNewPage) {
   SPage *pPage;
@@ -268,4 +261,18 @@ static int tdbPCacheOpenImpl(SPCache *pCache) {
   return 0;
 }
 
-int tdbPCacheGetPageSize(SPCache *pCache) { return pCache->pageSize; }
+static int tdbPCacheDestroyPage(SPage *pPage) {
+  // TODO
+  return 0;
+}
+
+static int tdbPCacheCloseImpl(SPCache *pCache) {
+  SPage *pPage;
+
+  for (pPage = pCache->pList; pPage; pPage = pCache->pList) {
+    pCache->pList = pPage->pCacheNext;
+    tdbPCacheDestroyPage(pPage);
+  }
+
+  tdbPCacheDestroyLock(pCache);
+}
diff --git a/source/libs/tdb/src/inc/tdbPCache.h b/source/libs/tdb/src/inc/tdbPCache.h
index f71d34ab53..eab5d7582a 100644
--- a/source/libs/tdb/src/inc/tdbPCache.h
+++ b/source/libs/tdb/src/inc/tdbPCache.h
@@ -25,6 +25,7 @@ extern "C" {
   u8      isLocalPage;  \
   u8      isDirty;      \
   i32     nRef;         \
+  SPage  *pCacheNext;   \
   SPage  *pFreeNext;    \
   SPage  *pHashNext;    \
   SPage  *pLruNext;     \
diff --git a/source/libs/tdb/src/inc/tdbTxn.h b/source/libs/tdb/src/inc/tdbTxn.h
index 0be2dad3c2..d979249f03 100644
--- a/source/libs/tdb/src/inc/tdbTxn.h
+++ b/source/libs/tdb/src/inc/tdbTxn.h
@@ -25,6 +25,7 @@ typedef struct STxn TXN;
 struct STxn {
   u64 txnId;
   void *(*xMalloc)(void *, int);
+  void (*xFree)(void *, void *);
   void *xArg;
 };
 

From 9dea86249cb1fce10aa80ec1cc9fa44707d4b205 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Sat, 2 Apr 2022 05:16:32 +0000
Subject: [PATCH 03/25] refact

---
 source/libs/tdb/src/db/tdbPCache.c  | 23 ++++++++---------------
 source/libs/tdb/src/db/tdbPager.c   |  4 ++--
 source/libs/tdb/src/inc/tdbPCache.h |  2 +-
 3 files changed, 11 insertions(+), 18 deletions(-)

diff --git a/source/libs/tdb/src/db/tdbPCache.c b/source/libs/tdb/src/db/tdbPCache.c
index 024143c8b2..5d16aeaa29 100644
--- a/source/libs/tdb/src/db/tdbPCache.c
+++ b/source/libs/tdb/src/db/tdbPCache.c
@@ -36,7 +36,7 @@ struct SPCache {
 #define PAGE_IS_PINNED(pPage) ((pPage)->pLruNext == NULL)
 
 static int    tdbPCacheOpenImpl(SPCache *pCache);
-static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid, bool alcNewPage);
+static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid);
 static void   tdbPCachePinPage(SPCache *pCache, SPage *pPage);
 static void   tdbPCacheRemovePageFromHash(SPCache *pCache, SPage *pPage);
 static void   tdbPCacheAddPageToHash(SPCache *pCache, SPage *pPage);
@@ -78,12 +78,12 @@ int tdbPCacheClose(SPCache *pCache) {
   return 0;
 }
 
-SPage *tdbPCacheFetch(SPCache *pCache, const SPgid *pPgid, bool alcNewPage) {
+SPage *tdbPCacheFetch(SPCache *pCache, const SPgid *pPgid) {
   SPage *pPage;
 
   tdbPCacheLock(pCache);
 
-  pPage = tdbPCacheFetchImpl(pCache, pPgid, alcNewPage);
+  pPage = tdbPCacheFetchImpl(pCache, pPgid);
   if (pPage) {
     TDB_REF_PAGE(pPage);
   }
@@ -106,7 +106,7 @@ void tdbPCacheRelease(SPCache *pCache, SPage *pPage) {
 
 int tdbPCacheGetPageSize(SPCache *pCache) { return pCache->pageSize; }
 
-static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid, bool alcNewPage) {
+static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid) {
   SPage *pPage;
 
   // 1. Search the hash table
@@ -116,12 +116,10 @@ static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid, bool alcNe
     pPage = pPage->pHashNext;
   }
 
-  if (pPage || !alcNewPage) {
-    if (pPage) {
-      tdbPCachePinPage(pCache, pPage);
-    }
-    return pPage;
+  if (pPage) {
+    tdbPCachePinPage(pCache, pPage);
   }
+  return pPage;
 
   // 2. Try to allocate a new page from the free list
   if (pCache->pFree) {
@@ -261,17 +259,12 @@ static int tdbPCacheOpenImpl(SPCache *pCache) {
   return 0;
 }
 
-static int tdbPCacheDestroyPage(SPage *pPage) {
-  // TODO
-  return 0;
-}
-
 static int tdbPCacheCloseImpl(SPCache *pCache) {
   SPage *pPage;
 
   for (pPage = pCache->pList; pPage; pPage = pCache->pList) {
     pCache->pList = pPage->pCacheNext;
-    tdbPCacheDestroyPage(pPage);
+    tdbPageDestroy(pPage, NULL, NULL);
   }
 
   tdbPCacheDestroyLock(pCache);
diff --git a/source/libs/tdb/src/db/tdbPager.c b/source/libs/tdb/src/db/tdbPager.c
index 2bc40a6aad..1953419bf4 100644
--- a/source/libs/tdb/src/db/tdbPager.c
+++ b/source/libs/tdb/src/db/tdbPager.c
@@ -227,7 +227,7 @@ int tdbPagerFetchPage(SPager *pPager, SPgno pgno, SPage **ppPage, int (*initPage
   // Fetch a page container from the page cache
   memcpy(&pgid, pPager->fid, TDB_FILE_ID_LEN);
   pgid.pgno = pgno;
-  pPage = tdbPCacheFetch(pPager->pCache, &pgid, 1);
+  pPage = tdbPCacheFetch(pPager->pCache, &pgid);
   if (pPage == NULL) {
     return -1;
   }
@@ -263,7 +263,7 @@ int tdbPagerNewPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage
   // Fetch a page container from the page cache
   memcpy(&pgid, pPager->fid, TDB_FILE_ID_LEN);
   pgid.pgno = *ppgno;
-  pPage = tdbPCacheFetch(pPager->pCache, &pgid, 1);
+  pPage = tdbPCacheFetch(pPager->pCache, &pgid);
   if (pPage == NULL) {
     return -1;
   }
diff --git a/source/libs/tdb/src/inc/tdbPCache.h b/source/libs/tdb/src/inc/tdbPCache.h
index eab5d7582a..c5eba88545 100644
--- a/source/libs/tdb/src/inc/tdbPCache.h
+++ b/source/libs/tdb/src/inc/tdbPCache.h
@@ -48,7 +48,7 @@ extern "C" {
 
 int    tdbPCacheOpen(int pageSize, int cacheSize, SPCache **ppCache);
 int    tdbPCacheClose(SPCache *pCache);
-SPage *tdbPCacheFetch(SPCache *pCache, const SPgid *pPgid, bool alcNewPage);
+SPage *tdbPCacheFetch(SPCache *pCache, const SPgid *pPgid);
 void   tdbPCacheRelease(SPCache *pCache, SPage *pPage);
 int    tdbPCacheGetPageSize(SPCache *pCache);
 

From f9e699bea5a3fdf42d57a000e9d0a0afea845496 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Sat, 2 Apr 2022 06:23:04 +0000
Subject: [PATCH 04/25] more TDB

---
 source/libs/tdb/src/db/tdbEnv.c     | 10 +++---
 source/libs/tdb/src/db/tdbPCache.c  | 53 +++++++++++++++++++++--------
 source/libs/tdb/src/db/tdbPager.c   |  8 ++---
 source/libs/tdb/src/inc/tdbEnv.h    |  6 ++--
 source/libs/tdb/src/inc/tdbInt.h    |  7 ++++
 source/libs/tdb/src/inc/tdbPCache.h |  4 +--
 source/libs/tdb/src/inc/tdbPager.h  |  4 +--
 source/libs/tdb/src/inc/tdbTxn.h    |  9 -----
 source/libs/tdb/test/tdbTest.cpp    |  6 ++--
 9 files changed, 63 insertions(+), 44 deletions(-)

diff --git a/source/libs/tdb/src/db/tdbEnv.c b/source/libs/tdb/src/db/tdbEnv.c
index 06d37df653..779435861e 100644
--- a/source/libs/tdb/src/db/tdbEnv.c
+++ b/source/libs/tdb/src/db/tdbEnv.c
@@ -73,12 +73,12 @@ int tdbEnvClose(TENV *pEnv) {
   return 0;
 }
 
-int tdbBegin(TENV *pEnv) {
+int tdbBegin(TENV *pEnv, TXN *pTxn) {
   SPager *pPager;
   int     ret;
 
   for (pPager = pEnv->pgrList; pPager; pPager = pPager->pNext) {
-    ret = tdbPagerBegin(pPager);
+    ret = tdbPagerBegin(pPager, pTxn);
     if (ret < 0) {
       ASSERT(0);
       return -1;
@@ -88,12 +88,12 @@ int tdbBegin(TENV *pEnv) {
   return 0;
 }
 
-int tdbCommit(TENV *pEnv) {
+int tdbCommit(TENV *pEnv, TXN *pTxn) {
   SPager *pPager;
   int     ret;
 
   for (pPager = pEnv->pgrList; pPager; pPager = pPager->pNext) {
-    ret = tdbPagerCommit(pPager);
+    ret = tdbPagerCommit(pPager, pTxn);
     if (ret < 0) {
       ASSERT(0);
       return -1;
@@ -103,7 +103,7 @@ int tdbCommit(TENV *pEnv) {
   return 0;
 }
 
-int tdbRollback(TENV *pEnv) {
+int tdbRollback(TENV *pEnv, TXN *pTxn) {
   ASSERT(0);
   return 0;
 }
diff --git a/source/libs/tdb/src/db/tdbPCache.c b/source/libs/tdb/src/db/tdbPCache.c
index 5d16aeaa29..b2d370a668 100644
--- a/source/libs/tdb/src/db/tdbPCache.c
+++ b/source/libs/tdb/src/db/tdbPCache.c
@@ -36,7 +36,7 @@ struct SPCache {
 #define PAGE_IS_PINNED(pPage) ((pPage)->pLruNext == NULL)
 
 static int    tdbPCacheOpenImpl(SPCache *pCache);
-static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid);
+static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid, TXN *pTxn);
 static void   tdbPCachePinPage(SPCache *pCache, SPage *pPage);
 static void   tdbPCacheRemovePageFromHash(SPCache *pCache, SPage *pPage);
 static void   tdbPCacheAddPageToHash(SPCache *pCache, SPage *pPage);
@@ -78,12 +78,12 @@ int tdbPCacheClose(SPCache *pCache) {
   return 0;
 }
 
-SPage *tdbPCacheFetch(SPCache *pCache, const SPgid *pPgid) {
+SPage *tdbPCacheFetch(SPCache *pCache, const SPgid *pPgid, TXN *pTxn) {
   SPage *pPage;
 
   tdbPCacheLock(pCache);
 
-  pPage = tdbPCacheFetchImpl(pCache, pPgid);
+  pPage = tdbPCacheFetchImpl(pCache, pPgid, pTxn);
   if (pPage) {
     TDB_REF_PAGE(pPage);
   }
@@ -106,7 +106,8 @@ void tdbPCacheRelease(SPCache *pCache, SPage *pPage) {
 
 int tdbPCacheGetPageSize(SPCache *pCache) { return pCache->pageSize; }
 
-static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid) {
+static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid, TXN *pTxn) {
+  int    ret;
   SPage *pPage;
 
   // 1. Search the hash table
@@ -117,9 +118,11 @@ static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid) {
   }
 
   if (pPage) {
+    // TODO: the page need to be copied and
+    // replaced the page in hash table
     tdbPCachePinPage(pCache, pPage);
+    return pPage;
   }
-  return pPage;
 
   // 2. Try to allocate a new page from the free list
   if (pCache->pFree) {
@@ -136,7 +139,20 @@ static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid) {
     tdbPCachePinPage(pCache, pPage);
   }
 
-  // 4. Try a stress allocation (TODO)
+  // 4. Try a create new page
+  if (pTxn && pTxn->xMalloc) {
+    ret = tdbPageCreate(pCache->pageSize, &pPage, pTxn->xMalloc, pTxn->xArg);
+    if (ret < 0) {
+      // TODO
+      ASSERT(0);
+      return NULL;
+    }
+
+    // init the page fields
+    pPage->isAnchor = 0;
+    pPage->isLocal = 0;
+    TDB_INIT_PAGE_REF(pPage);
+  }
 
   // 5. Page here are just created from a free list
   // or by recycling or allocated streesly,
@@ -145,6 +161,8 @@ static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid) {
     memcpy(&(pPage->pgid), pPgid, sizeof(*pPgid));
     pPage->pLruNext = NULL;
     pPage->pPager = NULL;
+
+    // TODO: allocated page may not add to hash
     tdbPCacheAddPageToHash(pCache, pPage);
   }
 
@@ -171,17 +189,21 @@ static void tdbPCacheUnpinPage(SPCache *pCache, SPage *pPage) {
   nRef = TDB_GET_PAGE_REF(pPage);
   ASSERT(nRef >= 0);
   if (nRef == 0) {
-    // Add the page to LRU list
-    ASSERT(pPage->pLruNext == NULL);
+    if (1) {
+      // Add the page to LRU list
+      ASSERT(pPage->pLruNext == NULL);
 
-    pPage->pLruPrev = &(pCache->lru);
-    pPage->pLruNext = pCache->lru.pLruNext;
-    pCache->lru.pLruNext->pLruPrev = pPage;
-    pCache->lru.pLruNext = pPage;
+      pPage->pLruPrev = &(pCache->lru);
+      pPage->pLruNext = pCache->lru.pLruNext;
+      pCache->lru.pLruNext->pLruPrev = pPage;
+      pCache->lru.pLruNext = pPage;
+
+      pCache->nRecyclable++;
+    } else {
+      // TODO: may need to free the page
+    }
   }
 
-  pCache->nRecyclable++;
-
   tdbPCacheUnlock(pCache);
 }
 
@@ -229,13 +251,14 @@ static int tdbPCacheOpenImpl(SPCache *pCache) {
 
     // pPage->pgid = 0;
     pPage->isAnchor = 0;
-    pPage->isLocalPage = 1;
+    pPage->isLocal = 1;
     TDB_INIT_PAGE_REF(pPage);
     pPage->pHashNext = NULL;
     pPage->pLruNext = NULL;
     pPage->pLruPrev = NULL;
     pPage->pDirtyNext = NULL;
 
+    // add page to free list
     pPage->pFreeNext = pCache->pFree;
     pCache->pFree = pPage;
     pCache->nFree++;
diff --git a/source/libs/tdb/src/db/tdbPager.c b/source/libs/tdb/src/db/tdbPager.c
index 1953419bf4..13db278c6e 100644
--- a/source/libs/tdb/src/db/tdbPager.c
+++ b/source/libs/tdb/src/db/tdbPager.c
@@ -157,7 +157,7 @@ int tdbPagerWrite(SPager *pPager, SPage *pPage) {
   return 0;
 }
 
-int tdbPagerBegin(SPager *pPager) {
+int tdbPagerBegin(SPager *pPager, TXN *pTxn) {
   if (pPager->inTran) {
     return 0;
   }
@@ -175,7 +175,7 @@ int tdbPagerBegin(SPager *pPager) {
   return 0;
 }
 
-int tdbPagerCommit(SPager *pPager) {
+int tdbPagerCommit(SPager *pPager, TXN *pTxn) {
   SPage *pPage;
   int    ret;
 
@@ -227,7 +227,7 @@ int tdbPagerFetchPage(SPager *pPager, SPgno pgno, SPage **ppPage, int (*initPage
   // Fetch a page container from the page cache
   memcpy(&pgid, pPager->fid, TDB_FILE_ID_LEN);
   pgid.pgno = pgno;
-  pPage = tdbPCacheFetch(pPager->pCache, &pgid);
+  pPage = tdbPCacheFetch(pPager->pCache, &pgid, NULL);
   if (pPage == NULL) {
     return -1;
   }
@@ -263,7 +263,7 @@ int tdbPagerNewPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage
   // Fetch a page container from the page cache
   memcpy(&pgid, pPager->fid, TDB_FILE_ID_LEN);
   pgid.pgno = *ppgno;
-  pPage = tdbPCacheFetch(pPager->pCache, &pgid);
+  pPage = tdbPCacheFetch(pPager->pCache, &pgid, NULL);
   if (pPage == NULL) {
     return -1;
   }
diff --git a/source/libs/tdb/src/inc/tdbEnv.h b/source/libs/tdb/src/inc/tdbEnv.h
index e10c5d54e0..54f5818be1 100644
--- a/source/libs/tdb/src/inc/tdbEnv.h
+++ b/source/libs/tdb/src/inc/tdbEnv.h
@@ -33,9 +33,9 @@ typedef struct STEnv {
 
 int tdbEnvOpen(const char *rootDir, int pageSize, int cacheSize, TENV **ppEnv);
 int tdbEnvClose(TENV *pEnv);
-int tdbBegin(TENV *pEnv);
-int tdbCommit(TENV *pEnv);
-int tdbRollback(TENV *pEnv);
+int tdbBegin(TENV *pEnv, TXN *pTxn);
+int tdbCommit(TENV *pEnv, TXN *pTxn);
+int tdbRollback(TENV *pEnv, TXN *pTxn);
 
 void    tdbEnvAddPager(TENV *pEnv, SPager *pPager);
 void    tdbEnvRemovePager(TENV *pEnv, SPager *pPager);
diff --git a/source/libs/tdb/src/inc/tdbInt.h b/source/libs/tdb/src/inc/tdbInt.h
index 57e01f904c..2a196b2cc2 100644
--- a/source/libs/tdb/src/inc/tdbInt.h
+++ b/source/libs/tdb/src/inc/tdbInt.h
@@ -111,6 +111,13 @@ typedef struct SPager  SPager;
 typedef struct SPCache SPCache;
 typedef struct SPage   SPage;
 
+typedef struct STxn {
+  u64 txnId;
+  void *(*xMalloc)(void *, size_t);
+  void (*xFree)(void *, void *);
+  void *xArg;
+} TXN;
+
 #include "tdbOs.h"
 
 #include "tdbUtil.h"
diff --git a/source/libs/tdb/src/inc/tdbPCache.h b/source/libs/tdb/src/inc/tdbPCache.h
index c5eba88545..5b24a53591 100644
--- a/source/libs/tdb/src/inc/tdbPCache.h
+++ b/source/libs/tdb/src/inc/tdbPCache.h
@@ -22,7 +22,7 @@ extern "C" {
 
 #define TDB_PCACHE_PAGE \
   u8      isAnchor;     \
-  u8      isLocalPage;  \
+  u8      isLocal;      \
   u8      isDirty;      \
   i32     nRef;         \
   SPage  *pCacheNext;   \
@@ -48,7 +48,7 @@ extern "C" {
 
 int    tdbPCacheOpen(int pageSize, int cacheSize, SPCache **ppCache);
 int    tdbPCacheClose(SPCache *pCache);
-SPage *tdbPCacheFetch(SPCache *pCache, const SPgid *pPgid);
+SPage *tdbPCacheFetch(SPCache *pCache, const SPgid *pPgid, TXN *pTxn);
 void   tdbPCacheRelease(SPCache *pCache, SPage *pPage);
 int    tdbPCacheGetPageSize(SPCache *pCache);
 
diff --git a/source/libs/tdb/src/inc/tdbPager.h b/source/libs/tdb/src/inc/tdbPager.h
index 81b6074431..cc8860ae7a 100644
--- a/source/libs/tdb/src/inc/tdbPager.h
+++ b/source/libs/tdb/src/inc/tdbPager.h
@@ -40,8 +40,8 @@ int  tdbPagerOpen(SPCache *pCache, const char *fileName, SPager **ppPager);
 int  tdbPagerClose(SPager *pPager);
 int  tdbPagerOpenDB(SPager *pPager, SPgno *ppgno, bool toCreate);
 int  tdbPagerWrite(SPager *pPager, SPage *pPage);
-int  tdbPagerBegin(SPager *pPager);
-int  tdbPagerCommit(SPager *pPager);
+int  tdbPagerBegin(SPager *pPager, TXN *pTxn);
+int  tdbPagerCommit(SPager *pPager, TXN *pTxn);
 int  tdbPagerFetchPage(SPager *pPager, SPgno pgno, SPage **ppPage, int (*initPage)(SPage *, void *), void *arg);
 int  tdbPagerNewPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage)(SPage *, void *), void *arg);
 void tdbPagerReturnPage(SPager *pPager, SPage *pPage);
diff --git a/source/libs/tdb/src/inc/tdbTxn.h b/source/libs/tdb/src/inc/tdbTxn.h
index d979249f03..13303bffb6 100644
--- a/source/libs/tdb/src/inc/tdbTxn.h
+++ b/source/libs/tdb/src/inc/tdbTxn.h
@@ -20,15 +20,6 @@
 extern "C" {
 #endif
 
-typedef struct STxn TXN;
-
-struct STxn {
-  u64 txnId;
-  void *(*xMalloc)(void *, int);
-  void (*xFree)(void *, void *);
-  void *xArg;
-};
-
 #ifdef __cplusplus
 }
 #endif
diff --git a/source/libs/tdb/test/tdbTest.cpp b/source/libs/tdb/test/tdbTest.cpp
index 9e1277a53d..7e5cfeff42 100644
--- a/source/libs/tdb/test/tdbTest.cpp
+++ b/source/libs/tdb/test/tdbTest.cpp
@@ -135,7 +135,7 @@ TEST(tdb_test, simple_test) {
 
     {  // Insert some data
       for (int i = 1; i <= nData;) {
-        tdbBegin(pEnv);
+        tdbBegin(pEnv, NULL);
 
         for (int k = 0; k < 2000; k++) {
           sprintf(key, "key%d", i);
@@ -145,12 +145,10 @@ TEST(tdb_test, simple_test) {
           i++;
         }
 
-        tdbCommit(pEnv);
+        tdbCommit(pEnv, NULL);
       }
     }
 
-    tdbCommit(pEnv);
-
     {  // Query the data
       void *pVal = NULL;
       int   vLen;

From 6b47fd8984f22266b852156d5c4c6c7ea96e707d Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Sat, 2 Apr 2022 11:20:56 +0000
Subject: [PATCH 05/25] TDB with txn

---
 source/libs/tdb/src/db/tdbBtree.c   | 45 ++++++++++++-----------
 source/libs/tdb/src/db/tdbDb.c      |  6 +--
 source/libs/tdb/src/db/tdbPCache.c  | 51 +++++++++++++++-----------
 source/libs/tdb/src/db/tdbPager.c   | 20 ++++++----
 source/libs/tdb/src/db/tdbTxn.c     | 36 ++++++------------
 source/libs/tdb/src/inc/tdbBtree.h  |  5 ++-
 source/libs/tdb/src/inc/tdbDb.h     |  2 +-
 source/libs/tdb/src/inc/tdbInt.h    | 10 ++++-
 source/libs/tdb/src/inc/tdbPCache.h |  2 +-
 source/libs/tdb/src/inc/tdbPager.h  |  9 +++--
 source/libs/tdb/src/inc/tdbTxn.h    |  4 ++
 source/libs/tdb/test/tdbTest.cpp    | 57 ++++++++++++++++++++---------
 12 files changed, 142 insertions(+), 105 deletions(-)

diff --git a/source/libs/tdb/src/db/tdbBtree.c b/source/libs/tdb/src/db/tdbBtree.c
index f4e1621742..285434b402 100644
--- a/source/libs/tdb/src/db/tdbBtree.c
+++ b/source/libs/tdb/src/db/tdbBtree.c
@@ -127,7 +127,7 @@ int tdbBtreeClose(SBTree *pBt) {
   return 0;
 }
 
-int tdbBtreeInsert(SBTree *pBt, const void *pKey, int kLen, const void *pVal, int vLen) {
+int tdbBtreeInsert(SBTree *pBt, const void *pKey, int kLen, const void *pVal, int vLen, TXN *pTxn) {
   SBTC   btc;
   SCell *pCell;
   void  *pBuf;
@@ -137,7 +137,7 @@ int tdbBtreeInsert(SBTree *pBt, const void *pKey, int kLen, const void *pVal, in
   int    idx;
   int    c;
 
-  tdbBtcOpen(&btc, pBt);
+  tdbBtcOpen(&btc, pBt, pTxn);
 
   // move to the position to insert
   ret = tdbBtcMoveTo(&btc, pKey, kLen, &c);
@@ -225,7 +225,7 @@ int tdbBtreePGet(SBTree *pBt, const void *pKey, int kLen, void **ppKey, int *pkL
   void        *pTVal = NULL;
   SCellDecoder cd;
 
-  tdbBtcOpen(&btc, pBt);
+  tdbBtcOpen(&btc, pBt, NULL);
 
   ret = tdbBtcMoveTo(&btc, pKey, kLen, &cret);
   if (ret < 0) {
@@ -307,13 +307,13 @@ static int tdbBtreeOpenImpl(SBTree *pBt) {
 
   // Try to create a new database
   SBtreeInitPageArg zArg = {.flags = TDB_BTREE_ROOT | TDB_BTREE_LEAF, .pBt = pBt};
-  ret = tdbPagerNewPage(pBt->pPager, &pgno, &pPage, tdbBtreeZeroPage, &zArg);
+  ret = tdbPagerNewPage(pBt->pPager, &pgno, &pPage, tdbBtreeZeroPage, &zArg, NULL);
   if (ret < 0) {
     return -1;
   }
 
   // TODO: here still has problem
-  tdbPagerReturnPage(pBt->pPager, pPage);
+  tdbPagerReturnPage(pBt->pPager, pPage, NULL);
 
   ASSERT(pgno != 0);
   pBt->root = pgno;
@@ -385,7 +385,7 @@ static int tdbBtreeZeroPage(SPage *pPage, void *arg) {
 }
 
 // TDB_BTREE_BALANCE =====================
-static int tdbBtreeBalanceDeeper(SBTree *pBt, SPage *pRoot, SPage **ppChild) {
+static int tdbBtreeBalanceDeeper(SBTree *pBt, SPage *pRoot, SPage **ppChild, TXN *pTxn) {
   SPager           *pPager;
   SPage            *pChild;
   SPgno             pgnoChild;
@@ -402,7 +402,7 @@ static int tdbBtreeBalanceDeeper(SBTree *pBt, SPage *pRoot, SPage **ppChild) {
   // Allocate a new child page
   zArg.flags = TDB_FLAG_REMOVE(flags, TDB_BTREE_ROOT);
   zArg.pBt = pBt;
-  ret = tdbPagerNewPage(pPager, &pgnoChild, &pChild, tdbBtreeZeroPage, &zArg);
+  ret = tdbPagerNewPage(pPager, &pgnoChild, &pChild, tdbBtreeZeroPage, &zArg, pTxn);
   if (ret < 0) {
     return -1;
   }
@@ -436,7 +436,7 @@ static int tdbBtreeBalanceDeeper(SBTree *pBt, SPage *pRoot, SPage **ppChild) {
   return 0;
 }
 
-static int tdbBtreeBalanceNonRoot(SBTree *pBt, SPage *pParent, int idx) {
+static int tdbBtreeBalanceNonRoot(SBTree *pBt, SPage *pParent, int idx, TXN *pTxn) {
   int ret;
 
   int    nOlds;
@@ -477,7 +477,7 @@ static int tdbBtreeBalanceNonRoot(SBTree *pBt, SPage *pParent, int idx) {
         pgno = *(SPgno *)pCell;
       }
 
-      ret = tdbPagerFetchPage(pBt->pPager, pgno, pOlds + i, tdbBtreeInitPage, pBt);
+      ret = tdbPagerFetchPage(pBt->pPager, pgno, pOlds + i, tdbBtreeInitPage, pBt, pTxn);
       if (ret < 0) {
         ASSERT(0);
         return -1;
@@ -640,7 +640,7 @@ static int tdbBtreeBalanceNonRoot(SBTree *pBt, SPage *pParent, int idx) {
       } else {
         iarg.pBt = pBt;
         iarg.flags = flags;
-        ret = tdbPagerNewPage(pBt->pPager, &pgno, pNews + iNew, tdbBtreeZeroPage, &iarg);
+        ret = tdbPagerNewPage(pBt->pPager, &pgno, pNews + iNew, tdbBtreeZeroPage, &iarg, pTxn);
         if (ret < 0) {
           ASSERT(0);
         }
@@ -767,9 +767,9 @@ static int tdbBtreeBalanceNonRoot(SBTree *pBt, SPage *pParent, int idx) {
   // TODO: here is not corrent for drop case
   for (int i = 0; i < nNews; i++) {
     if (i < nOlds) {
-      tdbPagerReturnPage(pBt->pPager, pOlds[i]);
+      tdbPagerReturnPage(pBt->pPager, pOlds[i], pTxn);
     } else {
-      tdbPagerReturnPage(pBt->pPager, pNews[i]);
+      tdbPagerReturnPage(pBt->pPager, pNews[i], pTxn);
     }
   }
 
@@ -805,7 +805,7 @@ static int tdbBtreeBalance(SBTC *pBtc) {
       // ignore the case of empty
       if (pPage->nOverflow == 0) break;
 
-      ret = tdbBtreeBalanceDeeper(pBtc->pBt, pPage, &(pBtc->pgStack[1]));
+      ret = tdbBtreeBalanceDeeper(pBtc->pBt, pPage, &(pBtc->pgStack[1]), pBtc->pTxn);
       if (ret < 0) {
         return -1;
       }
@@ -819,12 +819,12 @@ static int tdbBtreeBalance(SBTC *pBtc) {
       // Generalized balance step
       pParent = pBtc->pgStack[iPage - 1];
 
-      ret = tdbBtreeBalanceNonRoot(pBtc->pBt, pParent, pBtc->idxStack[pBtc->iPage - 1]);
+      ret = tdbBtreeBalanceNonRoot(pBtc->pBt, pParent, pBtc->idxStack[pBtc->iPage - 1], pBtc->pTxn);
       if (ret < 0) {
         return -1;
       }
 
-      tdbPagerReturnPage(pBtc->pBt->pPager, pBtc->pPage);
+      tdbPagerReturnPage(pBtc->pBt->pPager, pBtc->pPage, pBtc->pTxn);
 
       pBtc->iPage--;
       pBtc->pPage = pBtc->pgStack[pBtc->iPage];
@@ -1024,11 +1024,12 @@ static int tdbBtreeCellSize(const SPage *pPage, SCell *pCell) {
 // TDB_BTREE_CELL
 
 // TDB_BTREE_CURSOR =====================
-int tdbBtcOpen(SBTC *pBtc, SBTree *pBt) {
+int tdbBtcOpen(SBTC *pBtc, SBTree *pBt, TXN *pTxn) {
   pBtc->pBt = pBt;
   pBtc->iPage = -1;
   pBtc->pPage = NULL;
   pBtc->idx = -1;
+  pBtc->pTxn = pTxn;
 
   return 0;
 }
@@ -1045,7 +1046,7 @@ int tdbBtcMoveToFirst(SBTC *pBtc) {
 
   if (pBtc->iPage < 0) {
     // move a clean cursor
-    ret = tdbPagerFetchPage(pPager, pBt->root, &(pBtc->pPage), tdbBtreeInitPage, pBt);
+    ret = tdbPagerFetchPage(pPager, pBt->root, &(pBtc->pPage), tdbBtreeInitPage, pBt, pBtc->pTxn);
     if (ret < 0) {
       ASSERT(0);
       return -1;
@@ -1110,7 +1111,7 @@ int tdbBtcMoveToLast(SBTC *pBtc) {
 
   if (pBtc->iPage < 0) {
     // move a clean cursor
-    ret = tdbPagerFetchPage(pPager, pBt->root, &(pBtc->pPage), tdbBtreeInitPage, pBt);
+    ret = tdbPagerFetchPage(pPager, pBt->root, &(pBtc->pPage), tdbBtreeInitPage, pBt, pBtc->pTxn);
     if (ret < 0) {
       ASSERT(0);
       return -1;
@@ -1284,7 +1285,7 @@ static int tdbBtcMoveDownward(SBTC *pBtc) {
   pBtc->pPage = NULL;
   pBtc->idx = -1;
 
-  ret = tdbPagerFetchPage(pBtc->pBt->pPager, pgno, &pBtc->pPage, tdbBtreeInitPage, pBtc->pBt);
+  ret = tdbPagerFetchPage(pBtc->pBt->pPager, pgno, &pBtc->pPage, tdbBtreeInitPage, pBtc->pBt, pBtc->pTxn);
   if (ret < 0) {
     ASSERT(0);
     return -1;
@@ -1296,7 +1297,7 @@ static int tdbBtcMoveDownward(SBTC *pBtc) {
 static int tdbBtcMoveUpward(SBTC *pBtc) {
   if (pBtc->iPage == 0) return -1;
 
-  tdbPagerReturnPage(pBtc->pBt->pPager, pBtc->pPage);
+  tdbPagerReturnPage(pBtc->pBt->pPager, pBtc->pPage, pBtc->pTxn);
 
   pBtc->iPage--;
   pBtc->pPage = pBtc->pgStack[pBtc->iPage];
@@ -1319,7 +1320,7 @@ static int tdbBtcMoveTo(SBTC *pBtc, const void *pKey, int kLen, int *pCRst) {
 
   if (pBtc->iPage < 0) {
     // move from a clear cursor
-    ret = tdbPagerFetchPage(pPager, pBt->root, &(pBtc->pPage), tdbBtreeInitPage, pBt);
+    ret = tdbPagerFetchPage(pPager, pBt->root, &(pBtc->pPage), tdbBtreeInitPage, pBt, pBtc->pTxn);
     if (ret < 0) {
       // TODO
       ASSERT(0);
@@ -1456,7 +1457,7 @@ int tdbBtcClose(SBTC *pBtc) {
   for (;;) {
     ASSERT(pBtc->pPage);
 
-    tdbPagerReturnPage(pBtc->pBt->pPager, pBtc->pPage);
+    tdbPagerReturnPage(pBtc->pBt->pPager, pBtc->pPage, pBtc->pTxn);
 
     pBtc->iPage--;
     if (pBtc->iPage < 0) break;
diff --git a/source/libs/tdb/src/db/tdbDb.c b/source/libs/tdb/src/db/tdbDb.c
index fe7b8c6d48..f4468e7e82 100644
--- a/source/libs/tdb/src/db/tdbDb.c
+++ b/source/libs/tdb/src/db/tdbDb.c
@@ -75,8 +75,8 @@ int tdbDbDrop(TDB *pDb) {
   return 0;
 }
 
-int tdbDbInsert(TDB *pDb, const void *pKey, int keyLen, const void *pVal, int valLen) {
-  return tdbBtreeInsert(pDb->pBt, pKey, keyLen, pVal, valLen);
+int tdbDbInsert(TDB *pDb, const void *pKey, int keyLen, const void *pVal, int valLen, TXN *pTxn) {
+  return tdbBtreeInsert(pDb->pBt, pKey, keyLen, pVal, valLen, pTxn);
 }
 
 int tdbDbGet(TDB *pDb, const void *pKey, int kLen, void **ppVal, int *vLen) {
@@ -97,7 +97,7 @@ int tdbDbcOpen(TDB *pDb, TDBC **ppDbc) {
     return -1;
   }
 
-  tdbBtcOpen(&pDbc->btc, pDb->pBt);
+  tdbBtcOpen(&pDbc->btc, pDb->pBt, NULL);
 
   // TODO: move to first now, we can move to any key-value
   // and in any direction, design new APIs.
diff --git a/source/libs/tdb/src/db/tdbPCache.c b/source/libs/tdb/src/db/tdbPCache.c
index b2d370a668..b4e291f5e4 100644
--- a/source/libs/tdb/src/db/tdbPCache.c
+++ b/source/libs/tdb/src/db/tdbPCache.c
@@ -93,14 +93,33 @@ SPage *tdbPCacheFetch(SPCache *pCache, const SPgid *pPgid, TXN *pTxn) {
   return pPage;
 }
 
-void tdbPCacheRelease(SPCache *pCache, SPage *pPage) {
+void tdbPCacheRelease(SPCache *pCache, SPage *pPage, TXN *pTxn) {
   i32 nRef;
 
   nRef = TDB_UNREF_PAGE(pPage);
   ASSERT(nRef >= 0);
 
   if (nRef == 0) {
-    tdbPCacheUnpinPage(pCache, pPage);
+    tdbPCacheLock(pCache);
+
+    // test the nRef again to make sure
+    // it is safe th handle the page
+    nRef = TDB_GET_PAGE_REF(pPage);
+    if (nRef == 0) {
+      if (pPage->isLocal) {
+        tdbPCacheUnpinPage(pCache, pPage);
+      } else {
+        // remove from hash
+        tdbPCacheRemovePageFromHash(pCache, pPage);
+
+        // free the page
+        if (pTxn && pTxn->xFree) {
+          tdbPageDestroy(pPage, pTxn->xFree, pTxn->xArg);
+        }
+      }
+    }
+
+    tdbPCacheUnlock(pCache);
   }
 }
 
@@ -140,7 +159,7 @@ static SPage *tdbPCacheFetchImpl(SPCache *pCache, const SPgid *pPgid, TXN *pTxn)
   }
 
   // 4. Try a create new page
-  if (pTxn && pTxn->xMalloc) {
+  if (!pPage && pTxn && pTxn->xMalloc) {
     ret = tdbPageCreate(pCache->pageSize, &pPage, pTxn->xMalloc, pTxn->xArg);
     if (ret < 0) {
       // TODO
@@ -182,29 +201,17 @@ static void tdbPCachePinPage(SPCache *pCache, SPage *pPage) {
 static void tdbPCacheUnpinPage(SPCache *pCache, SPage *pPage) {
   i32 nRef;
 
-  tdbPCacheLock(pCache);
-
   ASSERT(!pPage->isDirty);
+  ASSERT(TDB_GET_PAGE_REF(pPage) == 0);
 
-  nRef = TDB_GET_PAGE_REF(pPage);
-  ASSERT(nRef >= 0);
-  if (nRef == 0) {
-    if (1) {
-      // Add the page to LRU list
-      ASSERT(pPage->pLruNext == NULL);
+  ASSERT(pPage->pLruNext == NULL);
 
-      pPage->pLruPrev = &(pCache->lru);
-      pPage->pLruNext = pCache->lru.pLruNext;
-      pCache->lru.pLruNext->pLruPrev = pPage;
-      pCache->lru.pLruNext = pPage;
+  pPage->pLruPrev = &(pCache->lru);
+  pPage->pLruNext = pCache->lru.pLruNext;
+  pCache->lru.pLruNext->pLruPrev = pPage;
+  pCache->lru.pLruNext = pPage;
 
-      pCache->nRecyclable++;
-    } else {
-      // TODO: may need to free the page
-    }
-  }
-
-  tdbPCacheUnlock(pCache);
+  pCache->nRecyclable++;
 }
 
 static void tdbPCacheRemovePageFromHash(SPCache *pCache, SPage *pPage) {
diff --git a/source/libs/tdb/src/db/tdbPager.c b/source/libs/tdb/src/db/tdbPager.c
index 13db278c6e..f78568b51c 100644
--- a/source/libs/tdb/src/db/tdbPager.c
+++ b/source/libs/tdb/src/db/tdbPager.c
@@ -27,7 +27,6 @@ TDB_STATIC_ASSERT(sizeof(SFileHdr) == 128, "Size of file header is not correct")
 
 #define TDB_PAGE_INITIALIZED(pPage) ((pPage)->pPager != NULL)
 
-static int tdbPagerAllocPage(SPager *pPager, SPgno *ppgno);
 static int tdbPagerInitPage(SPager *pPager, SPage *pPage, int (*initPage)(SPage *, void *), void *arg, u8 loadPage);
 static int tdbPagerWritePageToJournal(SPager *pPager, SPage *pPage);
 static int tdbPagerWritePageToDB(SPager *pPager, SPage *pPage);
@@ -204,7 +203,7 @@ int tdbPagerCommit(SPager *pPager, TXN *pTxn) {
 
     pPage->isDirty = 0;
 
-    tdbPCacheRelease(pPager->pCache, pPage);
+    tdbPCacheRelease(pPager->pCache, pPage, pTxn);
   }
 
   // sync the db file
@@ -219,7 +218,8 @@ int tdbPagerCommit(SPager *pPager, TXN *pTxn) {
   return 0;
 }
 
-int tdbPagerFetchPage(SPager *pPager, SPgno pgno, SPage **ppPage, int (*initPage)(SPage *, void *), void *arg) {
+int tdbPagerFetchPage(SPager *pPager, SPgno pgno, SPage **ppPage, int (*initPage)(SPage *, void *), void *arg,
+                      TXN *pTxn) {
   SPage *pPage;
   SPgid  pgid;
   int    ret;
@@ -227,7 +227,7 @@ int tdbPagerFetchPage(SPager *pPager, SPgno pgno, SPage **ppPage, int (*initPage
   // Fetch a page container from the page cache
   memcpy(&pgid, pPager->fid, TDB_FILE_ID_LEN);
   pgid.pgno = pgno;
-  pPage = tdbPCacheFetch(pPager->pCache, &pgid, NULL);
+  pPage = tdbPCacheFetch(pPager->pCache, &pgid, pTxn);
   if (pPage == NULL) {
     return -1;
   }
@@ -247,7 +247,8 @@ int tdbPagerFetchPage(SPager *pPager, SPgno pgno, SPage **ppPage, int (*initPage
   return 0;
 }
 
-int tdbPagerNewPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage)(SPage *, void *), void *arg) {
+int tdbPagerNewPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage)(SPage *, void *), void *arg,
+                    TXN *pTxn) {
   int    ret;
   SPage *pPage;
   SPgid  pgid;
@@ -255,6 +256,7 @@ int tdbPagerNewPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage
   // Allocate a page number
   ret = tdbPagerAllocPage(pPager, ppgno);
   if (ret < 0) {
+    ASSERT(0);
     return -1;
   }
 
@@ -263,8 +265,9 @@ int tdbPagerNewPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage
   // Fetch a page container from the page cache
   memcpy(&pgid, pPager->fid, TDB_FILE_ID_LEN);
   pgid.pgno = *ppgno;
-  pPage = tdbPCacheFetch(pPager->pCache, &pgid, NULL);
+  pPage = tdbPCacheFetch(pPager->pCache, &pgid, pTxn);
   if (pPage == NULL) {
+    ASSERT(0);
     return -1;
   }
 
@@ -273,6 +276,7 @@ int tdbPagerNewPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage
   // Initialize the page if need
   ret = tdbPagerInitPage(pPager, pPage, initPage, arg, 0);
   if (ret < 0) {
+    ASSERT(0);
     return -1;
   }
 
@@ -283,7 +287,7 @@ int tdbPagerNewPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage
   return 0;
 }
 
-void tdbPagerReturnPage(SPager *pPager, SPage *pPage) { tdbPCacheRelease(pPager->pCache, pPage); }
+void tdbPagerReturnPage(SPager *pPager, SPage *pPage, TXN *pTxn) { tdbPCacheRelease(pPager->pCache, pPage, pTxn); }
 
 static int tdbPagerAllocFreePage(SPager *pPager, SPgno *ppgno) {
   // TODO: Allocate a page from the free list
@@ -295,7 +299,7 @@ static int tdbPagerAllocNewPage(SPager *pPager, SPgno *ppgno) {
   return 0;
 }
 
-static int tdbPagerAllocPage(SPager *pPager, SPgno *ppgno) {
+int tdbPagerAllocPage(SPager *pPager, SPgno *ppgno) {
   int ret;
 
   *ppgno = 0;
diff --git a/source/libs/tdb/src/db/tdbTxn.c b/source/libs/tdb/src/db/tdbTxn.c
index 03bcbb44a7..b06fe05acd 100644
--- a/source/libs/tdb/src/db/tdbTxn.c
+++ b/source/libs/tdb/src/db/tdbTxn.c
@@ -15,29 +15,17 @@
 
 #include "tdbInt.h"
 
-// int tdbTxnBegin(TENV *pEnv) {
-//   // TODO
-//   return 0;
-// }
+int tdbTxnOpen(TXN *pTxn, int64_t txnid, void *(*xMalloc)(void *, size_t), void (*xFree)(void *, void *), void *xArg,
+               int flags) {
+  // not support read-committed version at the moment
+  ASSERT(flags == 0 || flags == TDB_TXN_WRITE | TDB_TXN_READ_UNCOMMITTED);
 
-// int tdbTxnCommit(TENV *pEnv) {
-//   SPager *pPager = NULL;
-//   int     ret;
+  pTxn->flags = flags;
+  pTxn->txnId = txnid;
+  pTxn->xMalloc = xMalloc;
+  pTxn->xFree = xFree;
+  pTxn->xArg = xArg;
+  return 0;
+}
 
-//   for (;;) {
-//     break;
-//     ret = tdbPagerCommit(pPager);
-//     if (ret < 0) {
-//       ASSERT(0);
-//       return -1;
-//     }
-//   }
-
-//   // TODO
-//   return 0;
-// }
-
-// int tdbTxnRollback(TENV *pEnv) {
-//   // TODO
-//   return 0;
-// }
\ No newline at end of file
+int tdbTxnClose(TXN *pTxn) { return 0; }
\ No newline at end of file
diff --git a/source/libs/tdb/src/inc/tdbBtree.h b/source/libs/tdb/src/inc/tdbBtree.h
index 2eba5f4f1a..3cdd30c7b5 100644
--- a/source/libs/tdb/src/inc/tdbBtree.h
+++ b/source/libs/tdb/src/inc/tdbBtree.h
@@ -35,17 +35,18 @@ struct SBTC {
   int     idx;
   int     idxStack[BTREE_MAX_DEPTH + 1];
   SPage  *pgStack[BTREE_MAX_DEPTH + 1];
+  TXN    *pTxn;
 };
 
 // SBTree
 int tdbBtreeOpen(int keyLen, int valLen, SPager *pFile, FKeyComparator kcmpr, SBTree **ppBt);
 int tdbBtreeClose(SBTree *pBt);
-int tdbBtreeInsert(SBTree *pBt, const void *pKey, int kLen, const void *pVal, int vLen);
+int tdbBtreeInsert(SBTree *pBt, const void *pKey, int kLen, const void *pVal, int vLen, TXN *pTxn);
 int tdbBtreeGet(SBTree *pBt, const void *pKey, int kLen, void **ppVal, int *vLen);
 int tdbBtreePGet(SBTree *pBt, const void *pKey, int kLen, void **ppKey, int *pkLen, void **ppVal, int *vLen);
 
 // SBTC
-int tdbBtcOpen(SBTC *pCur, SBTree *pBt);
+int tdbBtcOpen(SBTC *pBtc, SBTree *pBt, TXN *pTxn);
 int tdbBtcMoveToFirst(SBTC *pBtc);
 int tdbBtcMoveToLast(SBTC *pBtc);
 int tdbBtreeNext(SBTC *pBtc, void **ppKey, int *kLen, void **ppVal, int *vLen);
diff --git a/source/libs/tdb/src/inc/tdbDb.h b/source/libs/tdb/src/inc/tdbDb.h
index e60371c734..f9ff378633 100644
--- a/source/libs/tdb/src/inc/tdbDb.h
+++ b/source/libs/tdb/src/inc/tdbDb.h
@@ -27,7 +27,7 @@ typedef struct STDBC TDBC;
 int tdbDbOpen(const char *fname, int keyLen, int valLen, FKeyComparator keyCmprFn, TENV *pEnv, TDB **ppDb);
 int tdbDbClose(TDB *pDb);
 int tdbDbDrop(TDB *pDb);
-int tdbDbInsert(TDB *pDb, const void *pKey, int keyLen, const void *pVal, int valLen);
+int tdbDbInsert(TDB *pDb, const void *pKey, int keyLen, const void *pVal, int valLen, TXN *pTxn);
 int tdbDbGet(TDB *pDb, const void *pKey, int kLen, void **ppVal, int *vLen);
 int tdbDbPGet(TDB *pDb, const void *pKey, int kLen, void **ppKey, int *pkLen, void **ppVal, int *vLen);
 
diff --git a/source/libs/tdb/src/inc/tdbInt.h b/source/libs/tdb/src/inc/tdbInt.h
index 2a196b2cc2..b85b7174bc 100644
--- a/source/libs/tdb/src/inc/tdbInt.h
+++ b/source/libs/tdb/src/inc/tdbInt.h
@@ -111,13 +111,21 @@ typedef struct SPager  SPager;
 typedef struct SPCache SPCache;
 typedef struct SPage   SPage;
 
+// transaction
+#define TDB_TXN_WRITE            0x1
+#define TDB_TXN_READ_UNCOMMITTED 0x2
 typedef struct STxn {
-  u64 txnId;
+  int flags;
+  i64 txnId;
   void *(*xMalloc)(void *, size_t);
   void (*xFree)(void *, void *);
   void *xArg;
 } TXN;
 
+#define TDB_TXN_IS_WRITE(PTXN)            ((PTXN)->flags & TDB_TXN_WRITE)
+#define TDB_TXN_IS_READ(PTXN)             (!TDB_TXN_IS_WRITE(PTXN))
+#define TDB_TXN_IS_READ_UNCOMMITTED(PTXN) ((PTXN)->flags & TDB_TXN_READ_UNCOMMITTED)
+
 #include "tdbOs.h"
 
 #include "tdbUtil.h"
diff --git a/source/libs/tdb/src/inc/tdbPCache.h b/source/libs/tdb/src/inc/tdbPCache.h
index 5b24a53591..a77cf6248c 100644
--- a/source/libs/tdb/src/inc/tdbPCache.h
+++ b/source/libs/tdb/src/inc/tdbPCache.h
@@ -49,7 +49,7 @@ extern "C" {
 int    tdbPCacheOpen(int pageSize, int cacheSize, SPCache **ppCache);
 int    tdbPCacheClose(SPCache *pCache);
 SPage *tdbPCacheFetch(SPCache *pCache, const SPgid *pPgid, TXN *pTxn);
-void   tdbPCacheRelease(SPCache *pCache, SPage *pPage);
+void   tdbPCacheRelease(SPCache *pCache, SPage *pPage, TXN *pTxn);
 int    tdbPCacheGetPageSize(SPCache *pCache);
 
 #ifdef __cplusplus
diff --git a/source/libs/tdb/src/inc/tdbPager.h b/source/libs/tdb/src/inc/tdbPager.h
index cc8860ae7a..ca196785d0 100644
--- a/source/libs/tdb/src/inc/tdbPager.h
+++ b/source/libs/tdb/src/inc/tdbPager.h
@@ -42,9 +42,12 @@ int  tdbPagerOpenDB(SPager *pPager, SPgno *ppgno, bool toCreate);
 int  tdbPagerWrite(SPager *pPager, SPage *pPage);
 int  tdbPagerBegin(SPager *pPager, TXN *pTxn);
 int  tdbPagerCommit(SPager *pPager, TXN *pTxn);
-int  tdbPagerFetchPage(SPager *pPager, SPgno pgno, SPage **ppPage, int (*initPage)(SPage *, void *), void *arg);
-int  tdbPagerNewPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage)(SPage *, void *), void *arg);
-void tdbPagerReturnPage(SPager *pPager, SPage *pPage);
+int  tdbPagerFetchPage(SPager *pPager, SPgno pgno, SPage **ppPage, int (*initPage)(SPage *, void *), void *arg,
+                       TXN *pTxn);
+int  tdbPagerNewPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage)(SPage *, void *), void *arg,
+                     TXN *pTxn);
+void tdbPagerReturnPage(SPager *pPager, SPage *pPage, TXN *pTxn);
+int  tdbPagerAllocPage(SPager *pPager, SPgno *ppgno);
 
 #ifdef __cplusplus
 }
diff --git a/source/libs/tdb/src/inc/tdbTxn.h b/source/libs/tdb/src/inc/tdbTxn.h
index 13303bffb6..ed9784614c 100644
--- a/source/libs/tdb/src/inc/tdbTxn.h
+++ b/source/libs/tdb/src/inc/tdbTxn.h
@@ -20,6 +20,10 @@
 extern "C" {
 #endif
 
+int tdbTxnOpen(TXN *pTxn, int64_t txnid, void *(*xMalloc)(void *, size_t), void (*xFree)(void *, void *), void *xArg,
+               int flags);
+int tdbTxnClose(TXN *pTxn);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/source/libs/tdb/test/tdbTest.cpp b/source/libs/tdb/test/tdbTest.cpp
index 7e5cfeff42..904ad64fe7 100644
--- a/source/libs/tdb/test/tdbTest.cpp
+++ b/source/libs/tdb/test/tdbTest.cpp
@@ -19,7 +19,7 @@ static SPoolMem *openPool() {
   return pPool;
 }
 
-static void closePool(SPoolMem *pPool) {
+static void clearPool(SPoolMem *pPool) {
   SPoolMem *pMem;
 
   do {
@@ -35,13 +35,14 @@ static void closePool(SPoolMem *pPool) {
   } while (1);
 
   assert(pPool->size == 0);
+}
 
+static void closePool(SPoolMem *pPool) {
+  clearPool(pPool);
   tdbOsFree(pPool);
 }
 
-#define clearPool closePool
-
-static void *poolMalloc(void *arg, int size) {
+static void *poolMalloc(void *arg, size_t size) {
   void     *ptr = NULL;
   SPoolMem *pPool = (SPoolMem *)arg;
   SPoolMem *pMem;
@@ -118,7 +119,8 @@ TEST(tdb_test, simple_test) {
   TENV          *pEnv;
   TDB           *pDb;
   FKeyComparator compFunc;
-  int            nData = 50000000;
+  int            nData = 10000000;
+  TXN            txn;
 
   // Open Env
   ret = tdbEnvOpen("tdb", 4096, 64, &pEnv);
@@ -130,25 +132,44 @@ TEST(tdb_test, simple_test) {
   GTEST_ASSERT_EQ(ret, 0);
 
   {
-    char key[64];
-    char val[64];
+    char      key[64];
+    char      val[64];
+    int64_t   poolLimit = 4096;  // 1M pool limit
+    int64_t   txnid = 0;
+    SPoolMem *pPool;
 
-    {  // Insert some data
-      for (int i = 1; i <= nData;) {
-        tdbBegin(pEnv, NULL);
+    // open the pool
+    pPool = openPool();
 
-        for (int k = 0; k < 2000; k++) {
-          sprintf(key, "key%d", i);
-          sprintf(val, "value%d", i);
-          ret = tdbDbInsert(pDb, key, strlen(key), val, strlen(val));
-          GTEST_ASSERT_EQ(ret, 0);
-          i++;
-        }
+    // start a transaction
+    txnid++;
+    tdbTxnOpen(&txn, txnid, poolMalloc, poolFree, pPool, TDB_TXN_WRITE | TDB_TXN_READ_UNCOMMITTED);
+    tdbBegin(pEnv, &txn);
 
-        tdbCommit(pEnv, NULL);
+    for (int iData = 1; iData <= nData; iData++) {
+      sprintf(key, "key%d", iData);
+      sprintf(val, "value%d", iData);
+      ret = tdbDbInsert(pDb, key, strlen(key), val, strlen(val), &txn);
+      GTEST_ASSERT_EQ(ret, 0);
+
+      // if pool is full, commit the transaction and start a new one
+      if (pPool->size >= poolLimit) {
+        // commit current transaction
+        tdbCommit(pEnv, &txn);
+        tdbTxnClose(&txn);
+
+        // start a new transaction
+        clearPool(pPool);
+        txnid++;
+        tdbTxnOpen(&txn, txnid, poolMalloc, poolFree, pPool, TDB_TXN_WRITE | TDB_TXN_READ_UNCOMMITTED);
+        tdbBegin(pEnv, &txn);
       }
     }
 
+    // commit the transaction
+    tdbCommit(pEnv, &txn);
+    tdbTxnClose(&txn);
+
     {  // Query the data
       void *pVal = NULL;
       int   vLen;

From 27a3a1d782ae6e7fd9d21d5875b3763643420e57 Mon Sep 17 00:00:00 2001
From: dapan1121 <wpan@taosdata.com>
Date: Sat, 2 Apr 2022 19:31:52 +0800
Subject: [PATCH 06/25] feature/qnode

---
 include/common/tmsg.h                   |  12 +-
 include/libs/command/command.h          |   8 +-
 include/libs/nodes/nodes.h              |   1 +
 include/util/tlockfree.h                |   1 +
 source/common/src/tmsg.c                |  42 +++
 source/libs/command/inc/commandInt.h    |  76 +++--
 source/libs/command/src/explain.c       | 357 ++++++++++++++++++++----
 source/libs/executor/inc/executorimpl.h |   4 +-
 source/libs/executor/src/executorimpl.c |  18 +-
 source/libs/nodes/src/nodesUtilFuncs.c  |  22 ++
 source/libs/qworker/inc/qworkerMsg.h    |   3 +-
 source/libs/qworker/src/qworker.c       |   5 +-
 source/libs/qworker/src/qworkerMsg.c    |  21 ++
 source/libs/scheduler/src/scheduler.c   | 169 ++++++-----
 source/util/src/tlockfree.c             |  15 +
 15 files changed, 597 insertions(+), 157 deletions(-)

diff --git a/include/common/tmsg.h b/include/common/tmsg.h
index 4cfcf0e888..27b9a11625 100644
--- a/include/common/tmsg.h
+++ b/include/common/tmsg.h
@@ -915,11 +915,19 @@ typedef struct {
   char    data[];
 } SRetrieveMetaTableRsp;
 
+typedef struct SExplainExecInfo {
+  uint64_t startupCost;
+  uint64_t totalCost;
+  uint64_t numOfRows;
+} SExplainExecInfo;
+
 typedef struct {
-  int32_t numOfItems;
-  char    data[];
+  int32_t           numOfPlans;
+  SExplainExecInfo *subplanInfo;
 } SExplainRsp;
 
+int32_t tSerializeSExplainRsp(void* buf, int32_t bufLen, SExplainRsp* pRsp);
+int32_t tDeserializeSExplainRsp(void* buf, int32_t bufLen, SExplainRsp* pRsp);
 
 typedef struct {
   char    fqdn[TSDB_FQDN_LEN];  // end point, hostname:port
diff --git a/include/libs/command/command.h b/include/libs/command/command.h
index 31e370ac5c..fa229d7265 100644
--- a/include/libs/command/command.h
+++ b/include/libs/command/command.h
@@ -19,17 +19,11 @@
 
 typedef struct SExplainCtx SExplainCtx;
 
-typedef struct SExplainExecInfo {
-  uint64_t startupCost;
-  uint64_t totalCost;
-  uint64_t numOfRows;
-} SExplainExecInfo;
-
 int32_t qExecCommand(SNode* pStmt, SRetrieveTableRsp** pRsp);
 
 int32_t qExecStaticExplain(SQueryPlan *pDag, SRetrieveTableRsp **pRsp);
 int32_t qExecExplainBegin(SQueryPlan *pDag, SExplainCtx **pCtx, int32_t startTs);
-int32_t qExecExplainEnd(SExplainCtx *pCtx);
+int32_t qExecExplainEnd(SExplainCtx *pCtx, SRetrieveTableRsp **pRsp);
 int32_t qExplainUpdateExecInfo(SExplainCtx        *pCtx, SExplainRsp *pRspMsg, int32_t groupId, SRetrieveTableRsp **pRsp);
 void    qExplainFreeCtx(SExplainCtx *pCtx);
 
diff --git a/include/libs/nodes/nodes.h b/include/libs/nodes/nodes.h
index 2979040c07..1c229fabb7 100644
--- a/include/libs/nodes/nodes.h
+++ b/include/libs/nodes/nodes.h
@@ -217,6 +217,7 @@ int32_t nodesStringToList(const char* pStr, SNodeList** pList);
 
 int32_t nodesNodeToSQL(SNode *pNode, char *buf, int32_t bufSize, int32_t *len);
 char *nodesGetNameFromColumnNode(SNode *pNode);
+int32_t nodesGetOutputNumFromSlotList(SNodeList* pSlots);
 
 #ifdef __cplusplus
 }
diff --git a/include/util/tlockfree.h b/include/util/tlockfree.h
index e49e89c309..44e43f81cf 100644
--- a/include/util/tlockfree.h
+++ b/include/util/tlockfree.h
@@ -76,6 +76,7 @@ void taosWLockLatch(SRWLatch *pLatch);
 void taosWUnLockLatch(SRWLatch *pLatch);
 void taosRLockLatch(SRWLatch *pLatch);
 void taosRUnLockLatch(SRWLatch *pLatch);
+int32_t taosWTryLockLatch(SRWLatch *pLatch);
 
 // copy on read
 #define taosCorBeginRead(x)                     \
diff --git a/source/common/src/tmsg.c b/source/common/src/tmsg.c
index c85184ffba..0b25d7286f 100644
--- a/source/common/src/tmsg.c
+++ b/source/common/src/tmsg.c
@@ -2787,6 +2787,48 @@ int32_t tDecodeSMqCMCommitOffsetReq(SCoder *decoder, SMqCMCommitOffsetReq *pReq)
   return 0;
 }
 
+int32_t tSerializeSExplainRsp(void* buf, int32_t bufLen, SExplainRsp* pRsp) {
+  SCoder encoder = {0};
+  tCoderInit(&encoder, TD_LITTLE_ENDIAN, buf, bufLen, TD_ENCODER);
+
+  if (tStartEncode(&encoder) < 0) return -1;
+  if (tEncodeI32(&encoder, pRsp->numOfPlans) < 0) return -1;
+  for (int32_t i = 0; i < pRsp->numOfPlans; ++i) {
+    SExplainExecInfo *info = &pRsp->subplanInfo[i];
+    if (tEncodeU64(&encoder, info->startupCost) < 0) return -1;
+    if (tEncodeU64(&encoder, info->totalCost) < 0) return -1;
+    if (tEncodeU64(&encoder, info->numOfRows) < 0) return -1;
+  }
+
+  tEndEncode(&encoder);
+
+  int32_t tlen = encoder.pos;
+  tCoderClear(&encoder);
+  return tlen;
+}
+
+int32_t tDeserializeSExplainRsp(void* buf, int32_t bufLen, SExplainRsp* pRsp) {
+  SCoder decoder = {0};
+  tCoderInit(&decoder, TD_LITTLE_ENDIAN, buf, bufLen, TD_DECODER);
+
+  if (tStartDecode(&decoder) < 0) return -1;
+  if (tDecodeI32(&decoder, &pRsp->numOfPlans) < 0) return -1;
+  if (pRsp->numOfPlans > 0) {
+    pRsp->subplanInfo = taosMemoryMalloc(pRsp->numOfPlans * sizeof(SExplainExecInfo));
+    if (pRsp->subplanInfo == NULL) return -1;
+  }
+  for (int32_t i = 0; i < pRsp->numOfPlans; ++i) {
+    if (tDecodeU64(&decoder, &pRsp->subplanInfo[i].startupCost) < 0) return -1;
+    if (tDecodeU64(&decoder, &pRsp->subplanInfo[i].totalCost) < 0) return -1;
+    if (tDecodeU64(&decoder, &pRsp->subplanInfo[i].numOfRows) < 0) return -1;
+  }
+
+  tEndDecode(&decoder);
+
+  tCoderClear(&decoder);
+  return 0;
+}
+
 int32_t tSerializeSSchedulerHbReq(void *buf, int32_t bufLen, SSchedulerHbReq *pReq) {
   int32_t headLen = sizeof(SMsgHead);
   if (buf != NULL) {
diff --git a/source/libs/command/inc/commandInt.h b/source/libs/command/inc/commandInt.h
index de0a7c433a..7e9ea12ea4 100644
--- a/source/libs/command/inc/commandInt.h
+++ b/source/libs/command/inc/commandInt.h
@@ -26,39 +26,53 @@ extern "C" {
 #define EXPLAIN_MAX_GROUP_NUM 100
 
 //newline area
-#define EXPLAIN_TAG_SCAN_FORMAT "Tag Scan on %s columns=%d width=%d"
-#define EXPLAIN_TBL_SCAN_FORMAT "Table Scan on %s columns=%d width=%d"
-#define EXPLAIN_SYSTBL_SCAN_FORMAT "System Table Scan on %s columns=%d width=%d"
-#define EXPLAIN_PROJECTION_FORMAT "Projection columns=%d width=%d"
-#define EXPLAIN_JOIN_FORMAT "%s between %d tables width=%d"
-#define EXPLAIN_AGG_FORMAT "Aggragate functions=%d"
-#define EXPLAIN_EXCHANGE_FORMAT "Data Exchange %d:1 width=%d"
-#define EXPLAIN_SORT_FORMAT "Sort on %d Column(s) width=%d"
-#define EXPLAIN_INTERVAL_FORMAT "Interval on Column %s functions=%d interval=%" PRId64 "%c offset=%" PRId64 "%c sliding=%" PRId64 "%c width=%d"
-#define EXPLAIN_SESSION_FORMAT "Session gap=%" PRId64 " functions=%d width=%d"
+#define EXPLAIN_TAG_SCAN_FORMAT "Tag Scan on %s"
+#define EXPLAIN_TBL_SCAN_FORMAT "Table Scan on %s"
+#define EXPLAIN_SYSTBL_SCAN_FORMAT "System Table Scan on %s"
+#define EXPLAIN_PROJECTION_FORMAT "Projection"
+#define EXPLAIN_JOIN_FORMAT "%s"
+#define EXPLAIN_AGG_FORMAT "Aggragate"
+#define EXPLAIN_EXCHANGE_FORMAT "Data Exchange %d:1"
+#define EXPLAIN_SORT_FORMAT "Sort"
+#define EXPLAIN_INTERVAL_FORMAT "Interval on Column %s"
+#define EXPLAIN_SESSION_FORMAT "Session"
 #define EXPLAIN_ORDER_FORMAT "Order: %s"
 #define EXPLAIN_FILTER_FORMAT "Filter: "
 #define EXPLAIN_FILL_FORMAT "Fill: %s"
 #define EXPLAIN_ON_CONDITIONS_FORMAT "Join Cond: "
 #define EXPLAIN_TIMERANGE_FORMAT "Time Range: [%" PRId64 ", %" PRId64 "]"
+#define EXPLAIN_OUTPUT_FORMAT "Output: "
+#define EXPLAIN_TIME_WINDOWS_FORMAT "Time Window: interval=%" PRId64 "%c offset=%" PRId64 "%c sliding=%" PRId64 "%c"
+#define EXPLAIN_WINDOW_FORMAT "Window: gap=%" PRId64
 
 //append area
-#define EXPLAIN_GROUPS_FORMAT " groups=%d"
-#define EXPLAIN_WIDTH_FORMAT " width=%d"
-#define EXPLAIN_LOOPS_FORMAT " loops=%d"
-#define EXPLAIN_REVERSE_FORMAT " reverse=%d"
+#define EXPLAIN_LEFT_PARENTHESIS_FORMAT " ("
+#define EXPLAIN_RIGHT_PARENTHESIS_FORMAT ")"
+#define EXPLAIN_BLANK_FORMAT " "
+#define EXPLAIN_COST_FORMAT "cost=%.2f..%.2f"
+#define EXPLAIN_ROWS_FORMAT "rows=%" PRIu64
+#define EXPLAIN_COLUMNS_FORMAT "columns=%d"
+#define EXPLAIN_WIDTH_FORMAT "width=%d"
+#define EXPLAIN_GROUPS_FORMAT "groups=%d"
+#define EXPLAIN_WIDTH_FORMAT "width=%d"
+#define EXPLAIN_LOOPS_FORMAT "loops=%d"
+#define EXPLAIN_REVERSE_FORMAT "reverse=%d"
+#define EXPLAIN_FUNCTIONS_FORMAT "functions=%d"
 
 typedef struct SExplainGroup {
   int32_t   nodeNum;
+  int32_t   physiPlanExecNum;
+  int32_t   physiPlanNum;
+  int32_t   physiPlanExecIdx;
   SRWLatch  lock;
   SSubplan *plan;
-  SArray   *execInfo;
+  SArray   *nodeExecInfo;      //Array<SExplainRsp>
 } SExplainGroup;
 
 typedef struct SExplainResNode {
-  SNodeList*  pChildren;
-  SPhysiNode* pNode;
-  void*       pExecInfo;
+  SNodeList*        pChildren;
+  SPhysiNode*       pNode;
+  SArray*           pExecInfo; // Array<SExplainExecInfo>
 } SExplainResNode;
 
 typedef struct SQueryExplainRowInfo {
@@ -68,19 +82,21 @@ typedef struct SQueryExplainRowInfo {
 } SQueryExplainRowInfo;
 
 typedef struct SExplainCtx {
-  double    ratio;
-  bool      verbose;
+  EExplainMode mode;
+  double       ratio;
+  bool         verbose;
 
-  int32_t   rootGroupId;
-  int32_t   dataSize;
-  bool      execDone;
-  int64_t   reqStartTs;
-  int64_t   jobStartTs;
-  int64_t   jobDoneTs;
-  char     *tbuf;
-  SArray   *rows;
-  int32_t   groupDoneNum;
-  SHashObj *groupHash;
+  SRWLatch     lock;
+  int32_t      rootGroupId;
+  int32_t      dataSize;
+  bool         execDone;
+  int64_t      reqStartTs;
+  int64_t      jobStartTs;
+  int64_t      jobDoneTs;
+  char        *tbuf;
+  SArray      *rows;
+  int32_t      groupDoneNum;
+  SHashObj    *groupHash;     // Hash<SExplainGroup>
 } SExplainCtx;
 
 #define EXPLAIN_ORDER_STRING(_order) ((TSDB_ORDER_ASC == _order) ? "Ascending" : "Descending")
diff --git a/source/libs/command/src/explain.c b/source/libs/command/src/explain.c
index 4bf354dee5..c8502388ba 100644
--- a/source/libs/command/src/explain.c
+++ b/source/libs/command/src/explain.c
@@ -17,7 +17,7 @@
 #include "plannodes.h"
 #include "commandInt.h"
 
-int32_t qExplainGenerateResNode(SPhysiNode *pNode, void *pExecInfo, SExplainResNode **pRes);
+int32_t qExplainGenerateResNode(SPhysiNode *pNode, SExplainGroup *group, SExplainResNode **pRes);
 int32_t qExplainAppendGroupResRows(void *pCtx, int32_t groupId, int32_t level);
 
 
@@ -48,12 +48,30 @@ void qExplainFreeCtx(SExplainCtx *pCtx) {
     taosMemoryFreeClear(row->buf);
   }
 
+  if (EXPLAIN_MODE_ANALYZE == pCtx->mode && pCtx->groupHash) {
+    void *pIter = taosHashIterate(pCtx->groupHash, NULL);
+    while (pIter) {
+      SExplainGroup *group = (SExplainGroup *)pIter;
+      if (NULL == group->nodeExecInfo) {
+        continue;
+      }
+      
+      int32_t num = taosArrayGetSize(group->nodeExecInfo);
+      for (int32_t i = 0; i < num; ++i) {
+        SExplainRsp *rsp = taosArrayGet(group->nodeExecInfo, i);
+        taosMemoryFreeClear(rsp->subplanInfo);
+      }
+    
+      pIter = taosHashIterate(pCtx->groupHash, pIter);
+    }
+  }
+  
   taosHashCleanup(pCtx->groupHash);
   taosArrayDestroy(pCtx->rows);
   taosMemoryFree(pCtx);
 }
 
-int32_t qExplainInitCtx(SExplainCtx **pCtx, SHashObj *groupHash, bool verbose, double ratio) {
+int32_t qExplainInitCtx(SExplainCtx **pCtx, SHashObj *groupHash, bool verbose, double ratio, EExplainMode mode) {
   int32_t code = 0;
   SExplainCtx *ctx = taosMemoryCalloc(1, sizeof(SExplainCtx));
   if (NULL == ctx) {
@@ -73,6 +91,7 @@ int32_t qExplainInitCtx(SExplainCtx **pCtx, SHashObj *groupHash, bool verbose, d
     QRY_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
   }
 
+  ctx->mode = mode;
   ctx->verbose = verbose;
   ctx->ratio = ratio;
   ctx->tbuf = tbuf;
@@ -92,7 +111,7 @@ _return:
   QRY_RET(code);
 }
 
-int32_t qExplainGenerateResChildren(SPhysiNode *pNode, void *pExecInfo, SNodeList **pChildren) {
+int32_t qExplainGenerateResChildren(SPhysiNode *pNode, SExplainGroup *group, SNodeList **pChildren) {
   int32_t tlen = 0;
   SNodeList *pPhysiChildren = NULL;
   
@@ -164,20 +183,43 @@ int32_t qExplainGenerateResChildren(SPhysiNode *pNode, void *pExecInfo, SNodeLis
   SNode* node = NULL;
   SExplainResNode *pResNode = NULL;
   FOREACH(node, pPhysiChildren) {
-    QRY_ERR_RET(qExplainGenerateResNode((SPhysiNode *)node, pExecInfo, &pResNode));
+    QRY_ERR_RET(qExplainGenerateResNode((SPhysiNode *)node, group, &pResNode));
     QRY_ERR_RET(nodesListAppend(*pChildren, pResNode));
   }
 
   return TSDB_CODE_SUCCESS;
 }
 
-int32_t qExplainGenerateResNode(SPhysiNode *pNode, void *pExecInfo, SExplainResNode **pResNode) {
+int32_t qExplainGenerateResNodeExecInfo(SArray **pExecInfo, SExplainGroup *group) {
+  *pExecInfo = taosArrayInit(group->nodeNum, sizeof(SExplainExecInfo));
+  if (NULL == (*pExecInfo)) {
+    qError("taosArrayInit %d explainExecInfo failed", group->nodeNum);
+    return TSDB_CODE_QRY_OUT_OF_MEMORY;
+  }
+
+  SExplainRsp *rsp = NULL;  
+  for (int32_t i = 0; i < group->nodeNum; ++i) {
+    rsp = taosArrayGet(group->nodeExecInfo, i);
+    if (group->physiPlanExecIdx >= rsp->numOfPlans) {
+      qError("physiPlanIdx %d exceed plan num %d", group->physiPlanExecIdx, rsp->numOfPlans);
+      return TSDB_CODE_QRY_APP_ERROR;
+    }
+    
+    taosArrayPush(*pExecInfo, rsp->subplanInfo + group->physiPlanExecIdx);
+  }
+
+  ++group->physiPlanExecIdx;
+
+  return TSDB_CODE_SUCCESS;
+}
+
+int32_t qExplainGenerateResNode(SPhysiNode *pNode, SExplainGroup *group, SExplainResNode **pResNode) {
   if (NULL == pNode) {
     *pResNode = NULL;
     qError("physical node is NULL");
     return TSDB_CODE_QRY_APP_ERROR;
   }
-
+  
   SExplainResNode *resNode = taosMemoryCalloc(1, sizeof(SExplainResNode));
   if (NULL == resNode) {
     qError("calloc SPhysiNodeExplainRes failed");
@@ -186,8 +228,11 @@ int32_t qExplainGenerateResNode(SPhysiNode *pNode, void *pExecInfo, SExplainResN
 
   int32_t code = 0;
   resNode->pNode = pNode;
-  resNode->pExecInfo = pExecInfo;
-  QRY_ERR_JRET(qExplainGenerateResChildren(pNode, pExecInfo, &resNode->pChildren));
+  QRY_ERR_JRET(qExplainGenerateResNodeExecInfo(&resNode->pExecInfo, group));
+  
+  QRY_ERR_JRET(qExplainGenerateResChildren(pNode, group, &resNode->pChildren));
+
+  ++group->physiPlanNum;
   
   *pResNode = resNode;
 
@@ -200,7 +245,7 @@ _return:
   QRY_RET(code);
 }
 
-int32_t qExplainBufAppendExecInfo(void *pExecInfo, char *tbuf, int32_t *len) {
+int32_t qExplainBufAppendExecInfo(SArray *pExecInfo, char *tbuf, int32_t *len) {
   int32_t tlen = *len;
   
   EXPLAIN_ROW_APPEND("(exec info here)");
@@ -247,18 +292,33 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
   switch (pNode->type) {
     case QUERY_NODE_PHYSICAL_PLAN_TAG_SCAN: {
       STagScanPhysiNode *pTagScanNode = (STagScanPhysiNode *)pNode;
-      EXPLAIN_ROW_NEW(level, EXPLAIN_TAG_SCAN_FORMAT, pTagScanNode->tableName.tname, pTagScanNode->pScanCols->length, pTagScanNode->node.pOutputDataBlockDesc->outputRowSize);
+      EXPLAIN_ROW_NEW(level, EXPLAIN_TAG_SCAN_FORMAT, pTagScanNode->tableName.tname);
+      EXPLAIN_ROW_APPEND(EXPLAIN_LEFT_PARENTHESIS_FORMAT);
       if (pResNode->pExecInfo) {
         QRY_ERR_RET(qExplainBufAppendExecInfo(pResNode->pExecInfo, tbuf, &tlen));
-      }
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      }      
+      EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, pTagScanNode->pScanCols->length);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pTagScanNode->node.pOutputDataBlockDesc->totalRowSize);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
       EXPLAIN_ROW_APPEND(EXPLAIN_LOOPS_FORMAT, pTagScanNode->count);
       if (pTagScanNode->reverse) {
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
         EXPLAIN_ROW_APPEND(EXPLAIN_REVERSE_FORMAT, pTagScanNode->reverse);
       }
+      EXPLAIN_ROW_APPEND(EXPLAIN_RIGHT_PARENTHESIS_FORMAT);
       EXPLAIN_ROW_END();
       QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level));
 
       if (verbose) {
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_OUTPUT_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, nodesGetOutputNumFromSlotList(pTagScanNode->node.pOutputDataBlockDesc->pSlots));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pTagScanNode->node.pOutputDataBlockDesc->outputRowSize);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+        
         EXPLAIN_ROW_NEW(level + 1, EXPLAIN_ORDER_FORMAT, EXPLAIN_ORDER_STRING(pTagScanNode->order));
         EXPLAIN_ROW_END();
         QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
@@ -268,18 +328,33 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
     case QUERY_NODE_PHYSICAL_PLAN_TABLE_SEQ_SCAN:
     case QUERY_NODE_PHYSICAL_PLAN_TABLE_SCAN:{
       STableScanPhysiNode *pTblScanNode = (STableScanPhysiNode *)pNode;
-      EXPLAIN_ROW_NEW(level, EXPLAIN_TBL_SCAN_FORMAT, pTblScanNode->scan.tableName.tname, pTblScanNode->scan.pScanCols->length, pTblScanNode->scan.node.pOutputDataBlockDesc->outputRowSize);
+      EXPLAIN_ROW_NEW(level, EXPLAIN_TBL_SCAN_FORMAT, pTblScanNode->scan.tableName.tname);
+      EXPLAIN_ROW_APPEND(EXPLAIN_LEFT_PARENTHESIS_FORMAT);
       if (pResNode->pExecInfo) {
         QRY_ERR_RET(qExplainBufAppendExecInfo(pResNode->pExecInfo, tbuf, &tlen));
-      }
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      }      
+      EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, pTblScanNode->scan.pScanCols->length);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pTblScanNode->scan.node.pOutputDataBlockDesc->totalRowSize);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
       EXPLAIN_ROW_APPEND(EXPLAIN_LOOPS_FORMAT, pTblScanNode->scan.count);
       if (pTblScanNode->scan.reverse) {
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
         EXPLAIN_ROW_APPEND(EXPLAIN_REVERSE_FORMAT, pTblScanNode->scan.reverse);
       }
+      EXPLAIN_ROW_APPEND(EXPLAIN_RIGHT_PARENTHESIS_FORMAT);      
       EXPLAIN_ROW_END();
       QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level));
 
       if (verbose) {      
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_OUTPUT_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, nodesGetOutputNumFromSlotList(pTblScanNode->scan.node.pOutputDataBlockDesc->pSlots));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pTblScanNode->scan.node.pOutputDataBlockDesc->outputRowSize);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+
         EXPLAIN_ROW_NEW(level + 1, EXPLAIN_ORDER_FORMAT, EXPLAIN_ORDER_STRING(pTblScanNode->scan.order));
         EXPLAIN_ROW_END();
         QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
@@ -299,18 +374,33 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
     }
     case QUERY_NODE_PHYSICAL_PLAN_SYSTABLE_SCAN:{
       SSystemTableScanPhysiNode *pSTblScanNode = (SSystemTableScanPhysiNode *)pNode;
-      EXPLAIN_ROW_NEW(level, EXPLAIN_SYSTBL_SCAN_FORMAT, pSTblScanNode->scan.tableName.tname, pSTblScanNode->scan.pScanCols->length, pSTblScanNode->scan.node.pOutputDataBlockDesc->outputRowSize);
+      EXPLAIN_ROW_NEW(level, EXPLAIN_SYSTBL_SCAN_FORMAT, pSTblScanNode->scan.tableName.tname);
+      EXPLAIN_ROW_APPEND(EXPLAIN_LEFT_PARENTHESIS_FORMAT);
       if (pResNode->pExecInfo) {
         QRY_ERR_RET(qExplainBufAppendExecInfo(pResNode->pExecInfo, tbuf, &tlen));
-      }
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      }      
+      EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, pSTblScanNode->scan.pScanCols->length);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pSTblScanNode->scan.node.pOutputDataBlockDesc->totalRowSize);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
       EXPLAIN_ROW_APPEND(EXPLAIN_LOOPS_FORMAT, pSTblScanNode->scan.count);
       if (pSTblScanNode->scan.reverse) {
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
         EXPLAIN_ROW_APPEND(EXPLAIN_REVERSE_FORMAT, pSTblScanNode->scan.reverse);
       }
+      EXPLAIN_ROW_APPEND(EXPLAIN_RIGHT_PARENTHESIS_FORMAT);      
       EXPLAIN_ROW_END();
       QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level));
 
       if (verbose) {            
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_OUTPUT_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, nodesGetOutputNumFromSlotList(pSTblScanNode->scan.node.pOutputDataBlockDesc->pSlots));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pSTblScanNode->scan.node.pOutputDataBlockDesc->outputRowSize);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+
         EXPLAIN_ROW_NEW(level + 1, EXPLAIN_ORDER_FORMAT, EXPLAIN_ORDER_STRING(pSTblScanNode->scan.order));
         EXPLAIN_ROW_END();
         QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));   
@@ -327,14 +417,27 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
     }
     case QUERY_NODE_PHYSICAL_PLAN_PROJECT:{
       SProjectPhysiNode *pPrjNode = (SProjectPhysiNode *)pNode;
-      EXPLAIN_ROW_NEW(level, EXPLAIN_PROJECTION_FORMAT, pPrjNode->pProjections->length, pPrjNode->node.pOutputDataBlockDesc->outputRowSize);
+      EXPLAIN_ROW_NEW(level, EXPLAIN_PROJECTION_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_LEFT_PARENTHESIS_FORMAT);
       if (pResNode->pExecInfo) {
         QRY_ERR_RET(qExplainBufAppendExecInfo(pResNode->pExecInfo, tbuf, &tlen));
-      }
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      }      
+      EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, pPrjNode->pProjections->length);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pPrjNode->node.pOutputDataBlockDesc->totalRowSize);
+      EXPLAIN_ROW_APPEND(EXPLAIN_RIGHT_PARENTHESIS_FORMAT);      
       EXPLAIN_ROW_END();
       QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level));
 
       if (verbose) {
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_OUTPUT_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, nodesGetOutputNumFromSlotList(pPrjNode->node.pOutputDataBlockDesc->pSlots));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pPrjNode->node.pOutputDataBlockDesc->outputRowSize);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+      
         if (pPrjNode->node.pConditions) {
           EXPLAIN_ROW_NEW(level + 1, EXPLAIN_FILTER_FORMAT);      
           QRY_ERR_RET(nodesNodeToSQL(pPrjNode->node.pConditions, tbuf + VARSTR_HEADER_SIZE, TSDB_EXPLAIN_RESULT_ROW_SIZE, &tlen)); 
@@ -346,14 +449,27 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
     }
     case QUERY_NODE_PHYSICAL_PLAN_JOIN:{
       SJoinPhysiNode *pJoinNode = (SJoinPhysiNode *)pNode;
-      EXPLAIN_ROW_NEW(level, EXPLAIN_JOIN_FORMAT, EXPLAIN_JOIN_STRING(pJoinNode->joinType), pJoinNode->pTargets->length, pJoinNode->node.pOutputDataBlockDesc->outputRowSize);
+      EXPLAIN_ROW_NEW(level, EXPLAIN_JOIN_FORMAT, EXPLAIN_JOIN_STRING(pJoinNode->joinType));
+      EXPLAIN_ROW_APPEND(EXPLAIN_LEFT_PARENTHESIS_FORMAT);
       if (pResNode->pExecInfo) {
         QRY_ERR_RET(qExplainBufAppendExecInfo(pResNode->pExecInfo, tbuf, &tlen));
-      }
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      }      
+      EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, pJoinNode->pTargets->length);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pJoinNode->node.pOutputDataBlockDesc->totalRowSize);
+      EXPLAIN_ROW_APPEND(EXPLAIN_RIGHT_PARENTHESIS_FORMAT);      
       EXPLAIN_ROW_END();
       QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level));
 
       if (verbose) {
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_OUTPUT_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, nodesGetOutputNumFromSlotList(pJoinNode->node.pOutputDataBlockDesc->pSlots));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pJoinNode->node.pOutputDataBlockDesc->outputRowSize);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+
         if (pJoinNode->node.pConditions) {
           EXPLAIN_ROW_NEW(level + 1, EXPLAIN_FILTER_FORMAT);      
           QRY_ERR_RET(nodesNodeToSQL(pJoinNode->node.pConditions, tbuf + VARSTR_HEADER_SIZE, TSDB_EXPLAIN_RESULT_ROW_SIZE, &tlen)); 
@@ -370,19 +486,31 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
     }
     case QUERY_NODE_PHYSICAL_PLAN_AGG:{
       SAggPhysiNode *pAggNode = (SAggPhysiNode *)pNode;
-      EXPLAIN_ROW_NEW(level, EXPLAIN_AGG_FORMAT, pAggNode->pAggFuncs->length);
-      if (pAggNode->pGroupKeys) {
-        EXPLAIN_ROW_APPEND(EXPLAIN_GROUPS_FORMAT, pAggNode->pGroupKeys->length);
-      }
-      EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pAggNode->node.pOutputDataBlockDesc->outputRowSize);
-      
+      EXPLAIN_ROW_NEW(level, EXPLAIN_AGG_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_LEFT_PARENTHESIS_FORMAT);
       if (pResNode->pExecInfo) {
         QRY_ERR_RET(qExplainBufAppendExecInfo(pResNode->pExecInfo, tbuf, &tlen));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      }      
+      EXPLAIN_ROW_APPEND(EXPLAIN_FUNCTIONS_FORMAT, pAggNode->pAggFuncs->length);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pAggNode->node.pOutputDataBlockDesc->totalRowSize);
+      if (pAggNode->pGroupKeys) {
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_GROUPS_FORMAT, pAggNode->pGroupKeys->length);
       }
+      EXPLAIN_ROW_APPEND(EXPLAIN_RIGHT_PARENTHESIS_FORMAT);      
       EXPLAIN_ROW_END();
       QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level));
 
       if (verbose) {
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_OUTPUT_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, nodesGetOutputNumFromSlotList(pAggNode->node.pOutputDataBlockDesc->pSlots));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pAggNode->node.pOutputDataBlockDesc->outputRowSize);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+
         if (pAggNode->node.pConditions) {
           EXPLAIN_ROW_NEW(level + 1, EXPLAIN_FILTER_FORMAT);      
           QRY_ERR_RET(nodesNodeToSQL(pAggNode->node.pConditions, tbuf + VARSTR_HEADER_SIZE, TSDB_EXPLAIN_RESULT_ROW_SIZE, &tlen)); 
@@ -400,14 +528,25 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
         QRY_ERR_RET(TSDB_CODE_QRY_APP_ERROR);
       }
       
-      EXPLAIN_ROW_NEW(level, EXPLAIN_EXCHANGE_FORMAT, group->nodeNum, pExchNode->node.pOutputDataBlockDesc->outputRowSize);
+      EXPLAIN_ROW_NEW(level, EXPLAIN_EXCHANGE_FORMAT, group->nodeNum);
+      EXPLAIN_ROW_APPEND(EXPLAIN_LEFT_PARENTHESIS_FORMAT);
       if (pResNode->pExecInfo) {
         QRY_ERR_RET(qExplainBufAppendExecInfo(pResNode->pExecInfo, tbuf, &tlen));
-      }
+      }      
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pExchNode->node.pOutputDataBlockDesc->totalRowSize);
+      EXPLAIN_ROW_APPEND(EXPLAIN_RIGHT_PARENTHESIS_FORMAT);      
       EXPLAIN_ROW_END();
       QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level));
 
       if (verbose) {
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_OUTPUT_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, nodesGetOutputNumFromSlotList(pExchNode->node.pOutputDataBlockDesc->pSlots));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pExchNode->node.pOutputDataBlockDesc->outputRowSize);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+
         if (pExchNode->node.pConditions) {
           EXPLAIN_ROW_NEW(level + 1, EXPLAIN_FILTER_FORMAT);      
           QRY_ERR_RET(nodesNodeToSQL(pExchNode->node.pConditions, tbuf + VARSTR_HEADER_SIZE, TSDB_EXPLAIN_RESULT_ROW_SIZE, &tlen)); 
@@ -421,14 +560,28 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
     }
     case QUERY_NODE_PHYSICAL_PLAN_SORT:{
       SSortPhysiNode *pSortNode = (SSortPhysiNode *)pNode;
-      EXPLAIN_ROW_NEW(level, EXPLAIN_SORT_FORMAT, pSortNode->pSortKeys->length, pSortNode->node.pOutputDataBlockDesc->outputRowSize);
+      EXPLAIN_ROW_NEW(level, EXPLAIN_SORT_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_LEFT_PARENTHESIS_FORMAT);
       if (pResNode->pExecInfo) {
         QRY_ERR_RET(qExplainBufAppendExecInfo(pResNode->pExecInfo, tbuf, &tlen));
-      }
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      }      
+      EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, pSortNode->pSortKeys->length);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pSortNode->node.pOutputDataBlockDesc->totalRowSize);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_RIGHT_PARENTHESIS_FORMAT);
       EXPLAIN_ROW_END();
       QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level));
 
       if (verbose) {
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_OUTPUT_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, nodesGetOutputNumFromSlotList(pSortNode->node.pOutputDataBlockDesc->pSlots));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pSortNode->node.pOutputDataBlockDesc->outputRowSize);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+
         if (pSortNode->node.pConditions) {
           EXPLAIN_ROW_NEW(level + 1, EXPLAIN_FILTER_FORMAT);      
           QRY_ERR_RET(nodesNodeToSQL(pSortNode->node.pConditions, tbuf + VARSTR_HEADER_SIZE, TSDB_EXPLAIN_RESULT_ROW_SIZE, &tlen)); 
@@ -440,18 +593,32 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
     }
     case QUERY_NODE_PHYSICAL_PLAN_INTERVAL:{
       SIntervalPhysiNode *pIntNode = (SIntervalPhysiNode *)pNode;
-      EXPLAIN_ROW_NEW(level, EXPLAIN_INTERVAL_FORMAT, nodesGetNameFromColumnNode(pIntNode->pTspk), pIntNode->window.pFuncs->length, 
-        INVERAL_TIME_FROM_PRECISION_TO_UNIT(pIntNode->interval, pIntNode->intervalUnit, pIntNode->precision), pIntNode->intervalUnit, 
-        pIntNode->offset, getPrecisionUnit(pIntNode->precision), 
-        INVERAL_TIME_FROM_PRECISION_TO_UNIT(pIntNode->sliding, pIntNode->slidingUnit, pIntNode->precision), pIntNode->slidingUnit, 
-        pIntNode->window.node.pOutputDataBlockDesc->outputRowSize);
+      EXPLAIN_ROW_NEW(level, EXPLAIN_INTERVAL_FORMAT, nodesGetNameFromColumnNode(pIntNode->pTspk));
+      EXPLAIN_ROW_APPEND(EXPLAIN_LEFT_PARENTHESIS_FORMAT);
       if (pResNode->pExecInfo) {
         QRY_ERR_RET(qExplainBufAppendExecInfo(pResNode->pExecInfo, tbuf, &tlen));
-      }
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      }      
+      EXPLAIN_ROW_APPEND(EXPLAIN_FUNCTIONS_FORMAT, pIntNode->window.pFuncs->length);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pIntNode->window.node.pOutputDataBlockDesc->totalRowSize);
       EXPLAIN_ROW_END();
       QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level));
 
       if (verbose) {
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_OUTPUT_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, nodesGetOutputNumFromSlotList(pIntNode->window.node.pOutputDataBlockDesc->pSlots));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pIntNode->window.node.pOutputDataBlockDesc->outputRowSize);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_TIME_WINDOWS_FORMAT, INVERAL_TIME_FROM_PRECISION_TO_UNIT(pIntNode->interval, pIntNode->intervalUnit, pIntNode->precision),  
+              pIntNode->intervalUnit, pIntNode->offset, getPrecisionUnit(pIntNode->precision), 
+              INVERAL_TIME_FROM_PRECISION_TO_UNIT(pIntNode->sliding, pIntNode->slidingUnit, pIntNode->precision), pIntNode->slidingUnit);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+        
         if (pIntNode->pFill) {
           EXPLAIN_ROW_NEW(level + 1, EXPLAIN_FILL_FORMAT, getFillModeString(pIntNode->pFill->mode));      
           EXPLAIN_ROW_END();
@@ -468,18 +635,36 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
       break;
     }
     case QUERY_NODE_PHYSICAL_PLAN_SESSION_WINDOW:{
-      SSessionWinodwPhysiNode *pIntNode = (SSessionWinodwPhysiNode *)pNode;
-      EXPLAIN_ROW_NEW(level, EXPLAIN_SESSION_FORMAT, pIntNode->gap, pIntNode->window.pFuncs->length, pIntNode->window.node.pOutputDataBlockDesc->outputRowSize);
+      SSessionWinodwPhysiNode *pSessNode = (SSessionWinodwPhysiNode *)pNode;
+      EXPLAIN_ROW_NEW(level, EXPLAIN_SESSION_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_LEFT_PARENTHESIS_FORMAT);
       if (pResNode->pExecInfo) {
         QRY_ERR_RET(qExplainBufAppendExecInfo(pResNode->pExecInfo, tbuf, &tlen));
-      }
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      }      
+      EXPLAIN_ROW_APPEND(EXPLAIN_FUNCTIONS_FORMAT, pSessNode->window.pFuncs->length);
+      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+      EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pSessNode->window.node.pOutputDataBlockDesc->totalRowSize);
+      EXPLAIN_ROW_APPEND(EXPLAIN_RIGHT_PARENTHESIS_FORMAT);
       EXPLAIN_ROW_END();
       QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level));
 
+
       if (verbose) {
-        if (pIntNode->window.node.pConditions) {
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_OUTPUT_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_COLUMNS_FORMAT, nodesGetOutputNumFromSlotList(pSessNode->window.node.pOutputDataBlockDesc->pSlots));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
+        EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pSessNode->window.node.pOutputDataBlockDesc->outputRowSize);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+
+        EXPLAIN_ROW_NEW(level + 1, EXPLAIN_WINDOW_FORMAT, pSessNode->gap);
+        EXPLAIN_ROW_END();
+        QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+
+        if (pSessNode->window.node.pConditions) {
           EXPLAIN_ROW_NEW(level + 1, EXPLAIN_FILTER_FORMAT);      
-          QRY_ERR_RET(nodesNodeToSQL(pIntNode->window.node.pConditions, tbuf + VARSTR_HEADER_SIZE, TSDB_EXPLAIN_RESULT_ROW_SIZE, &tlen)); 
+          QRY_ERR_RET(nodesNodeToSQL(pSessNode->window.node.pConditions, tbuf + VARSTR_HEADER_SIZE, TSDB_EXPLAIN_RESULT_ROW_SIZE, &tlen)); 
           EXPLAIN_ROW_END();
           QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
         }
@@ -523,7 +708,12 @@ int32_t qExplainAppendGroupResRows(void *pCtx, int32_t groupId, int32_t level) {
     QRY_ERR_RET(TSDB_CODE_QRY_APP_ERROR);
   }
   
-  QRY_ERR_RET(qExplainGenerateResNode(group->plan->pNode, group->execInfo, &node));
+  QRY_ERR_RET(qExplainGenerateResNode(group->plan->pNode, group, &node));
+
+  if (group->physiPlanNum != group->physiPlanExecNum) {
+    qError("physiPlanNum %d mismatch with physiExecNum %d in group %d", group->physiPlanNum, group->physiPlanExecNum, groupId);
+    QRY_ERR_JRET(TSDB_CODE_QRY_APP_ERROR);
+  }
 
   QRY_ERR_JRET(qExplainResNodeToRows(node, ctx, level));
 
@@ -601,7 +791,7 @@ int32_t qExplainPrepareCtx(SQueryPlan *pDag, SExplainCtx **pCtx) {
     QRY_ERR_RET(TSDB_CODE_QRY_OUT_OF_MEMORY);
   }
 
-  QRY_ERR_JRET(qExplainInitCtx(&ctx, groupHash, pDag->explainInfo.verbose, pDag->explainInfo.ratio));
+  QRY_ERR_JRET(qExplainInitCtx(&ctx, groupHash, pDag->explainInfo.verbose, pDag->explainInfo.ratio, pDag->explainInfo.mode));
 
   for (int32_t i = 0; i < levelNum; ++i) {
     plans = (SNodeListNode *)nodesListGetNode(pDag->pSubplans, i);
@@ -625,7 +815,10 @@ int32_t qExplainPrepareCtx(SQueryPlan *pDag, SExplainCtx **pCtx) {
         continue;
       }
 
-      SExplainGroup group = {.nodeNum = 1, .plan = plan, .execInfo = NULL};
+      SExplainGroup group = {0};
+      group.nodeNum = 1;
+      group.plan = plan;
+      
       if (0 != taosHashPut(groupHash, &plan->id.groupId, sizeof(plan->id.groupId), &group, sizeof(group))) {
         qError("taosHashPut to explainGroupHash failed, taskIdx:%d", n);
         QRY_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
@@ -656,8 +849,70 @@ _return:
 }
 
 
-int32_t qExplainUpdateExecInfo(SExplainCtx        *pCtx, SExplainRsp *pRspMsg, int32_t groupId, SRetrieveTableRsp **pRsp) {
+int32_t qExplainGenerateRsp(SExplainCtx *pCtx, SRetrieveTableRsp **pRsp) {
+  QRY_ERR_RET(qExplainAppendGroupResRows(pCtx, pCtx->rootGroupId, 0));
+  
+  QRY_ERR_RET(qExplainGetRspFromCtx(pCtx, pRsp));
 
+  return TSDB_CODE_SUCCESS;
+}
+
+int32_t qExplainUpdateExecInfo(SExplainCtx        *pCtx, SExplainRsp *pRspMsg, int32_t groupId, SRetrieveTableRsp **pRsp) {
+  SExplainResNode *node = NULL;
+  int32_t code = 0;
+  bool groupDone = false;
+  SExplainCtx *ctx = (SExplainCtx *)pCtx;
+
+  SExplainGroup *group = taosHashGet(ctx->groupHash, &groupId, sizeof(groupId));
+  if (NULL == group) {
+    qError("group %d not in groupHash", groupId);
+    taosMemoryFreeClear(pRspMsg->subplanInfo);
+    QRY_ERR_RET(TSDB_CODE_QRY_APP_ERROR);
+  }
+
+  taosWLockLatch(&group->lock);
+  if (NULL == group->nodeExecInfo) {
+    group->nodeExecInfo = taosArrayInit(group->nodeNum, sizeof(SExplainRsp));
+    if (NULL == group->nodeExecInfo) {
+      qError("taosArrayInit %d explainExecInfo failed", group->nodeNum);
+      taosMemoryFreeClear(pRspMsg->subplanInfo);
+      taosWUnLockLatch(&group->lock);
+      
+      QRY_ERR_RET(TSDB_CODE_QRY_OUT_OF_MEMORY);
+    }
+
+    group->physiPlanExecNum = pRspMsg->numOfPlans;
+  } else if (taosArrayGetSize(group->nodeExecInfo) >= group->nodeNum) {
+    qError("group execInfo already full, size:%d, nodeNum:%d", (int32_t)taosArrayGetSize(group->nodeExecInfo), group->nodeNum);
+    taosMemoryFreeClear(pRspMsg->subplanInfo);
+    taosWUnLockLatch(&group->lock);
+    
+    QRY_ERR_RET(TSDB_CODE_QRY_APP_ERROR);
+  }
+
+  if (group->physiPlanExecNum != pRspMsg->numOfPlans) {
+    qError("physiPlanExecNum %d mismatch with others %d in group %d", pRspMsg->numOfPlans, group->physiPlanExecNum, groupId);
+    taosMemoryFreeClear(pRspMsg->subplanInfo);
+    taosWUnLockLatch(&group->lock);
+    
+    QRY_ERR_RET(TSDB_CODE_QRY_APP_ERROR);
+  }
+
+  taosArrayPush(group->nodeExecInfo, pRspMsg);
+  groupDone = (taosArrayGetSize(group->nodeExecInfo) >= group->nodeNum);
+  
+  taosWUnLockLatch(&group->lock);
+
+  if (groupDone && (taosHashGetSize(pCtx->groupHash) == atomic_add_fetch_32(&pCtx->groupDoneNum, 1))) {
+    if (atomic_load_8((int8_t *)&pCtx->execDone)) {
+      if (0 == taosWTryLockLatch(&pCtx->lock)) {
+        QRY_ERR_RET(qExplainGenerateRsp(pCtx, pRsp));
+        // LEAVE LOCK THERE
+      }
+    }
+  }
+
+  return TSDB_CODE_SUCCESS;
 }
 
 
@@ -666,10 +921,8 @@ int32_t qExecStaticExplain(SQueryPlan *pDag, SRetrieveTableRsp **pRsp) {
   SExplainCtx *pCtx = NULL;
 
   QRY_ERR_RET(qExplainPrepareCtx(pDag, &pCtx));
-
-  QRY_ERR_JRET(qExplainAppendGroupResRows(pCtx, pCtx->rootGroupId, 0));
   
-  QRY_ERR_JRET(qExplainGetRspFromCtx(pCtx, pRsp));
+  QRY_ERR_JRET(qExplainGenerateRsp(pCtx, pRsp));
   
 _return:
 
@@ -687,11 +940,19 @@ int32_t qExecExplainBegin(SQueryPlan *pDag, SExplainCtx **pCtx, int32_t startTs)
   return TSDB_CODE_SUCCESS;
 }
 
-int32_t qExecExplainEnd(SExplainCtx *pCtx) {
+int32_t qExecExplainEnd(SExplainCtx *pCtx, SRetrieveTableRsp **pRsp) {
+  int32_t code = 0;
   pCtx->jobDoneTs = taosGetTimestampMs();
   
   atomic_store_8((int8_t *)&pCtx->execDone, true);
 
+  if (taosHashGetSize(pCtx->groupHash) == atomic_load_32(&pCtx->groupDoneNum)) {
+    if (0 == taosWTryLockLatch(&pCtx->lock)) {
+      QRY_ERR_RET(qExplainGenerateRsp(pCtx, pRsp));
+      // LEAVE LOCK THERE
+    }
+  }
+
   return TSDB_CODE_SUCCESS;
 }
 
diff --git a/source/libs/executor/inc/executorimpl.h b/source/libs/executor/inc/executorimpl.h
index a92813812a..f73ef1af38 100644
--- a/source/libs/executor/inc/executorimpl.h
+++ b/source/libs/executor/inc/executorimpl.h
@@ -36,6 +36,7 @@ extern "C" {
 #include "thash.h"
 #include "tlockfree.h"
 #include "tpagedbuf.h"
+#include "tmsg.h"
 
 struct SColumnFilterElem;
 
@@ -160,7 +161,7 @@ typedef struct STaskCostInfo {
 
 typedef struct SOperatorCostInfo {
   uint64_t openCost;
-  uint64_t execCost;
+  uint64_t totalCost;
 } SOperatorCostInfo;
 
 // The basic query information extracted from the SQueryInfo tree to support the
@@ -714,6 +715,7 @@ int32_t getMaximumIdleDurationSec();
 void    doInvokeUdf(struct SUdfInfo* pUdfInfo, SqlFunctionCtx* pCtx, int32_t idx, int32_t type);
 void    setTaskStatus(SExecTaskInfo* pTaskInfo, int8_t status);
 int32_t createExecTaskInfoImpl(SSubplan* pPlan, SExecTaskInfo** pTaskInfo, SReadHandle* pHandle, uint64_t taskId, EOPTR_EXEC_MODEL model);
+int32_t getOperatorExplainExecInfo(SOperatorInfo *operator, SExplainExecInfo **pRes, int32_t *capacity, int32_t *resNum);
 
 #ifdef __cplusplus
 }
diff --git a/source/libs/executor/src/executorimpl.c b/source/libs/executor/src/executorimpl.c
index e423690740..498426d1a4 100644
--- a/source/libs/executor/src/executorimpl.c
+++ b/source/libs/executor/src/executorimpl.c
@@ -9523,11 +9523,27 @@ int32_t getOperatorExplainExecInfo(SOperatorInfo *operator, SExplainExecInfo **p
     
     *pRes = taosMemoryRealloc(*pRes, (*capacity) * sizeof(SExplainExecInfo));
     if (NULL == *pRes) {
-      qError("malloc %d failed", capacity * sizeof(SExplainExecInfo));
+      qError("malloc %d failed", (*capacity) * (int32_t)sizeof(SExplainExecInfo));
       return TSDB_CODE_QRY_OUT_OF_MEMORY;
     }
   }
 
+  (*pRes)[*resNum].numOfRows = operator->resultInfo.totalRows;
+  (*pRes)[*resNum].startupCost = operator->cost.openCost;
+  (*pRes)[*resNum].totalCost = operator->cost.totalCost;
+  
+  ++(*resNum);
+  
+  int32_t code = 0;
+  for (int32_t i = 0; i < operator->numOfDownstream; ++i) {
+    code = getOperatorExplainExecInfo(operator->pDownstream[i], pRes, capacity, resNum);
+    if (code) {
+      taosMemoryFreeClear(*pRes);
+      return TSDB_CODE_QRY_OUT_OF_MEMORY;
+    }
+  }
+
+  return TSDB_CODE_SUCCESS;
 }
 
 
diff --git a/source/libs/nodes/src/nodesUtilFuncs.c b/source/libs/nodes/src/nodesUtilFuncs.c
index 812fa333b6..2d5b38a1fe 100644
--- a/source/libs/nodes/src/nodesUtilFuncs.c
+++ b/source/libs/nodes/src/nodesUtilFuncs.c
@@ -1052,4 +1052,26 @@ char *nodesGetNameFromColumnNode(SNode *pNode) {
   return ((SColumnNode *)pNode)->colName;
 }
 
+int32_t nodesGetOutputNumFromSlotList(SNodeList* pSlots) {
+  if (NULL == pSlots || pSlots->length <= 0) {
+    return 0;
+  }
+
+  SNode* pNode = NULL;
+  int32_t num = 0;
+  FOREACH(pNode, pSlots) {
+    if (QUERY_NODE_SLOT_DESC != pNode->type) {
+      continue;
+    }
+
+    SSlotDescNode *descNode = (SSlotDescNode *)pNode;
+    if (descNode->output) {
+      ++num;
+    }
+  }
+
+  return num;
+}
+
+
 
diff --git a/source/libs/qworker/inc/qworkerMsg.h b/source/libs/qworker/inc/qworkerMsg.h
index 313b9459e2..c0f4d6d157 100644
--- a/source/libs/qworker/inc/qworkerMsg.h
+++ b/source/libs/qworker/inc/qworkerMsg.h
@@ -23,7 +23,7 @@ extern "C" {
 #include "qworkerInt.h"
 #include "dataSinkMgt.h"
 
-int32_t qwProcessQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg, int8_t taskType);
+int32_t qwProcessQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg, int8_t taskType, int8_t explain);
 int32_t qwProcessCQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg);
 int32_t qwProcessReady(QW_FPARAMS_DEF, SQWMsg *qwMsg);
 int32_t qwProcessFetch(QW_FPARAMS_DEF, SQWMsg *qwMsg);
@@ -38,6 +38,7 @@ void    qwBuildFetchRsp(void *msg, SOutputData *input, int32_t len, bool qComple
 int32_t qwBuildAndSendCQueryMsg(QW_FPARAMS_DEF, SQWConnInfo *pConn);
 int32_t qwBuildAndSendReadyRsp(SQWConnInfo *pConn, int32_t code);
 int32_t qwBuildAndSendQueryRsp(SQWConnInfo *pConn, int32_t code);
+int32_t qwBuildAndSendExplainRsp(SQWConnInfo *pConn, SExplainExecInfo *execInfo, int32_t num);
 void    qwFreeFetchRsp(void *msg);
 int32_t qwMallocFetchRsp(int32_t length, SRetrieveTableRsp **rsp);
 int32_t qwGetSchTasksStatus(SQWorkerMgmt *mgmt, uint64_t sId, SSchedulerStatusRsp **rsp);
diff --git a/source/libs/qworker/src/qworker.c b/source/libs/qworker/src/qworker.c
index 726d0ebf79..a83fcc26f2 100644
--- a/source/libs/qworker/src/qworker.c
+++ b/source/libs/qworker/src/qworker.c
@@ -549,7 +549,10 @@ int32_t qwHandleTaskComplete(QW_FPARAMS_DEF, SQWTaskCtx *ctx) {
       int32_t resNum = 0;
       QW_ERR_RET(qGetExplainExecInfo(ctx->taskHandle, &resNum, &execInfo));
 
-      QW_ERR_RET(qwBuildAndSendExplainRsp(&ctx->ctrlConnInfo, &rsp));
+      SQWConnInfo connInfo = {0};
+      connInfo.handle = ctx->ctrlConnInfo.handle;
+      
+      QW_ERR_RET(qwBuildAndSendExplainRsp(&connInfo, execInfo, resNum));
     }
     
     qwFreeTaskHandle(QW_FPARAMS(), taskHandle);
diff --git a/source/libs/qworker/src/qworkerMsg.c b/source/libs/qworker/src/qworkerMsg.c
index c9779f9252..bf8f97aa4c 100644
--- a/source/libs/qworker/src/qworkerMsg.c
+++ b/source/libs/qworker/src/qworkerMsg.c
@@ -85,6 +85,27 @@ int32_t qwBuildAndSendReadyRsp(SQWConnInfo *pConn, int32_t code) {
   return TSDB_CODE_SUCCESS;
 }
 
+int32_t qwBuildAndSendExplainRsp(SQWConnInfo *pConn, SExplainExecInfo *execInfo, int32_t num) {  
+  SExplainRsp rsp = {.numOfPlans = num, .subplanInfo = execInfo};
+
+  int32_t contLen = tSerializeSExplainRsp(NULL, 0, &rsp);
+  void *pRsp = rpcMallocCont(contLen);
+  tSerializeSExplainRsp(pRsp, contLen, &rsp);
+
+  SRpcMsg rpcRsp = {
+    .msgType = TDMT_VND_EXPLAIN_RSP,
+    .handle  = pConn->handle,
+    .ahandle = pConn->ahandle,
+    .pCont   = pRsp,
+    .contLen = contLen,
+    .code    = 0,
+  };
+
+  tmsgSendRsp(&rpcRsp);
+
+  return TSDB_CODE_SUCCESS;
+}
+
 int32_t qwBuildAndSendHbRsp(SQWConnInfo *pConn, SSchedulerHbRsp *pStatus, int32_t code) {
   int32_t contLen = tSerializeSSchedulerHbRsp(NULL, 0, pStatus);
   void *pRsp = rpcMallocCont(contLen);
diff --git a/source/libs/scheduler/src/scheduler.c b/source/libs/scheduler/src/scheduler.c
index 658d924f38..55eaae1b54 100644
--- a/source/libs/scheduler/src/scheduler.c
+++ b/source/libs/scheduler/src/scheduler.c
@@ -1043,6 +1043,19 @@ _return:
   SCH_RET(schProcessOnTaskFailure(pJob, pJob->fetchTask, code));
 }
 
+int32_t schProcessOnExplainDone(SSchJob *pJob, SSchTask *pTask, SRetrieveTableRsp *pRsp) {
+  SCH_TASK_DLOG("got explain rsp, rows:%d, complete:%d", htonl(pRsp->numOfRows), pRsp->completed);
+  
+  atomic_store_32(&pJob->resNumOfRows, htonl(pRsp->numOfRows));
+  atomic_store_ptr(&pJob->resData, pRsp);
+
+  SCH_SET_TASK_STATUS(pTask, JOB_TASK_STATUS_SUCCEED);
+  
+  schProcessOnDataFetched(pJob);
+
+  return TSDB_CODE_SUCCESS;
+}
+
 // Note: no more task error processing, handled in function internal
 int32_t schHandleResponseMsg(SSchJob *pJob, SSchTask *pTask, int32_t msgType, char *msg, int32_t msgSize,
                              int32_t rspCode) {
@@ -1061,7 +1074,7 @@ int32_t schHandleResponseMsg(SSchJob *pJob, SSchTask *pTask, int32_t msgType, ch
     case TDMT_VND_CREATE_TABLE_RSP: {
       SVCreateTbBatchRsp batchRsp = {0};
       if (msg) {
-        tDeserializeSVCreateTbBatchRsp(msg, msgSize, &batchRsp);
+        SCH_ERR_JRET(tDeserializeSVCreateTbBatchRsp(msg, msgSize, &batchRsp));
         if (batchRsp.rspList) {
           int32_t num = taosArrayGetSize(batchRsp.rspList);
           for (int32_t i = 0; i < num; ++i) {
@@ -1099,7 +1112,7 @@ int32_t schHandleResponseMsg(SSchJob *pJob, SSchTask *pTask, int32_t msgType, ch
     case TDMT_VND_QUERY_RSP: {
       SQueryTableRsp rsp = {0};
       if (msg) {
-        tDeserializeSQueryTableRsp(msg, msgSize, &rsp);
+        SCH_ERR_JRET(tDeserializeSQueryTableRsp(msg, msgSize, &rsp));
         SCH_ERR_JRET(rsp.code);
       }
 
@@ -1126,13 +1139,11 @@ int32_t schHandleResponseMsg(SSchJob *pJob, SSchTask *pTask, int32_t msgType, ch
       break;
     }
     case TDMT_VND_EXPLAIN_RSP: {
-      SExplainRsp *taskRsp = (SExplainRsp *)msg;
-
       SCH_ERR_JRET(rspCode);
       if (NULL == msg) {
         SCH_ERR_JRET(TSDB_CODE_QRY_INVALID_INPUT);
       }
-
+      
       if (!SCH_IS_EXPLAIN_JOB(pJob)) {
         SCH_TASK_ELOG("invalid msg received for none explain query, msg type:%s", TMSG_INFO(msgType));
         SCH_ERR_JRET(TSDB_CODE_QRY_INVALID_INPUT);
@@ -1140,22 +1151,20 @@ int32_t schHandleResponseMsg(SSchJob *pJob, SSchTask *pTask, int32_t msgType, ch
 
       if (pJob->resData) {
         SCH_TASK_ELOG("explain result is already generated, res:%p", pJob->resData);
-        taosMemoryFreeClear(taskRsp);
         SCH_ERR_JRET(TSDB_CODE_SCH_STATUS_ERROR);
       }
 
+      SExplainRsp rsp = {0};
+      if (tDeserializeSExplainRsp(msg, msgSize, &rsp)) {
+        taosMemoryFree(rsp.subplanInfo);
+        SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
+      }
+
       SRetrieveTableRsp *pRsp = NULL;
-      SCH_ERR_JRET(qExplainUpdateExecInfo(pJob->explainCtx, taskRsp, pTask->plan->id.groupId, &pRsp));
+      SCH_ERR_JRET(qExplainUpdateExecInfo(pJob->explainCtx, &rsp, pTask->plan->id.groupId, &pRsp));
 
       if (pRsp) {
-        atomic_store_ptr(&pJob->resData, pRsp);
-        atomic_store_32(&pJob->resNumOfRows, htonl(pRsp->numOfRows));
-
-        SCH_SET_TASK_STATUS(pTask, JOB_TASK_STATUS_SUCCEED);
-
-        SCH_TASK_DLOG("got explain rsp, rows:%d, complete:%d", htonl(pRsp->numOfRows), pRsp->completed);
-
-        schProcessOnDataFetched(pJob);
+        SCH_ERR_JRET(schProcessOnExplainDone(pJob, pTask, pRsp));
       }
       break;
     }
@@ -1168,8 +1177,13 @@ int32_t schHandleResponseMsg(SSchJob *pJob, SSchTask *pTask, int32_t msgType, ch
       }
 
       if (SCH_IS_EXPLAIN_JOB(pJob)) {
-        if (rsp->completed) {
-          SCH_ERR_JRET(qExecExplainEnd(pJob->explainCtx));
+        if (rsp->completed) {          
+          SRetrieveTableRsp *pRsp = NULL;
+          SCH_ERR_JRET(qExecExplainEnd(pJob->explainCtx, &pRsp));
+          if (pRsp) {
+            SCH_ERR_JRET(schProcessOnExplainDone(pJob, pTask, pRsp));
+          }
+          
           return TSDB_CODE_SUCCESS;
         }
 
@@ -1278,6 +1292,10 @@ int32_t schHandleReadyCallback(void *param, const SDataBuf *pMsg, int32_t code)
   return schHandleCallback(param, pMsg, TDMT_VND_RES_READY_RSP, code);
 }
 
+int32_t schHandleExplainCallback(void *param, const SDataBuf *pMsg, int32_t code) {
+  return schHandleCallback(param, pMsg, TDMT_VND_EXPLAIN_RSP, code);
+}
+
 int32_t schHandleDropCallback(void *param, const SDataBuf *pMsg, int32_t code) {
   SSchTaskCallbackParam *pParam = (SSchTaskCallbackParam *)param;
   qDebug("QID:%" PRIx64 ",TID:%" PRIx64 " drop task rsp received, code:%x", pParam->queryId, pParam->taskId, code);
@@ -1366,6 +1384,9 @@ int32_t schGetCallbackFp(int32_t msgType, __async_send_cb_fn_t *fp) {
     case TDMT_VND_RES_READY:
       *fp = schHandleReadyCallback;
       break;
+    case TDMT_VND_EXPLAIN:
+      *fp = schHandleExplainCallback;
+      break;
     case TDMT_VND_FETCH:
       *fp = schHandleFetchCallback;
       break;
@@ -1386,6 +1407,43 @@ int32_t schGetCallbackFp(int32_t msgType, __async_send_cb_fn_t *fp) {
   return TSDB_CODE_SUCCESS;
 }
 
+int32_t schGenerateTaskCallBackAHandle(SSchJob *pJob, SSchTask *pTask, int32_t msgType, SMsgSendInfo **pMsgSendInfo) {
+  int32_t code = 0;
+  SMsgSendInfo *msgSendInfo = taosMemoryCalloc(1, sizeof(SMsgSendInfo));
+  if (NULL == msgSendInfo) {
+    SCH_TASK_ELOG("calloc %d failed", (int32_t)sizeof(SMsgSendInfo));
+    SCH_ERR_RET(TSDB_CODE_QRY_OUT_OF_MEMORY);
+  }
+
+  SSchTaskCallbackParam *param = taosMemoryCalloc(1, sizeof(SSchTaskCallbackParam));
+  if (NULL == param) {
+    SCH_TASK_ELOG("calloc %d failed", (int32_t)sizeof(SSchTaskCallbackParam));
+    SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
+  }
+
+  __async_send_cb_fn_t fp = NULL;
+  SCH_ERR_JRET(schGetCallbackFp(msgType, &fp));
+
+  param->queryId = pJob->queryId;
+  param->refId = pJob->refId;
+  param->taskId = SCH_TASK_ID(pTask);
+  param->transport = pJob->transport;
+
+  msgSendInfo->param = param;
+  msgSendInfo->fp = fp;
+
+  *pMsgSendInfo = msgSendInfo;
+
+  return TSDB_CODE_SUCCESS;
+
+_return:
+
+  taosMemoryFree(param);
+  taosMemoryFree(msgSendInfo);
+
+  SCH_RET(code);
+}
+
 void schFreeRpcCtxVal(const void *arg) {
   if (NULL == arg) {
     return;
@@ -1472,8 +1530,8 @@ _return:
 
 int32_t schMakeQueryRpcCtx(SSchJob *pJob, SSchTask *pTask, SRpcCtx *pCtx) {
   int32_t                code = 0;
-  SSchTaskCallbackParam *param = NULL;
-  SMsgSendInfo          *pMsgSendInfo = NULL;
+  SMsgSendInfo *pReadyMsgSendInfo = NULL;
+  SMsgSendInfo *pExplainMsgSendInfo = NULL;
 
   pCtx->args = taosHashInit(1, taosGetDefaultHashFunction(TSDB_DATA_TYPE_INT), false, HASH_ENTRY_LOCK);
   if (NULL == pCtx->args) {
@@ -1481,31 +1539,18 @@ int32_t schMakeQueryRpcCtx(SSchJob *pJob, SSchTask *pTask, SRpcCtx *pCtx) {
     SCH_ERR_RET(TSDB_CODE_QRY_OUT_OF_MEMORY);
   }
 
-  pMsgSendInfo = taosMemoryCalloc(1, sizeof(SMsgSendInfo));
-  if (NULL == pMsgSendInfo) {
-    SCH_TASK_ELOG("calloc %d failed", (int32_t)sizeof(SMsgSendInfo));
+  SCH_ERR_JRET(schGenerateTaskCallBackAHandle(pJob, pTask, TDMT_VND_RES_READY, &pReadyMsgSendInfo));
+  SCH_ERR_JRET(schGenerateTaskCallBackAHandle(pJob, pTask, TDMT_VND_EXPLAIN, &pExplainMsgSendInfo));
+
+  int32_t msgType = TDMT_VND_RES_READY_RSP;
+  SRpcCtxVal ctxVal = {.val = pReadyMsgSendInfo, .clone = schCloneSMsgSendInfo, .freeFunc = schFreeRpcCtxVal};
+  if (taosHashPut(pCtx->args, &msgType, sizeof(msgType), &ctxVal, sizeof(ctxVal))) {
+    SCH_TASK_ELOG("taosHashPut msg %d to rpcCtx failed", msgType);
     SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
   }
 
-  param = taosMemoryCalloc(1, sizeof(SSchTaskCallbackParam));
-  if (NULL == param) {
-    SCH_TASK_ELOG("calloc %d failed", (int32_t)sizeof(SSchTaskCallbackParam));
-    SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
-  }
-
-  int32_t              msgType = TDMT_VND_RES_READY_RSP;
-  __async_send_cb_fn_t fp = NULL;
-  SCH_ERR_JRET(schGetCallbackFp(TDMT_VND_RES_READY, &fp));
-
-  param->queryId = pJob->queryId;
-  param->refId = pJob->refId;
-  param->taskId = SCH_TASK_ID(pTask);
-  param->transport = pJob->transport;
-
-  pMsgSendInfo->param = param;
-  pMsgSendInfo->fp = fp;
-
-  SRpcCtxVal ctxVal = {.val = pMsgSendInfo, .clone = schCloneSMsgSendInfo, .freeFunc = schFreeRpcCtxVal};
+  msgType = TDMT_VND_EXPLAIN_RSP;
+  ctxVal.val = pExplainMsgSendInfo;
   if (taosHashPut(pCtx->args, &msgType, sizeof(msgType), &ctxVal, sizeof(ctxVal))) {
     SCH_TASK_ELOG("taosHashPut msg %d to rpcCtx failed", msgType);
     SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
@@ -1518,8 +1563,16 @@ int32_t schMakeQueryRpcCtx(SSchJob *pJob, SSchTask *pTask, SRpcCtx *pCtx) {
 _return:
 
   taosHashCleanup(pCtx->args);
-  taosMemoryFreeClear(param);
-  taosMemoryFreeClear(pMsgSendInfo);
+  
+  if (pReadyMsgSendInfo) {
+    taosMemoryFreeClear(pReadyMsgSendInfo->param);
+    taosMemoryFreeClear(pReadyMsgSendInfo);
+  }
+
+  if (pExplainMsgSendInfo) {
+    taosMemoryFreeClear(pExplainMsgSendInfo->param);
+    taosMemoryFreeClear(pExplainMsgSendInfo);
+  }
 
   SCH_RET(code);
 }
@@ -1702,32 +1755,13 @@ int32_t schAsyncSendMsg(SSchJob *pJob, SSchTask *pTask, void *transport, SEpSet
 
   SSchTrans *trans = (SSchTrans *)transport;
 
-  SMsgSendInfo *pMsgSendInfo = taosMemoryCalloc(1, sizeof(SMsgSendInfo));
-  if (NULL == pMsgSendInfo) {
-    SCH_TASK_ELOG("calloc %d failed", (int32_t)sizeof(SMsgSendInfo));
-    SCH_ERR_RET(TSDB_CODE_QRY_OUT_OF_MEMORY);
-  }
+  SMsgSendInfo *pMsgSendInfo = NULL;
+  SCH_ERR_JRET(schGenerateTaskCallBackAHandle(pJob, pTask, msgType, &pMsgSendInfo));
 
-  SSchTaskCallbackParam *param = taosMemoryCalloc(1, sizeof(SSchTaskCallbackParam));
-  if (NULL == param) {
-    SCH_TASK_ELOG("calloc %d failed", (int32_t)sizeof(SSchTaskCallbackParam));
-    SCH_ERR_JRET(TSDB_CODE_QRY_OUT_OF_MEMORY);
-  }
-
-  __async_send_cb_fn_t fp = NULL;
-  SCH_ERR_JRET(schGetCallbackFp(msgType, &fp));
-
-  param->queryId = pJob->queryId;
-  param->refId = pJob->refId;
-  param->taskId = SCH_TASK_ID(pTask);
-  param->transport = trans->transInst;
-
-  pMsgSendInfo->param = param;
   pMsgSendInfo->msgInfo.pData = msg;
   pMsgSendInfo->msgInfo.len = msgSize;
   pMsgSendInfo->msgInfo.handle = trans->transHandle;
   pMsgSendInfo->msgType = msgType;
-  pMsgSendInfo->fp = fp;
 
   qDebug("start to send %s msg to node[%d,%s,%d], refId:%" PRIx64 "instance:%p, handle:%p", TMSG_INFO(msgType),
          ntohl(((SMsgHead *)msg)->vgId), epSet->eps[epSet->inUse].fqdn, epSet->eps[epSet->inUse].port, pJob->refId,
@@ -1744,8 +1778,11 @@ int32_t schAsyncSendMsg(SSchJob *pJob, SSchTask *pTask, void *transport, SEpSet
 
 _return:
 
-  taosMemoryFreeClear(param);
-  taosMemoryFreeClear(pMsgSendInfo);
+  if (pMsgSendInfo) {
+    taosMemoryFreeClear(pMsgSendInfo->param);
+    taosMemoryFreeClear(pMsgSendInfo);
+  }
+  
   SCH_RET(code);
 }
 
diff --git a/source/util/src/tlockfree.c b/source/util/src/tlockfree.c
index 8edcca0c16..a755a67cc8 100644
--- a/source/util/src/tlockfree.c
+++ b/source/util/src/tlockfree.c
@@ -53,6 +53,21 @@ void taosWLockLatch(SRWLatch *pLatch) {
   }
 }
 
+int32_t taosWTryLockLatch(SRWLatch *pLatch) {
+  SRWLatch oLatch, nLatch;
+  oLatch = atomic_load_32(pLatch);
+  if (oLatch) {
+    return -1;
+  }
+
+  nLatch = oLatch | TD_RWLATCH_WRITE_FLAG;
+  if (atomic_val_compare_exchange_32(pLatch, oLatch, nLatch) == oLatch) {
+    return 0;
+  }
+
+  return -1;
+}
+
 void taosWUnLockLatch(SRWLatch *pLatch) { atomic_store_32(pLatch, 0); }
 
 void taosRLockLatch(SRWLatch *pLatch) {

From 4540f53b32cedfc3b15646913cdfa6bcd0ea2d44 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 01:45:22 +0000
Subject: [PATCH 07/25] make feature/meta compile

---
 source/dnode/vnode/src/meta/metaTDBImpl.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/source/dnode/vnode/src/meta/metaTDBImpl.c b/source/dnode/vnode/src/meta/metaTDBImpl.c
index 8e82cf1abc..a004a96ad0 100644
--- a/source/dnode/vnode/src/meta/metaTDBImpl.c
+++ b/source/dnode/vnode/src/meta/metaTDBImpl.c
@@ -212,7 +212,7 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
   pVal = pBuf = buf;
   metaEncodeTbInfo(&pBuf, pTbCfg);
   vLen = POINTER_DISTANCE(pBuf, buf);
-  ret = tdbDbInsert(pMetaDb->pTbDB, pKey, kLen, pVal, vLen);
+  ret = tdbDbInsert(pMetaDb->pTbDB, pKey, kLen, pVal, vLen, NULL);
   if (ret < 0) {
     return -1;
   }
@@ -234,7 +234,7 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
     pVal = pBuf = buf;
     metaEncodeSchemaEx(&pBuf, &schemaWrapper);
     vLen = POINTER_DISTANCE(pBuf, buf);
-    ret = tdbDbInsert(pMetaDb->pSchemaDB, pKey, kLen, pVal, vLen);
+    ret = tdbDbInsert(pMetaDb->pSchemaDB, pKey, kLen, pVal, vLen, NULL);
     if (ret < 0) {
       return -1;
     }
@@ -248,7 +248,7 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
   kLen = nameLen + 1 + sizeof(uid);
   pVal = NULL;
   vLen = 0;
-  ret = tdbDbInsert(pMetaDb->pNameIdx, pKey, kLen, pVal, vLen);
+  ret = tdbDbInsert(pMetaDb->pNameIdx, pKey, kLen, pVal, vLen, NULL);
   if (ret < 0) {
     return -1;
   }
@@ -259,7 +259,7 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
     kLen = sizeof(uid);
     pVal = NULL;
     vLen = 0;
-    ret = tdbDbInsert(pMetaDb->pStbIdx, pKey, kLen, pVal, vLen);
+    ret = tdbDbInsert(pMetaDb->pStbIdx, pKey, kLen, pVal, vLen, NULL);
     if (ret < 0) {
       return -1;
     }
@@ -270,7 +270,7 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
     kLen = sizeof(ctbIdxKey);
     pVal = NULL;
     vLen = 0;
-    ret = tdbDbInsert(pMetaDb->pCtbIdx, pKey, kLen, pVal, vLen);
+    ret = tdbDbInsert(pMetaDb->pCtbIdx, pKey, kLen, pVal, vLen, NULL);
     if (ret < 0) {
       return -1;
     }
@@ -279,7 +279,7 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
     kLen = sizeof(uid);
     pVal = NULL;
     vLen = 0;
-    ret = tdbDbInsert(pMetaDb->pNtbIdx, pKey, kLen, pVal, vLen);
+    ret = tdbDbInsert(pMetaDb->pNtbIdx, pKey, kLen, pVal, vLen, NULL);
     if (ret < 0) {
       return -1;
     }

From 906b615c9a138501f0f8e683db86f840fb6b022e Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 01:55:37 +0000
Subject: [PATCH 08/25] make TDB compile

---
 source/dnode/vnode/CMakeLists.txt         | 2 +-
 source/dnode/vnode/src/meta/metaTDBImpl.c | 6 +++---
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/source/dnode/vnode/CMakeLists.txt b/source/dnode/vnode/CMakeLists.txt
index 56a8698d77..6b21f935c4 100644
--- a/source/dnode/vnode/CMakeLists.txt
+++ b/source/dnode/vnode/CMakeLists.txt
@@ -1,5 +1,5 @@
 set(META_DB_IMPL_LIST "BDB" "TDB")
-set(META_DB_IMPL "BDB" CACHE STRING "Use BDB as the default META implementation")
+set(META_DB_IMPL "TDB" CACHE STRING "Use BDB as the default META implementation")
 set_property(CACHE META_DB_IMPL PROPERTY STRINGS ${META_DB_IMPL_LIST})
 
 if(META_DB_IMPL IN_LIST META_DB_IMPL_LIST)
diff --git a/source/dnode/vnode/src/meta/metaTDBImpl.c b/source/dnode/vnode/src/meta/metaTDBImpl.c
index a004a96ad0..cfbd83df5c 100644
--- a/source/dnode/vnode/src/meta/metaTDBImpl.c
+++ b/source/dnode/vnode/src/meta/metaTDBImpl.c
@@ -226,10 +226,10 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
 
     if (pTbCfg->type == META_SUPER_TABLE) {
       schemaWrapper.nCols = pTbCfg->stbCfg.nCols;
-      schemaWrapper.pSchema = pTbCfg->stbCfg.pSchema;
+      schemaWrapper.pSchemaEx = pTbCfg->stbCfg.pSchema;
     } else {
       schemaWrapper.nCols = pTbCfg->ntbCfg.nCols;
-      schemaWrapper.pSchema = pTbCfg->ntbCfg.pSchema;
+      schemaWrapper.pSchemaEx = pTbCfg->ntbCfg.pSchema;
     }
     pVal = pBuf = buf;
     metaEncodeSchemaEx(&pBuf, &schemaWrapper);
@@ -349,7 +349,7 @@ STbCfg *metaGetTbInfoByName(SMeta *pMeta, char *tbname, tb_uid_t *uid) {
 }
 
 SSchemaWrapper *metaGetTableSchema(SMeta *pMeta, tb_uid_t uid, int32_t sver, bool isinline) {
-  return *metaGetTableSchemaImpl(pMeta, uid, sver, isinline, false);
+  return metaGetTableSchemaImpl(pMeta, uid, sver, isinline, false);
 }
 
 static SSchemaWrapper *metaGetTableSchemaImpl(SMeta *pMeta, tb_uid_t uid, int32_t sver, bool isinline, bool isGetEx) {

From 317d4ff2eb8dd83c34d6d9fdb6ea6c52df181b5c Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 02:06:59 +0000
Subject: [PATCH 09/25] make tdb work with meta

---
 source/dnode/vnode/src/meta/metaCommit.c  | 21 ---------------------
 source/dnode/vnode/src/meta/metaTDBImpl.c |  7 +++++++
 2 files changed, 7 insertions(+), 21 deletions(-)
 delete mode 100644 source/dnode/vnode/src/meta/metaCommit.c

diff --git a/source/dnode/vnode/src/meta/metaCommit.c b/source/dnode/vnode/src/meta/metaCommit.c
deleted file mode 100644
index 805e4f3e3d..0000000000
--- a/source/dnode/vnode/src/meta/metaCommit.c
+++ /dev/null
@@ -1,21 +0,0 @@
-/*
- * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
- *
- * This program is free software: you can use, redistribute, and/or modify
- * it under the terms of the GNU Affero General Public License, version 3
- * or later ("AGPL"), as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but WITHOUT
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
- * FITNESS FOR A PARTICULAR PURPOSE.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program. If not, see <http://www.gnu.org/licenses/>.
- */
-
-#include "meta.h"
-
-int metaCommit(SMeta *pMeta) {
-  // TODO
-  return 0;
-}
\ No newline at end of file
diff --git a/source/dnode/vnode/src/meta/metaTDBImpl.c b/source/dnode/vnode/src/meta/metaTDBImpl.c
index cfbd83df5c..e934caa7e3 100644
--- a/source/dnode/vnode/src/meta/metaTDBImpl.c
+++ b/source/dnode/vnode/src/meta/metaTDBImpl.c
@@ -167,6 +167,8 @@ int metaOpenDB(SMeta *pMeta) {
     return -1;
   }
 
+  tdbBegin(pMetaDb->pEnv, NULL);
+
   pMeta->pDB = pMetaDb;
   return 0;
 }
@@ -701,3 +703,8 @@ static void *metaDecodeTbInfo(void *buf, STbCfg *pTbCfg) {
   }
   return buf;
 }
+
+int metaCommit(SMeta *pMeta) {
+  tdbCommit(pMeta->pDB->pEnv, NULL);
+  return 0;
+}
\ No newline at end of file

From 2c6dec89d2215a04b2ce605cc701258bc7d9fcdc Mon Sep 17 00:00:00 2001
From: dapan1121 <wpan@taosdata.com>
Date: Wed, 6 Apr 2022 11:32:21 +0800
Subject: [PATCH 10/25] feature/qnode

---
 include/common/tmsg.h                         |  1 +
 include/libs/scalar/filter.h                  |  2 +-
 source/libs/command/inc/commandInt.h          |  1 +
 source/libs/command/src/explain.c             | 57 +++++++++++++++++--
 source/libs/executor/inc/executorimpl.h       | 44 +++++++-------
 source/libs/executor/src/executorimpl.c       | 24 ++++++--
 source/libs/scalar/src/filter.c               |  3 +-
 .../libs/scalar/test/filter/filterTests.cpp   | 22 +++----
 source/libs/scheduler/src/scheduler.c         |  3 +
 9 files changed, 114 insertions(+), 43 deletions(-)

diff --git a/include/common/tmsg.h b/include/common/tmsg.h
index 27b9a11625..9e04b5a618 100644
--- a/include/common/tmsg.h
+++ b/include/common/tmsg.h
@@ -919,6 +919,7 @@ typedef struct SExplainExecInfo {
   uint64_t startupCost;
   uint64_t totalCost;
   uint64_t numOfRows;
+  void    *verboseInfo;
 } SExplainExecInfo;
 
 typedef struct {
diff --git a/include/libs/scalar/filter.h b/include/libs/scalar/filter.h
index a93180800e..c81cb49b64 100644
--- a/include/libs/scalar/filter.h
+++ b/include/libs/scalar/filter.h
@@ -40,7 +40,7 @@ extern int32_t filterInitFromNode(SNode *pNode, SFilterInfo **pinfo, uint32_t op
 extern bool filterExecute(SFilterInfo *info, SSDataBlock *pSrc, int8_t** p, SColumnDataAgg *statis, int16_t numOfCols);
 extern int32_t filterSetDataFromSlotId(SFilterInfo *info, void *param);
 extern int32_t filterSetDataFromColId(SFilterInfo *info, void *param);
-extern int32_t filterGetTimeRange(SFilterInfo *info, STimeWindow *win);
+extern int32_t filterGetTimeRange(SNode *pNode, STimeWindow *win, bool *isStrict);
 extern int32_t filterConverNcharColumns(SFilterInfo* pFilterInfo, int32_t rows, bool *gotNchar);
 extern int32_t filterFreeNcharColumns(SFilterInfo* pFilterInfo);
 extern void filterFreeInfo(SFilterInfo *info);
diff --git a/source/libs/command/inc/commandInt.h b/source/libs/command/inc/commandInt.h
index 7e9ea12ea4..23e4f2b24f 100644
--- a/source/libs/command/inc/commandInt.h
+++ b/source/libs/command/inc/commandInt.h
@@ -58,6 +58,7 @@ extern "C" {
 #define EXPLAIN_LOOPS_FORMAT "loops=%d"
 #define EXPLAIN_REVERSE_FORMAT "reverse=%d"
 #define EXPLAIN_FUNCTIONS_FORMAT "functions=%d"
+#define EXPLAIN_EXECINFO_FORMAT "cost=%" PRIu64 "..%" PRIu64 " rows=%" PRIu64
 
 typedef struct SExplainGroup {
   int32_t   nodeNum;
diff --git a/source/libs/command/src/explain.c b/source/libs/command/src/explain.c
index c8502388ba..2a36113749 100644
--- a/source/libs/command/src/explain.c
+++ b/source/libs/command/src/explain.c
@@ -228,7 +228,10 @@ int32_t qExplainGenerateResNode(SPhysiNode *pNode, SExplainGroup *group, SExplai
 
   int32_t code = 0;
   resNode->pNode = pNode;
-  QRY_ERR_JRET(qExplainGenerateResNodeExecInfo(&resNode->pExecInfo, group));
+  
+  if (group->nodeExecInfo) {
+    QRY_ERR_JRET(qExplainGenerateResNodeExecInfo(&resNode->pExecInfo, group));
+  }
   
   QRY_ERR_JRET(qExplainGenerateResChildren(pNode, group, &resNode->pChildren));
 
@@ -247,14 +250,52 @@ _return:
 
 int32_t qExplainBufAppendExecInfo(SArray *pExecInfo, char *tbuf, int32_t *len) {
   int32_t tlen = *len;
+  int32_t nodeNum = taosArrayGetSize(pExecInfo);
+  SExplainExecInfo maxExecInfo = {0};
   
-  EXPLAIN_ROW_APPEND("(exec info here)");
+  for (int32_t i = 0; i < nodeNum; ++i) {
+    SExplainExecInfo *execInfo = taosArrayGet(pExecInfo, i);
+    if (execInfo->startupCost > maxExecInfo.startupCost) {
+      maxExecInfo.startupCost = execInfo->startupCost;
+    }
+    if (execInfo->totalCost > maxExecInfo.totalCost) {
+      maxExecInfo.totalCost = execInfo->totalCost;
+    }
+    if (execInfo->numOfRows > maxExecInfo.numOfRows) {
+      maxExecInfo.numOfRows = execInfo->numOfRows;
+    }
+  }
   
+  EXPLAIN_ROW_APPEND(EXPLAIN_EXECINFO_FORMAT, maxExecInfo.startupCost, maxExecInfo.totalCost, maxExecInfo.numOfRows);
+
   *len = tlen;
   
   return TSDB_CODE_SUCCESS;
 }
 
+int32_t qExplainBufAppendVerboseExecInfo(SArray *pExecInfo, char *tbuf, int32_t *len) {
+  int32_t tlen = 0;
+  bool gotVerbose = false;
+  int32_t nodeNum = taosArrayGetSize(pExecInfo);
+  SExplainExecInfo maxExecInfo = {0};
+  
+  for (int32_t i = 0; i < nodeNum; ++i) {
+    SExplainExecInfo *execInfo = taosArrayGet(pExecInfo, i);
+    if (execInfo->verboseInfo) {
+      gotVerbose = true;
+    }
+  }
+
+  if (gotVerbose) {
+    EXPLAIN_ROW_APPEND("exec verbose info");
+  }
+
+  *len = tlen;
+  
+  return TSDB_CODE_SUCCESS;
+}
+
+
 int32_t qExplainResAppendRow(SExplainCtx *ctx, char *tbuf, int32_t len, int32_t level) {
   SQueryExplainRowInfo row = {0};
   row.buf = taosMemoryMalloc(len);
@@ -322,6 +363,14 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
         EXPLAIN_ROW_NEW(level + 1, EXPLAIN_ORDER_FORMAT, EXPLAIN_ORDER_STRING(pTagScanNode->order));
         EXPLAIN_ROW_END();
         QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+
+        if (pResNode->pExecInfo) {
+          QRY_ERR_RET(qExplainBufAppendVerboseExecInfo(pResNode->pExecInfo, tbuf, &tlen));
+          if (tlen) {
+            EXPLAIN_ROW_END();
+            QRY_ERR_RET(qExplainResAppendRow(ctx, tbuf, tlen, level + 1));
+          }
+        }      
       }
       break;
     }
@@ -532,8 +581,8 @@ int32_t qExplainResNodeToRowsImpl(SExplainResNode *pResNode, SExplainCtx *ctx, i
       EXPLAIN_ROW_APPEND(EXPLAIN_LEFT_PARENTHESIS_FORMAT);
       if (pResNode->pExecInfo) {
         QRY_ERR_RET(qExplainBufAppendExecInfo(pResNode->pExecInfo, tbuf, &tlen));
+        EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
       }      
-      EXPLAIN_ROW_APPEND(EXPLAIN_BLANK_FORMAT);
       EXPLAIN_ROW_APPEND(EXPLAIN_WIDTH_FORMAT, pExchNode->node.pOutputDataBlockDesc->totalRowSize);
       EXPLAIN_ROW_APPEND(EXPLAIN_RIGHT_PARENTHESIS_FORMAT);      
       EXPLAIN_ROW_END();
@@ -710,7 +759,7 @@ int32_t qExplainAppendGroupResRows(void *pCtx, int32_t groupId, int32_t level) {
   
   QRY_ERR_RET(qExplainGenerateResNode(group->plan->pNode, group, &node));
 
-  if (group->physiPlanNum != group->physiPlanExecNum) {
+  if ((EXPLAIN_MODE_ANALYZE == ctx->mode) && (group->physiPlanNum != group->physiPlanExecNum)) {
     qError("physiPlanNum %d mismatch with physiExecNum %d in group %d", group->physiPlanNum, group->physiPlanExecNum, groupId);
     QRY_ERR_JRET(TSDB_CODE_QRY_APP_ERROR);
   }
diff --git a/source/libs/executor/inc/executorimpl.h b/source/libs/executor/inc/executorimpl.h
index f73ef1af38..2f667582e5 100644
--- a/source/libs/executor/inc/executorimpl.h
+++ b/source/libs/executor/inc/executorimpl.h
@@ -237,6 +237,7 @@ typedef bool (*__optr_decode_fn_t)(struct SOperatorInfo* pOperator, char *result
 typedef int32_t (*__optr_open_fn_t)(struct SOperatorInfo* pOptr);
 typedef SSDataBlock* (*__optr_fn_t)(struct SOperatorInfo* pOptr, bool* newgroup);
 typedef void (*__optr_close_fn_t)(void* param, int32_t num);
+typedef int32_t (*__optr_get_explain_fn_t)(struct SOperatorInfo* pOptr, void **pOptrExplain);
 
 typedef struct STaskIdInfo {
   uint64_t queryId;  // this is also a request id
@@ -305,26 +306,27 @@ enum {
 };
 
 typedef struct SOperatorInfo {
-  uint8_t                operatorType;
-  bool                   blockingOptr;  // block operator or not
-  uint8_t                status;        // denote if current operator is completed
-  int32_t                numOfOutput;   // number of columns of the current operator results
-  char*                  name;          // name, used to show the query execution plan
-  void*                  info;          // extension attribution
-  SExprInfo*             pExpr;
-  STaskRuntimeEnv*       pRuntimeEnv;   // todo remove it
-  SExecTaskInfo*         pTaskInfo;
-  SOperatorCostInfo      cost;
-  SResultInfo            resultInfo;
-  struct SOperatorInfo** pDownstream;      // downstram pointer list
-  int32_t                numOfDownstream;  // number of downstream. The value is always ONE expect for join operator
-  __optr_open_fn_t       _openFn;          // DO NOT invoke this function directly
-  __optr_fn_t            getNextFn;
-  __optr_fn_t            getStreamResFn;   // execute the aggregate in the stream model.
-  __optr_fn_t            cleanupFn;        // call this function to release the allocated resources ASAP
-  __optr_close_fn_t      closeFn;
-  __optr_encode_fn_t     encodeResultRow;
-  __optr_decode_fn_t     decodeResultRow;
+  uint8_t                 operatorType;
+  bool                    blockingOptr;  // block operator or not
+  uint8_t                 status;        // denote if current operator is completed
+  int32_t                 numOfOutput;   // number of columns of the current operator results
+  char*                   name;          // name, used to show the query execution plan
+  void*                   info;          // extension attribution
+  SExprInfo*              pExpr;
+  STaskRuntimeEnv*        pRuntimeEnv;   // todo remove it
+  SExecTaskInfo*          pTaskInfo;
+  SOperatorCostInfo       cost;
+  SResultInfo             resultInfo;
+  struct SOperatorInfo**  pDownstream;      // downstram pointer list
+  int32_t                 numOfDownstream;  // number of downstream. The value is always ONE expect for join operator
+  __optr_open_fn_t        _openFn;          // DO NOT invoke this function directly
+  __optr_fn_t             getNextFn;
+  __optr_fn_t             getStreamResFn;   // execute the aggregate in the stream model.
+  __optr_fn_t             cleanupFn;        // call this function to release the allocated resources ASAP
+  __optr_close_fn_t       closeFn;
+  __optr_encode_fn_t      encodeResultRow;
+  __optr_decode_fn_t      decodeResultRow;
+  __optr_get_explain_fn_t getExplainFn;
 } SOperatorInfo;
 
 typedef struct {
@@ -715,7 +717,7 @@ int32_t getMaximumIdleDurationSec();
 void    doInvokeUdf(struct SUdfInfo* pUdfInfo, SqlFunctionCtx* pCtx, int32_t idx, int32_t type);
 void    setTaskStatus(SExecTaskInfo* pTaskInfo, int8_t status);
 int32_t createExecTaskInfoImpl(SSubplan* pPlan, SExecTaskInfo** pTaskInfo, SReadHandle* pHandle, uint64_t taskId, EOPTR_EXEC_MODEL model);
-int32_t getOperatorExplainExecInfo(SOperatorInfo *operator, SExplainExecInfo **pRes, int32_t *capacity, int32_t *resNum);
+int32_t getOperatorExplainExecInfo(SOperatorInfo *operatorInfo, SExplainExecInfo **pRes, int32_t *capacity, int32_t *resNum);
 
 #ifdef __cplusplus
 }
diff --git a/source/libs/executor/src/executorimpl.c b/source/libs/executor/src/executorimpl.c
index 498426d1a4..729d9efd86 100644
--- a/source/libs/executor/src/executorimpl.c
+++ b/source/libs/executor/src/executorimpl.c
@@ -5585,6 +5585,10 @@ SOperatorInfo* createTableScanOperatorInfo(void* pTsdbReadHandle, int32_t order,
   pOperator->getNextFn = doTableScan;
   pOperator->pTaskInfo = pTaskInfo;
 
+  //pOperator->cost.openCost = 1314;
+  //pOperator->cost.totalCost = 2324;
+  //pOperator->resultInfo.totalRows = 3334;
+  
   return pOperator;
 }
 
@@ -9517,7 +9521,7 @@ void releaseQueryBuf(size_t numOfTables) {
   atomic_add_fetch_64(&tsQueryBufferSizeBytes, t);
 }
 
-int32_t getOperatorExplainExecInfo(SOperatorInfo *operator, SExplainExecInfo **pRes, int32_t *capacity, int32_t *resNum) {
+int32_t getOperatorExplainExecInfo(SOperatorInfo *operatorInfo, SExplainExecInfo **pRes, int32_t *capacity, int32_t *resNum) {
   if (*resNum >= *capacity) {
     *capacity += 10;
     
@@ -9528,15 +9532,23 @@ int32_t getOperatorExplainExecInfo(SOperatorInfo *operator, SExplainExecInfo **p
     }
   }
 
-  (*pRes)[*resNum].numOfRows = operator->resultInfo.totalRows;
-  (*pRes)[*resNum].startupCost = operator->cost.openCost;
-  (*pRes)[*resNum].totalCost = operator->cost.totalCost;
+  (*pRes)[*resNum].numOfRows = operatorInfo->resultInfo.totalRows;
+  (*pRes)[*resNum].startupCost = operatorInfo->cost.openCost;
+  (*pRes)[*resNum].totalCost = operatorInfo->cost.totalCost;
+
+  if (operatorInfo->getExplainFn) {
+    int32_t code = (*operatorInfo->getExplainFn)(operatorInfo, &(*pRes)->verboseInfo);
+    if (code) {
+      qError("operator getExplainFn failed, error:%s", tstrerror(code));
+      return code;
+    }
+  }
   
   ++(*resNum);
   
   int32_t code = 0;
-  for (int32_t i = 0; i < operator->numOfDownstream; ++i) {
-    code = getOperatorExplainExecInfo(operator->pDownstream[i], pRes, capacity, resNum);
+  for (int32_t i = 0; i < operatorInfo->numOfDownstream; ++i) {
+    code = getOperatorExplainExecInfo(operatorInfo->pDownstream[i], pRes, capacity, resNum);
     if (code) {
       taosMemoryFreeClear(*pRes);
       return TSDB_CODE_QRY_OUT_OF_MEMORY;
diff --git a/source/libs/scalar/src/filter.c b/source/libs/scalar/src/filter.c
index b0632bbc34..0502774d1b 100644
--- a/source/libs/scalar/src/filter.c
+++ b/source/libs/scalar/src/filter.c
@@ -3314,7 +3314,8 @@ bool filterRangeExecute(SFilterInfo *info, SColumnDataAgg *pDataStatis, int32_t
 
 
 
-int32_t filterGetTimeRange(SFilterInfo *info, STimeWindow       *win) {
+int32_t filterGetTimeRange(SNode *pNode, STimeWindow *win, bool *isStrict) {
+  SFilterInfo *info = NULL;
   SFilterRange ra = {0};
   SFilterRangeCtx *prev = filterInitRangeCtx(TSDB_DATA_TYPE_TIMESTAMP, FLT_OPTION_TIMESTAMP);
   SFilterRangeCtx *tmpc = filterInitRangeCtx(TSDB_DATA_TYPE_TIMESTAMP, FLT_OPTION_TIMESTAMP);
diff --git a/source/libs/scalar/test/filter/filterTests.cpp b/source/libs/scalar/test/filter/filterTests.cpp
index 54f82eae2d..696a697a91 100644
--- a/source/libs/scalar/test/filter/filterTests.cpp
+++ b/source/libs/scalar/test/filter/filterTests.cpp
@@ -234,15 +234,16 @@ TEST(timerangeTest, greater) {
   flttMakeValueNode(&pval, TSDB_DATA_TYPE_TIMESTAMP, &tsmall);
   flttMakeOpNode(&opNode1, OP_TYPE_GREATER_THAN, TSDB_DATA_TYPE_BOOL, pcol, pval);
 
-  SFilterInfo *filter = NULL;
-  int32_t code = filterInitFromNode(opNode1, &filter, FLT_OPTION_NO_REWRITE|FLT_OPTION_TIMESTAMP);
-  ASSERT_EQ(code, 0);
+  //SFilterInfo *filter = NULL;
+  //int32_t code = filterInitFromNode(opNode1, &filter, FLT_OPTION_NO_REWRITE|FLT_OPTION_TIMESTAMP);
+  //ASSERT_EQ(code, 0);
   STimeWindow win = {0};
-  code = filterGetTimeRange(filter, &win);
+  bool isStrict = false;
+  int32_t code = filterGetTimeRange(opNode1, &win, &isStrict);
   ASSERT_EQ(code, 0);
   ASSERT_EQ(win.skey, tsmall);
   ASSERT_EQ(win.ekey, INT64_MAX); 
-  filterFreeInfo(filter);
+  //filterFreeInfo(filter);
   nodesDestroyNode(opNode1);
 }
 
@@ -263,15 +264,16 @@ TEST(timerangeTest, greater_and_lower) {
   
   flttMakeLogicNode(&logicNode, LOGIC_COND_TYPE_AND, list, 2);
 
-  SFilterInfo *filter = NULL;
-  int32_t code = filterInitFromNode(logicNode, &filter, FLT_OPTION_NO_REWRITE|FLT_OPTION_TIMESTAMP);
-  ASSERT_EQ(code, 0);
+  //SFilterInfo *filter = NULL;
+  //int32_t code = filterInitFromNode(logicNode, &filter, FLT_OPTION_NO_REWRITE|FLT_OPTION_TIMESTAMP);
+  //ASSERT_EQ(code, 0);
   STimeWindow win = {0};
-  code = filterGetTimeRange(filter, &win);
+  bool isStrict = false;
+  int32_t code = filterGetTimeRange(logicNode, &win, &isStrict);
   ASSERT_EQ(code, 0);
   ASSERT_EQ(win.skey, tsmall);
   ASSERT_EQ(win.ekey, tbig); 
-  filterFreeInfo(filter);
+  //filterFreeInfo(filter);
   nodesDestroyNode(logicNode);
 }
 
diff --git a/source/libs/scheduler/src/scheduler.c b/source/libs/scheduler/src/scheduler.c
index 55eaae1b54..1675f91330 100644
--- a/source/libs/scheduler/src/scheduler.c
+++ b/source/libs/scheduler/src/scheduler.c
@@ -199,6 +199,7 @@ int32_t schValidateTaskReceivedMsgType(SSchJob *pJob, SSchTask *pTask, int32_t m
   int32_t reqMsgType = msgType - 1;
   switch (msgType) {
     case TDMT_SCH_LINK_BROKEN:
+    case TDMT_VND_EXPLAIN_RSP:
       return TSDB_CODE_SUCCESS;
     case TDMT_VND_QUERY_RSP:  // query_rsp may be processed later than ready_rsp
       if (lastMsgType != reqMsgType && -1 != lastMsgType && TDMT_VND_FETCH != lastMsgType) {
@@ -1187,6 +1188,8 @@ int32_t schHandleResponseMsg(SSchJob *pJob, SSchTask *pTask, int32_t msgType, ch
           return TSDB_CODE_SUCCESS;
         }
 
+        atomic_val_compare_exchange_32(&pJob->remoteFetch, 1, 0);
+
         SCH_ERR_JRET(schFetchFromRemote(pJob));
 
         return TSDB_CODE_SUCCESS;

From 69ff3cc67b4980bb96aa97852b03f3d1eeb3ca09 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 06:37:22 +0000
Subject: [PATCH 11/25] fix an assert bug

---
 source/libs/tdb/src/db/tdbPage.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/source/libs/tdb/src/db/tdbPage.c b/source/libs/tdb/src/db/tdbPage.c
index cc4ce20802..ea75f7fbeb 100644
--- a/source/libs/tdb/src/db/tdbPage.c
+++ b/source/libs/tdb/src/db/tdbPage.c
@@ -278,7 +278,7 @@ static int tdbPageAllocate(SPage *pPage, int szCell, SCell **ppCell) {
 
   // 2. Try to allocate from the page free list
   cellFree = TDB_PAGE_FCELL(pPage);
-  ASSERT(cellFree == 0 || cellFree > pPage->pFreeEnd - pPage->pData);
+  ASSERT(cellFree == 0 || cellFree >= pPage->pFreeEnd - pPage->pData);
   if (cellFree && pPage->pFreeEnd - pPage->pFreeStart >= TDB_PAGE_OFFSET_SIZE(pPage)) {
     SCell *pPrevFreeCell = NULL;
     int    szPrevFreeCell;

From 61d0180b852f93333d086cba388e33c0918675f9 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 06:37:37 +0000
Subject: [PATCH 12/25] more TDB integration

---
 source/dnode/vnode/src/meta/metaTDBImpl.c | 125 +++++++++++++++++++---
 source/dnode/vnode/src/meta/metaTbUid.c   |   2 +-
 source/dnode/vnode/src/vnd/vnodeCommit.c  |   2 +-
 3 files changed, 112 insertions(+), 17 deletions(-)

diff --git a/source/dnode/vnode/src/meta/metaTDBImpl.c b/source/dnode/vnode/src/meta/metaTDBImpl.c
index e934caa7e3..282e0a51f6 100644
--- a/source/dnode/vnode/src/meta/metaTDBImpl.c
+++ b/source/dnode/vnode/src/meta/metaTDBImpl.c
@@ -16,15 +16,28 @@
 #include "metaDef.h"
 
 #include "tdbInt.h"
+typedef struct SPoolMem {
+  int64_t          size;
+  struct SPoolMem *prev;
+  struct SPoolMem *next;
+} SPoolMem;
+
+static SPoolMem *openPool();
+static void      clearPool(SPoolMem *pPool);
+static void      closePool(SPoolMem *pPool);
+static void     *poolMalloc(void *arg, size_t size);
+static void      poolFree(void *arg, void *ptr);
 
 struct SMetaDB {
-  TENV *pEnv;
-  TDB  *pTbDB;
-  TDB  *pSchemaDB;
-  TDB  *pNameIdx;
-  TDB  *pStbIdx;
-  TDB  *pNtbIdx;
-  TDB  *pCtbIdx;
+  TXN       txn;
+  TENV     *pEnv;
+  TDB      *pTbDB;
+  TDB      *pSchemaDB;
+  TDB      *pNameIdx;
+  TDB      *pStbIdx;
+  TDB      *pNtbIdx;
+  TDB      *pCtbIdx;
+  SPoolMem *pPool;
 };
 
 typedef struct __attribute__((__packed__)) {
@@ -167,6 +180,8 @@ int metaOpenDB(SMeta *pMeta) {
     return -1;
   }
 
+  pMetaDb->pPool = openPool();
+  tdbTxnOpen(&pMetaDb->txn, 0, poolMalloc, poolFree, pMetaDb->pPool, TDB_TXN_WRITE | TDB_TXN_READ_UNCOMMITTED);
   tdbBegin(pMetaDb->pEnv, NULL);
 
   pMeta->pDB = pMetaDb;
@@ -214,7 +229,7 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
   pVal = pBuf = buf;
   metaEncodeTbInfo(&pBuf, pTbCfg);
   vLen = POINTER_DISTANCE(pBuf, buf);
-  ret = tdbDbInsert(pMetaDb->pTbDB, pKey, kLen, pVal, vLen, NULL);
+  ret = tdbDbInsert(pMetaDb->pTbDB, pKey, kLen, pVal, vLen, &pMetaDb->txn);
   if (ret < 0) {
     return -1;
   }
@@ -236,7 +251,7 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
     pVal = pBuf = buf;
     metaEncodeSchemaEx(&pBuf, &schemaWrapper);
     vLen = POINTER_DISTANCE(pBuf, buf);
-    ret = tdbDbInsert(pMetaDb->pSchemaDB, pKey, kLen, pVal, vLen, NULL);
+    ret = tdbDbInsert(pMetaDb->pSchemaDB, pKey, kLen, pVal, vLen, &pMeta->pDB->txn);
     if (ret < 0) {
       return -1;
     }
@@ -250,7 +265,7 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
   kLen = nameLen + 1 + sizeof(uid);
   pVal = NULL;
   vLen = 0;
-  ret = tdbDbInsert(pMetaDb->pNameIdx, pKey, kLen, pVal, vLen, NULL);
+  ret = tdbDbInsert(pMetaDb->pNameIdx, pKey, kLen, pVal, vLen, &pMetaDb->txn);
   if (ret < 0) {
     return -1;
   }
@@ -261,7 +276,7 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
     kLen = sizeof(uid);
     pVal = NULL;
     vLen = 0;
-    ret = tdbDbInsert(pMetaDb->pStbIdx, pKey, kLen, pVal, vLen, NULL);
+    ret = tdbDbInsert(pMetaDb->pStbIdx, pKey, kLen, pVal, vLen, &pMetaDb->txn);
     if (ret < 0) {
       return -1;
     }
@@ -272,7 +287,7 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
     kLen = sizeof(ctbIdxKey);
     pVal = NULL;
     vLen = 0;
-    ret = tdbDbInsert(pMetaDb->pCtbIdx, pKey, kLen, pVal, vLen, NULL);
+    ret = tdbDbInsert(pMetaDb->pCtbIdx, pKey, kLen, pVal, vLen, &pMetaDb->txn);
     if (ret < 0) {
       return -1;
     }
@@ -281,12 +296,16 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
     kLen = sizeof(uid);
     pVal = NULL;
     vLen = 0;
-    ret = tdbDbInsert(pMetaDb->pNtbIdx, pKey, kLen, pVal, vLen, NULL);
+    ret = tdbDbInsert(pMetaDb->pNtbIdx, pKey, kLen, pVal, vLen, &pMetaDb->txn);
     if (ret < 0) {
       return -1;
     }
   }
 
+  if (pMeta->pDB->pPool->size > 0) {
+    metaCommit(pMeta);
+  }
+
   return 0;
 }
 
@@ -705,6 +724,82 @@ static void *metaDecodeTbInfo(void *buf, STbCfg *pTbCfg) {
 }
 
 int metaCommit(SMeta *pMeta) {
-  tdbCommit(pMeta->pDB->pEnv, NULL);
+  TXN *pTxn = &pMeta->pDB->txn;
+
+  // Commit current txn
+  tdbCommit(pMeta->pDB->pEnv, pTxn);
+  tdbTxnClose(pTxn);
+  clearPool(pMeta->pDB->pPool);
+
+  // start a new txn
+  tdbTxnOpen(&pMeta->pDB->txn, 0, poolMalloc, poolFree, pMeta->pDB->pPool, TDB_TXN_WRITE | TDB_TXN_READ_UNCOMMITTED);
+  tdbBegin(pMeta->pDB->pEnv, pTxn);
   return 0;
-}
\ No newline at end of file
+}
+
+static SPoolMem *openPool() {
+  SPoolMem *pPool = (SPoolMem *)tdbOsMalloc(sizeof(*pPool));
+
+  pPool->prev = pPool->next = pPool;
+  pPool->size = 0;
+
+  return pPool;
+}
+
+static void clearPool(SPoolMem *pPool) {
+  SPoolMem *pMem;
+
+  do {
+    pMem = pPool->next;
+
+    if (pMem == pPool) break;
+
+    pMem->next->prev = pMem->prev;
+    pMem->prev->next = pMem->next;
+    pPool->size -= pMem->size;
+
+    tdbOsFree(pMem);
+  } while (1);
+
+  assert(pPool->size == 0);
+}
+
+static void closePool(SPoolMem *pPool) {
+  clearPool(pPool);
+  tdbOsFree(pPool);
+}
+
+static void *poolMalloc(void *arg, size_t size) {
+  void     *ptr = NULL;
+  SPoolMem *pPool = (SPoolMem *)arg;
+  SPoolMem *pMem;
+
+  pMem = (SPoolMem *)tdbOsMalloc(sizeof(*pMem) + size);
+  if (pMem == NULL) {
+    assert(0);
+  }
+
+  pMem->size = sizeof(*pMem) + size;
+  pMem->next = pPool->next;
+  pMem->prev = pPool;
+
+  pPool->next->prev = pMem;
+  pPool->next = pMem;
+  pPool->size += pMem->size;
+
+  ptr = (void *)(&pMem[1]);
+  return ptr;
+}
+
+static void poolFree(void *arg, void *ptr) {
+  SPoolMem *pPool = (SPoolMem *)arg;
+  SPoolMem *pMem;
+
+  pMem = &(((SPoolMem *)ptr)[-1]);
+
+  pMem->next->prev = pMem->prev;
+  pMem->prev->next = pMem->next;
+  pPool->size -= pMem->size;
+
+  tdbOsFree(pMem);
+}
diff --git a/source/dnode/vnode/src/meta/metaTbUid.c b/source/dnode/vnode/src/meta/metaTbUid.c
index 1f57d1396a..b488630024 100644
--- a/source/dnode/vnode/src/meta/metaTbUid.c
+++ b/source/dnode/vnode/src/meta/metaTbUid.c
@@ -27,5 +27,5 @@ void metaCloseUidGnrt(SMeta *pMeta) { /* TODO */
 
 tb_uid_t metaGenerateUid(SMeta *pMeta) {
   // Generate a new table UID
-  return tGenIdPI32();
+  return tGenIdPI64();
 }
diff --git a/source/dnode/vnode/src/vnd/vnodeCommit.c b/source/dnode/vnode/src/vnd/vnodeCommit.c
index 696c5f39f6..1717721bb2 100644
--- a/source/dnode/vnode/src/vnd/vnodeCommit.c
+++ b/source/dnode/vnode/src/vnd/vnodeCommit.c
@@ -47,7 +47,7 @@ int vnodeSyncCommit(SVnode *pVnode) {
 static int vnodeCommit(void *arg) {
   SVnode *pVnode = (SVnode *)arg;
 
-  metaCommit(pVnode->pMeta);
+  // metaCommit(pVnode->pMeta);
   tqCommit(pVnode->pTq);
   tsdbCommit(pVnode->pTsdb);
 

From 33bcee62c966c8339441b46980ae594ffb2132fb Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Wed, 6 Apr 2022 15:09:29 +0800
Subject: [PATCH 13/25] [td-14393] refactor.

---
 source/client/inc/clientInt.h      | 1 +
 source/client/src/clientImpl.c     | 1 +
 source/client/src/clientMain.c     | 7 ++++++-
 source/client/test/clientTests.cpp | 4 ++--
 4 files changed, 10 insertions(+), 3 deletions(-)

diff --git a/source/client/inc/clientInt.h b/source/client/inc/clientInt.h
index 0749ce9e7d..ae1b34a3bd 100644
--- a/source/client/inc/clientInt.h
+++ b/source/client/inc/clientInt.h
@@ -163,6 +163,7 @@ typedef struct SReqResultInfo {
   uint64_t       totalRows;
   uint32_t       current;
   bool           completed;
+  int32_t        precision;
   int32_t        payloadLen;
 } SReqResultInfo;
 
diff --git a/source/client/src/clientImpl.c b/source/client/src/clientImpl.c
index 80559c025f..4a2687ed36 100644
--- a/source/client/src/clientImpl.c
+++ b/source/client/src/clientImpl.c
@@ -825,6 +825,7 @@ int32_t setQueryResultFromRsp(SReqResultInfo* pResultInfo, const SRetrieveTableR
   pResultInfo->current    = 0;
   pResultInfo->completed  = (pRsp->completed == 1);
   pResultInfo->payloadLen = htonl(pRsp->compLen);
+  pResultInfo->precision  = pRsp->precision;
 
   // TODO handle the compressed case
   pResultInfo->totalRows += pResultInfo->numOfRows;
diff --git a/source/client/src/clientMain.c b/source/client/src/clientMain.c
index ee4fc7fcb5..b2c32a4366 100644
--- a/source/client/src/clientMain.c
+++ b/source/client/src/clientMain.c
@@ -323,7 +323,12 @@ int taos_affected_rows(TAOS_RES *res) {
 }
 
 int taos_result_precision(TAOS_RES *res) {
-  return TSDB_TIME_PRECISION_MILLI;
+  SRequestObj* pRequest = (SRequestObj*) res;
+  if (pRequest == NULL) {
+    return TSDB_TIME_PRECISION_MILLI;
+  }
+
+  return pRequest->body.resInfo.precision;
 }
 
 int taos_select_db(TAOS *taos, const char *db) {
diff --git a/source/client/test/clientTests.cpp b/source/client/test/clientTests.cpp
index 629dd90da3..772f9049e5 100644
--- a/source/client/test/clientTests.cpp
+++ b/source/client/test/clientTests.cpp
@@ -400,6 +400,7 @@ TEST(testCase, show_vgroup_Test) {
   taos_free_result(pRes);
   taos_close(pConn);
 }
+#endif
 
 TEST(testCase, create_multiple_tables) {
   TAOS* pConn = taos_connect("localhost", "root", "taosdata", NULL, 0);
@@ -458,7 +459,7 @@ TEST(testCase, create_multiple_tables) {
 
   taos_free_result(pRes);
 
-  for (int32_t i = 0; i < 20; ++i) {
+  for (int32_t i = 0; i < 25000; ++i) {
     char sql[512] = {0};
     snprintf(sql, tListLen(sql),
              "create table t_x_%d using st1 tags(2) t_x_%d using st1 tags(5) t_x_%d using st1 tags(911)", i,
@@ -652,7 +653,6 @@ TEST(testCase, projection_query_stables) {
   taos_free_result(pRes);
   taos_close(pConn);
 }
-#endif
 
 TEST(testCase, agg_query_tables) {
   TAOS* pConn = taos_connect("localhost", "root", "taosdata", NULL, 0);

From 398a41d09bfdd62af9ba9e98fa7dec4e2927d463 Mon Sep 17 00:00:00 2001
From: dapan1121 <wpan@taosdata.com>
Date: Wed, 6 Apr 2022 15:20:38 +0800
Subject: [PATCH 14/25] feature/qnode

---
 include/common/ttypes.h                 |  1 +
 include/common/tvariant.h               |  1 +
 include/libs/nodes/querynodes.h         |  2 ++
 source/common/src/tvariant.c            | 26 +++++++++++++-
 source/libs/executor/src/executorimpl.c | 22 +++++++++++-
 source/libs/nodes/src/nodesUtilFuncs.c  | 41 +++++++++++++++++++++
 source/libs/parser/src/parTranslater.c  | 48 ++++---------------------
 source/libs/scalar/inc/sclvector.h      |  2 ++
 source/libs/scalar/src/scalar.c         |  3 +-
 source/libs/scalar/src/sclvector.c      | 30 +++++++++++++++-
 10 files changed, 130 insertions(+), 46 deletions(-)

diff --git a/include/common/ttypes.h b/include/common/ttypes.h
index 1cb398fb85..37d688a0ef 100644
--- a/include/common/ttypes.h
+++ b/include/common/ttypes.h
@@ -146,6 +146,7 @@ typedef struct {
 #define IS_FLOAT_TYPE(_t)            ((_t) == TSDB_DATA_TYPE_FLOAT || (_t) == TSDB_DATA_TYPE_DOUBLE)
 
 #define IS_NUMERIC_TYPE(_t) ((IS_SIGNED_NUMERIC_TYPE(_t)) || (IS_UNSIGNED_NUMERIC_TYPE(_t)) || (IS_FLOAT_TYPE(_t)))
+#define IS_MATHABLE_TYPE(_t) (IS_NUMERIC_TYPE(_t) || (_t) == (TSDB_DATA_TYPE_BOOL) || (_t) == (TSDB_DATA_TYPE_TIMESTAMP))
 
 #define IS_VALID_TINYINT(_t)   ((_t) > INT8_MIN && (_t) <= INT8_MAX)
 #define IS_VALID_SMALLINT(_t)  ((_t) > INT16_MIN && (_t) <= INT16_MAX)
diff --git a/include/common/tvariant.h b/include/common/tvariant.h
index 63f305ab2d..83dccd0092 100644
--- a/include/common/tvariant.h
+++ b/include/common/tvariant.h
@@ -59,6 +59,7 @@ int32_t taosVariantDumpEx(SVariant *pVariant, char *payload, int16_t type, bool
 #endif
 
 int32_t taosVariantTypeSetType(SVariant *pVariant, char type);
+char * taosVariantGet(SVariant *pVar, int32_t type);
 
 #ifdef __cplusplus
 }
diff --git a/include/libs/nodes/querynodes.h b/include/libs/nodes/querynodes.h
index d035f7de8f..87aadfd244 100644
--- a/include/libs/nodes/querynodes.h
+++ b/include/libs/nodes/querynodes.h
@@ -22,6 +22,7 @@ extern "C" {
 
 #include "nodes.h"
 #include "tmsg.h"
+#include "tvariant.h"
 
 #define TABLE_TOTAL_COL_NUM(pMeta) ((pMeta)->tableInfo.numOfColumns + (pMeta)->tableInfo.numOfTags)
 #define TABLE_META_SIZE(pMeta) (NULL == (pMeta) ? 0 : (sizeof(STableMeta) + TABLE_TOTAL_COL_NUM((pMeta)) * sizeof(SSchema)))
@@ -317,6 +318,7 @@ bool nodesIsTimelineQuery(const SNode* pQuery);
 void* nodesGetValueFromNode(SValueNode *pNode);
 char* nodesGetStrValueFromNode(SValueNode *pNode);
 char *getFillModeString(EFillMode mode);
+void valueNodeToVariant(const SValueNode* pNode, SVariant* pVal);
 
 #ifdef __cplusplus
 }
diff --git a/source/common/src/tvariant.c b/source/common/src/tvariant.c
index 3995db89b6..77feb04f6c 100644
--- a/source/common/src/tvariant.c
+++ b/source/common/src/tvariant.c
@@ -1014,4 +1014,28 @@ int32_t taosVariantTypeSetType(SVariant *pVariant, char type) {
   }
 
   return 0;
-}
\ No newline at end of file
+}
+
+char * taosVariantGet(SVariant *pVar, int32_t type) {
+  switch (type) {
+    case TSDB_DATA_TYPE_BOOL: 
+    case TSDB_DATA_TYPE_TINYINT:
+    case TSDB_DATA_TYPE_SMALLINT:
+    case TSDB_DATA_TYPE_BIGINT:
+    case TSDB_DATA_TYPE_INT: 
+    case TSDB_DATA_TYPE_TIMESTAMP:
+      return (char *)&pVar->i;
+    case TSDB_DATA_TYPE_DOUBLE:
+    case TSDB_DATA_TYPE_FLOAT:
+      return (char *)&pVar->d;
+    case TSDB_DATA_TYPE_BINARY:
+      return (char *)pVar->pz;
+    case TSDB_DATA_TYPE_NCHAR:
+      return (char *)pVar->ucs4;
+    default: 
+      return NULL;
+  }
+
+  return NULL;
+}
+
diff --git a/source/libs/executor/src/executorimpl.c b/source/libs/executor/src/executorimpl.c
index a1642fc2c8..bb2432c07a 100644
--- a/source/libs/executor/src/executorimpl.c
+++ b/source/libs/executor/src/executorimpl.c
@@ -1289,6 +1289,12 @@ static void projectApplyFunctions(SExprInfo* pExpr, SSDataBlock* pResult, SSData
 
       taosArrayDestroy(pBlockList);
 
+    } else if (pExpr[k].pExpr->nodeType == QUERY_NODE_VALUE) {
+      SColumnInfoData* pColInfoData = taosArrayGet(pResult->pDataBlock, k);
+      for (int32_t i = 0; i < pSrcBlock->info.rows; ++i) {
+        colDataAppend(pColInfoData, i, taosVariantGet(&pExpr[k].base.pParam[0].param, pExpr[k].base.pParam[0].type), TSDB_DATA_TYPE_NULL == pExpr[k].base.pParam[0].param.nType);
+      }
+      pResult->info.rows = pSrcBlock->info.rows;
     } else {
       ASSERT(0);
     }
@@ -2093,7 +2099,7 @@ static SqlFunctionCtx* createSqlFunctionCtx_rv(SExprInfo* pExprInfo, int32_t num
         }
       }
       pCtx->resDataInfo.interBufSize = env.calcMemSize;
-    } else if (pExpr->pExpr->nodeType == QUERY_NODE_COLUMN || pExpr->pExpr->nodeType == QUERY_NODE_OPERATOR) {
+    } else if (pExpr->pExpr->nodeType == QUERY_NODE_COLUMN || pExpr->pExpr->nodeType == QUERY_NODE_OPERATOR || pExpr->pExpr->nodeType == QUERY_NODE_VALUE) {
       pCtx->resDataInfo.interBufSize = pFunct->resSchema.bytes; // for simple column, the intermediate buffer needs to hold one element.
     }
 
@@ -2864,6 +2870,9 @@ int32_t loadDataBlock(SExecTaskInfo* pTaskInfo, STableScanInfo* pTableScanInfo,
   if (pTableScanInfo->pFilterNode != NULL) {
     SFilterInfo* filter = NULL;
     int32_t      code = filterInitFromNode((SNode*)pTableScanInfo->pFilterNode, &filter, 0);
+    if (code) {
+      return code;
+    }
 
     SFilterColumnParam param1 = {.numOfCols = pBlock->info.numOfCols, .pDataBlock = pBlock->pDataBlock};
     code = filterSetDataFromSlotId(filter, &param1);
@@ -8309,6 +8318,17 @@ SExprInfo* createExprInfo(SNodeList* pNodeList, SNodeList* pGroupKeys, int32_t*
 
 //      pExp->base.pParam[0].type = FUNC_PARAM_TYPE_COLUMN;
 //      pExp->base.pParam[0].pCol = createColumn(pTargetNode->dataBlockId, pTargetNode->slotId, pType);
+    } else if (nodeType(pTargetNode->pExpr) == QUERY_NODE_VALUE) {
+      pExp->pExpr->nodeType = QUERY_NODE_VALUE;
+      SValueNode* pValNode = (SValueNode*)pTargetNode->pExpr;
+
+      pExp->base.pParam = taosMemoryCalloc(1, sizeof(SFunctParam));
+      pExp->base.numOfParams = 1;
+
+      SDataType* pType = &pValNode->node.resType;
+      pExp->base.resSchema = createResSchema(pType->type, pType->bytes, pTargetNode->slotId, pType->scale, pType->precision, pValNode->node.aliasName);
+      pExp->base.pParam[0].type = FUNC_PARAM_TYPE_VALUE;
+      valueNodeToVariant(pValNode, &pExp->base.pParam[0].param);
     } else {
       ASSERT(0);
     }
diff --git a/source/libs/nodes/src/nodesUtilFuncs.c b/source/libs/nodes/src/nodesUtilFuncs.c
index cfee083f46..6bf2c99d0d 100644
--- a/source/libs/nodes/src/nodesUtilFuncs.c
+++ b/source/libs/nodes/src/nodesUtilFuncs.c
@@ -1092,4 +1092,45 @@ int32_t nodesGetOutputNumFromSlotList(SNodeList* pSlots) {
 }
 
 
+void valueNodeToVariant(const SValueNode* pNode, SVariant* pVal) {
+  pVal->nType = pNode->node.resType.type;
+  pVal->nLen = pNode->node.resType.bytes;
+  switch (pNode->node.resType.type) {
+    case TSDB_DATA_TYPE_NULL:
+        break;
+    case TSDB_DATA_TYPE_BOOL:
+      pVal->i = pNode->datum.b;
+      break;
+    case TSDB_DATA_TYPE_TINYINT:
+    case TSDB_DATA_TYPE_SMALLINT:
+    case TSDB_DATA_TYPE_INT:
+    case TSDB_DATA_TYPE_BIGINT:
+    case TSDB_DATA_TYPE_TIMESTAMP:
+      pVal->i = pNode->datum.i;
+      break;
+    case TSDB_DATA_TYPE_UTINYINT:
+    case TSDB_DATA_TYPE_USMALLINT:
+    case TSDB_DATA_TYPE_UINT:
+    case TSDB_DATA_TYPE_UBIGINT:
+      pVal->u = pNode->datum.u;
+      break;
+    case TSDB_DATA_TYPE_FLOAT:
+    case TSDB_DATA_TYPE_DOUBLE:
+      pVal->d = pNode->datum.d;
+      break;
+    case TSDB_DATA_TYPE_NCHAR:
+    case TSDB_DATA_TYPE_VARCHAR:
+    case TSDB_DATA_TYPE_VARBINARY:
+      pVal->pz = pNode->datum.p;
+      break;
+    case TSDB_DATA_TYPE_JSON:
+    case TSDB_DATA_TYPE_DECIMAL:
+    case TSDB_DATA_TYPE_BLOB:
+      // todo
+    default:
+      break;
+  }
+}
+
+
 
diff --git a/source/libs/parser/src/parTranslater.c b/source/libs/parser/src/parTranslater.c
index bf909a0eaf..97fce407b6 100644
--- a/source/libs/parser/src/parTranslater.c
+++ b/source/libs/parser/src/parTranslater.c
@@ -438,8 +438,12 @@ static EDealRes translateValue(STranslateContext* pCxt, SValueNode* pVal) {
 
 static EDealRes translateOperator(STranslateContext* pCxt, SOperatorNode* pOp) {
   if (nodesIsUnaryOp(pOp)) {
-    if (OP_TYPE_MINUS == pOp->opType && !IS_NUMERIC_TYPE(((SExprNode*)(pOp->pLeft))->resType.type)) {
-      return generateDealNodeErrMsg(pCxt, TSDB_CODE_PAR_WRONG_VALUE_TYPE, ((SExprNode*)(pOp->pLeft))->aliasName);
+    if (OP_TYPE_MINUS == pOp->opType) {
+      if (!IS_MATHABLE_TYPE(((SExprNode*)(pOp->pLeft))->resType.type)) {
+        return generateDealNodeErrMsg(pCxt, TSDB_CODE_PAR_WRONG_VALUE_TYPE, ((SExprNode*)(pOp->pLeft))->aliasName);
+      }
+      pOp->node.resType.type = TSDB_DATA_TYPE_DOUBLE;
+      pOp->node.resType.bytes = tDataTypes[TSDB_DATA_TYPE_DOUBLE].bytes;
     }
     return DEAL_RES_CONTINUE;
   }
@@ -2338,46 +2342,6 @@ static void addCreateTbReqIntoVgroup(int32_t acctId, SHashObj* pVgroupHashmap,
   }
 }
 
-static void valueNodeToVariant(const SValueNode* pNode, SVariant* pVal) {
-  pVal->nType = pNode->node.resType.type;
-  pVal->nLen = pNode->node.resType.bytes;
-  switch (pNode->node.resType.type) {
-    case TSDB_DATA_TYPE_NULL:
-        break;
-    case TSDB_DATA_TYPE_BOOL:
-      pVal->i = pNode->datum.b;
-      break;
-    case TSDB_DATA_TYPE_TINYINT:
-    case TSDB_DATA_TYPE_SMALLINT:
-    case TSDB_DATA_TYPE_INT:
-    case TSDB_DATA_TYPE_BIGINT:
-    case TSDB_DATA_TYPE_TIMESTAMP:
-      pVal->i = pNode->datum.i;
-      break;
-    case TSDB_DATA_TYPE_UTINYINT:
-    case TSDB_DATA_TYPE_USMALLINT:
-    case TSDB_DATA_TYPE_UINT:
-    case TSDB_DATA_TYPE_UBIGINT:
-      pVal->u = pNode->datum.u;
-      break;
-    case TSDB_DATA_TYPE_FLOAT:
-    case TSDB_DATA_TYPE_DOUBLE:
-      pVal->d = pNode->datum.d;
-      break;
-    case TSDB_DATA_TYPE_NCHAR:
-    case TSDB_DATA_TYPE_VARCHAR:
-    case TSDB_DATA_TYPE_VARBINARY:
-      pVal->pz = pNode->datum.p;
-      break;
-    case TSDB_DATA_TYPE_JSON:
-    case TSDB_DATA_TYPE_DECIMAL:
-    case TSDB_DATA_TYPE_BLOB:
-      // todo
-    default:
-      break;
-  }
-}
-
 static int32_t addValToKVRow(STranslateContext* pCxt, SValueNode* pVal, const SSchema* pSchema, SKVRowBuilder* pBuilder) {
   if (DEAL_RES_ERROR == translateValue(pCxt, pVal)) {
     return pCxt->errCode;
diff --git a/source/libs/scalar/inc/sclvector.h b/source/libs/scalar/inc/sclvector.h
index 3f41ad875c..e51115e9c0 100644
--- a/source/libs/scalar/inc/sclvector.h
+++ b/source/libs/scalar/inc/sclvector.h
@@ -75,6 +75,8 @@ static FORCE_INLINE _getDoubleValue_fn_t getVectorDoubleValueFn(int32_t srcType)
     p = getVectorDoubleValue_FLOAT;
   } else if (srcType == TSDB_DATA_TYPE_DOUBLE) {
     p = getVectorDoubleValue_DOUBLE;
+  } else if (srcType == TSDB_DATA_TYPE_TIMESTAMP) {
+    p = getVectorDoubleValue_BIGINT;
   } else {
     assert(0);
   }
diff --git a/source/libs/scalar/src/scalar.c b/source/libs/scalar/src/scalar.c
index d77ada3b4e..07358f905f 100644
--- a/source/libs/scalar/src/scalar.c
+++ b/source/libs/scalar/src/scalar.c
@@ -10,7 +10,8 @@
 
 int32_t scalarGetOperatorParamNum(EOperatorType type) {
   if (OP_TYPE_IS_NULL == type || OP_TYPE_IS_NOT_NULL == type || OP_TYPE_IS_TRUE == type || OP_TYPE_IS_NOT_TRUE == type 
-   || OP_TYPE_IS_FALSE == type || OP_TYPE_IS_NOT_FALSE == type || OP_TYPE_IS_UNKNOWN == type || OP_TYPE_IS_NOT_UNKNOWN == type) {
+   || OP_TYPE_IS_FALSE == type || OP_TYPE_IS_NOT_FALSE == type || OP_TYPE_IS_UNKNOWN == type || OP_TYPE_IS_NOT_UNKNOWN == type
+   || OP_TYPE_MINUS == type) {
     return 1;
   }
 
diff --git a/source/libs/scalar/src/sclvector.c b/source/libs/scalar/src/sclvector.c
index ef3be80893..c5f18fb1a9 100644
--- a/source/libs/scalar/src/sclvector.c
+++ b/source/libs/scalar/src/sclvector.c
@@ -773,6 +773,32 @@ void vectorMathRemainder(SScalarParam* pLeft, SScalarParam* pRight, SScalarParam
   doReleaseVec(pRightCol, rightConvert);
 }
 
+void vectorMathMinus(SScalarParam* pLeft, SScalarParam* pRight, SScalarParam *pOut, int32_t _ord) {
+  SColumnInfoData *pOutputCol = pOut->columnData;
+
+  pOut->numOfRows = pLeft->numOfRows;
+
+  int32_t i = ((_ord) == TSDB_ORDER_ASC)? 0 : (pLeft->numOfRows - 1);
+  int32_t step = ((_ord) == TSDB_ORDER_ASC)? 1 : -1;
+
+  int32_t leftConvert = 0;
+  SColumnInfoData *pLeftCol   = doVectorConvert(pLeft, &leftConvert);
+
+  _getDoubleValue_fn_t getVectorDoubleValueFnLeft  = getVectorDoubleValueFn(pLeftCol->info.type);
+
+  double *output = (double *)pOutputCol->pData;
+  for (; i < pLeft->numOfRows && i >= 0; i += step, output += 1) {
+    *output = - getVectorDoubleValueFnLeft(pLeftCol->pData, i);
+  }
+
+  pOutputCol->hasNull = pLeftCol->hasNull;
+  if (pOutputCol->hasNull) {
+    memcpy(pOutputCol->nullbitmap, pLeftCol->nullbitmap, BitmapLen(pLeft->numOfRows));
+  }
+
+  doReleaseVec(pLeftCol,  leftConvert);
+}
+
 void vectorConcat(SScalarParam* pLeft, SScalarParam* pRight, void *out, int32_t _ord) {
 #if 0
   int32_t len = pLeft->bytes + pRight->bytes;
@@ -1102,6 +1128,8 @@ _bin_scalar_fn_t getBinScalarOperatorFn(int32_t binFunctionId) {
       return vectorMathDivide;
     case OP_TYPE_MOD:
       return vectorMathRemainder;
+    case OP_TYPE_MINUS:
+      return vectorMathMinus;
     case OP_TYPE_GREATER_THAN:
       return vectorGreater;
     case OP_TYPE_GREATER_EQUAL:
@@ -1140,4 +1168,4 @@ _bin_scalar_fn_t getBinScalarOperatorFn(int32_t binFunctionId) {
       assert(0);
       return NULL;
   }
-}
\ No newline at end of file
+}

From 709dda39b7ea2f4e19ac690003b8055f0bee9247 Mon Sep 17 00:00:00 2001
From: Xiaoyu Wang <xiaoyuwang@taosdata.com>
Date: Wed, 6 Apr 2022 03:22:41 -0400
Subject: [PATCH 15/25] TD-13037 distinct, state window and partition by plan
 implement

---
 include/libs/nodes/nodes.h                 |   2 +
 include/libs/nodes/plannodes.h             |  26 +
 include/libs/nodes/querynodes.h            |   2 +-
 include/util/taoserror.h                   |   3 +
 source/libs/nodes/src/nodesCloneFuncs.c    |   8 +
 source/libs/nodes/src/nodesCodeFuncs.c     |  67 +++
 source/libs/nodes/src/nodesTraverseFuncs.c |   7 +-
 source/libs/nodes/src/nodesUtilFuncs.c     |   8 +-
 source/libs/parser/inc/parAst.h            |   2 +-
 source/libs/parser/inc/sql.y               |   2 +-
 source/libs/parser/src/parAstCreater.c     |   4 +-
 source/libs/parser/src/sql.c               | 664 +++++++++++----------
 source/libs/planner/src/planLogicCreater.c |  67 ++-
 source/libs/planner/src/planPhysiCreater.c | 114 +++-
 source/libs/planner/test/plannerTest.cpp   |  39 ++
 source/util/src/terror.c                   |   3 +
 16 files changed, 676 insertions(+), 342 deletions(-)

diff --git a/include/libs/nodes/nodes.h b/include/libs/nodes/nodes.h
index d2217f7e01..1d7bd09175 100644
--- a/include/libs/nodes/nodes.h
+++ b/include/libs/nodes/nodes.h
@@ -163,6 +163,8 @@ typedef enum ENodeType {
   QUERY_NODE_PHYSICAL_PLAN_SORT,
   QUERY_NODE_PHYSICAL_PLAN_INTERVAL,
   QUERY_NODE_PHYSICAL_PLAN_SESSION_WINDOW,
+  QUERY_NODE_PHYSICAL_PLAN_STATE_WINDOW,
+  QUERY_NODE_PHYSICAL_PLAN_PARTITION,
   QUERY_NODE_PHYSICAL_PLAN_DISPATCH,
   QUERY_NODE_PHYSICAL_PLAN_INSERT,
   QUERY_NODE_PHYSICAL_SUBPLAN,
diff --git a/include/libs/nodes/plannodes.h b/include/libs/nodes/plannodes.h
index 37163f60dd..9f8cddfd54 100644
--- a/include/libs/nodes/plannodes.h
+++ b/include/libs/nodes/plannodes.h
@@ -104,6 +104,7 @@ typedef struct SWindowLogicNode {
   SFillNode* pFill;
   int64_t sessionGap;
   SNode* pTspk;
+  SNode* pStateExpr;
 } SWindowLogicNode;
 
 typedef struct SSortLogicNode {
@@ -194,11 +195,20 @@ typedef struct SSystemTableScanPhysiNode {
   int32_t accountId;
 } SSystemTableScanPhysiNode;
 
+typedef enum EScanRequired {
+  SCAN_REQUIRED_DATA_NO_NEEDED = 1,
+  SCAN_REQUIRED_DATA_STATIS_NEEDED,
+  SCAN_REQUIRED_DATA_ALL_NEEDED,
+  SCAN_REQUIRED_DATA_DISCARD,
+} EScanRequired;
+
 typedef struct STableScanPhysiNode {
   SScanPhysiNode scan;
   uint8_t scanFlag;         // denotes reversed scan of data or not
   STimeWindow scanRange;
   double ratio;
+  EScanRequired scanRequired;
+  SNodeList* pScanReferFuncs;
 } STableScanPhysiNode;
 
 typedef STableScanPhysiNode STableSeqScanPhysiNode;
@@ -257,17 +267,33 @@ typedef struct SIntervalPhysiNode {
   SFillNode* pFill;
 } SIntervalPhysiNode;
 
+typedef struct SMultiTableIntervalPhysiNode {
+  SIntervalPhysiNode interval;
+  SNodeList* pPartitionKeys;
+} SMultiTableIntervalPhysiNode;
+
 typedef struct SSessionWinodwPhysiNode {
   SWinodwPhysiNode window;
   int64_t    gap;
 } SSessionWinodwPhysiNode;
 
+typedef struct SStateWinodwPhysiNode {
+  SWinodwPhysiNode window;
+  SNode* pStateKey;
+} SStateWinodwPhysiNode;
+
 typedef struct SSortPhysiNode {
   SPhysiNode node;
   SNodeList* pExprs;   // these are expression list of order_by_clause and parameter expression of aggregate function
   SNodeList* pSortKeys; // element is SOrderByExprNode, and SOrderByExprNode::pExpr is SColumnNode
 } SSortPhysiNode;
 
+typedef struct SPartitionPhysiNode {
+  SPhysiNode node;
+  SNodeList* pExprs;   // these are expression list of partition_by_clause
+  SNodeList* pPartitionKeys;
+} SPartitionPhysiNode;
+
 typedef struct SDataSinkNode {
   ENodeType type;
   SDataBlockDescNode* pInputDataBlockDesc;
diff --git a/include/libs/nodes/querynodes.h b/include/libs/nodes/querynodes.h
index a8ff48072f..82d05853c4 100644
--- a/include/libs/nodes/querynodes.h
+++ b/include/libs/nodes/querynodes.h
@@ -188,7 +188,7 @@ typedef struct SLimitNode {
 
 typedef struct SStateWindowNode {
   ENodeType type; // QUERY_NODE_STATE_WINDOW
-  SNode* pCol;
+  SNode* pExpr;
 } SStateWindowNode;
 
 typedef struct SSessionWindowNode {
diff --git a/include/util/taoserror.h b/include/util/taoserror.h
index 4e2cb7944a..e7f440e26f 100644
--- a/include/util/taoserror.h
+++ b/include/util/taoserror.h
@@ -483,6 +483,9 @@ int32_t* taosGetErrno();
 #define TSDB_CODE_PAR_INVALID_IDENTIFIER_NAME   TAOS_DEF_ERROR_CODE(0, 0x2617)
 #define TSDB_CODE_PAR_CORRESPONDING_STABLE_ERR  TAOS_DEF_ERROR_CODE(0, 0x2618)
 
+//planner
+#define TSDB_CODE_PLAN_INTERNAL_ERROR              TAOS_DEF_ERROR_CODE(0, 0x2700)
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/source/libs/nodes/src/nodesCloneFuncs.c b/source/libs/nodes/src/nodesCloneFuncs.c
index ac78988900..12fe645473 100644
--- a/source/libs/nodes/src/nodesCloneFuncs.c
+++ b/source/libs/nodes/src/nodesCloneFuncs.c
@@ -283,6 +283,12 @@ static SNode* logicSortCopy(const SSortLogicNode* pSrc, SSortLogicNode* pDst) {
   return (SNode*)pDst;
 }
 
+static SNode* logicPartitionCopy(const SPartitionLogicNode* pSrc, SPartitionLogicNode* pDst) {
+  COPY_BASE_OBJECT_FIELD(node, logicNodeCopy);
+  CLONE_NODE_LIST_FIELD(pPartitionKeys);
+  return (SNode*)pDst;
+}
+
 static SNode* logicSubplanCopy(const SLogicSubplan* pSrc, SLogicSubplan* pDst) {
   CLONE_NODE_FIELD(pNode);
   COPY_SCALAR_FIELD(subplanType);
@@ -367,6 +373,8 @@ SNodeptr nodesCloneNode(const SNodeptr pNode) {
       return logicWindowCopy((const SWindowLogicNode*)pNode, (SWindowLogicNode*)pDst);
     case QUERY_NODE_LOGIC_PLAN_SORT:
       return logicSortCopy((const SSortLogicNode*)pNode, (SSortLogicNode*)pDst);
+    case QUERY_NODE_LOGIC_PLAN_PARTITION:
+      return logicPartitionCopy((const SPartitionLogicNode*)pNode, (SPartitionLogicNode*)pDst);
     case QUERY_NODE_LOGIC_SUBPLAN:
       return logicSubplanCopy((const SLogicSubplan*)pNode, (SLogicSubplan*)pDst);
     default:
diff --git a/source/libs/nodes/src/nodesCodeFuncs.c b/source/libs/nodes/src/nodesCodeFuncs.c
index 1b16ba9206..6cb1955307 100644
--- a/source/libs/nodes/src/nodesCodeFuncs.c
+++ b/source/libs/nodes/src/nodesCodeFuncs.c
@@ -198,6 +198,10 @@ const char* nodesNodeName(ENodeType type) {
       return "PhysiInterval";
     case QUERY_NODE_PHYSICAL_PLAN_SESSION_WINDOW:
       return "PhysiSessionWindow";
+    case QUERY_NODE_PHYSICAL_PLAN_STATE_WINDOW:
+      return "PhysiStateWindow";
+    case QUERY_NODE_PHYSICAL_PLAN_PARTITION:
+      return "PhysiPartition";
     case QUERY_NODE_PHYSICAL_PLAN_DISPATCH:
       return "PhysiDispatch";
     case QUERY_NODE_PHYSICAL_PLAN_INSERT:
@@ -1147,6 +1151,61 @@ static int32_t jsonToPhysiSessionWindowNode(const SJson* pJson, void* pObj) {
   return code;
 }
 
+static const char* jkStateWindowPhysiPlanStateKey = "StateKey";
+
+static int32_t physiStateWindowNodeToJson(const void* pObj, SJson* pJson) {
+  const SStateWinodwPhysiNode* pNode = (const SStateWinodwPhysiNode*)pObj;
+
+  int32_t code = physiWindowNodeToJson(pObj, pJson);
+  if (TSDB_CODE_SUCCESS == code) {
+    code = tjsonAddObject(pJson, jkStateWindowPhysiPlanStateKey, nodeToJson, pNode->pStateKey);
+  }
+
+  return code;
+}
+
+static int32_t jsonToPhysiStateWindowNode(const SJson* pJson, void* pObj) {
+  SStateWinodwPhysiNode* pNode = (SStateWinodwPhysiNode*)pObj;
+
+  int32_t code = jsonToPhysiWindowNode(pJson, pObj);
+  if (TSDB_CODE_SUCCESS == code) {
+    code = jsonToNodeObject(pJson, jkStateWindowPhysiPlanStateKey, &pNode->pStateKey);
+  }
+
+  return code;
+}
+
+static const char* jkPartitionPhysiPlanExprs = "Exprs";
+static const char* jkPartitionPhysiPlanPartitionKeys = "PartitionKeys";
+
+static int32_t physiPartitionNodeToJson(const void* pObj, SJson* pJson) {
+  const SPartitionPhysiNode* pNode = (const SPartitionPhysiNode*)pObj;
+
+  int32_t code = physicPlanNodeToJson(pObj, pJson);
+  if (TSDB_CODE_SUCCESS == code) {
+    code = nodeListToJson(pJson, jkPartitionPhysiPlanExprs, pNode->pExprs);
+  }
+  if (TSDB_CODE_SUCCESS == code) {
+    code = nodeListToJson(pJson, jkPartitionPhysiPlanPartitionKeys, pNode->pPartitionKeys);
+  }
+
+  return code;
+}
+
+static int32_t jsonToPhysiPartitionNode(const SJson* pJson, void* pObj) {
+  SPartitionPhysiNode* pNode = (SPartitionPhysiNode*)pObj;
+
+  int32_t code = jsonToPhysicPlanNode(pJson, pObj);
+  if (TSDB_CODE_SUCCESS == code) {
+    code = jsonToNodeList(pJson, jkPartitionPhysiPlanExprs, &pNode->pExprs);
+  }
+  if (TSDB_CODE_SUCCESS == code) {
+    code = jsonToNodeList(pJson, jkPartitionPhysiPlanPartitionKeys, &pNode->pPartitionKeys);
+  }
+
+  return code;
+}
+
 static const char* jkDataSinkInputDataBlockDesc = "InputDataBlockDesc";
 
 static int32_t physicDataSinkNodeToJson(const void* pObj, SJson* pJson) {
@@ -2420,6 +2479,10 @@ static int32_t specificNodeToJson(const void* pObj, SJson* pJson) {
       return physiIntervalNodeToJson(pObj, pJson);
     case QUERY_NODE_PHYSICAL_PLAN_SESSION_WINDOW:
       return physiSessionWindowNodeToJson(pObj, pJson);
+    case QUERY_NODE_PHYSICAL_PLAN_STATE_WINDOW:
+      return physiStateWindowNodeToJson(pObj, pJson);
+    case QUERY_NODE_PHYSICAL_PLAN_PARTITION:
+      return physiPartitionNodeToJson(pObj, pJson);
     case QUERY_NODE_PHYSICAL_PLAN_DISPATCH:
       return physiDispatchNodeToJson(pObj, pJson);
     case QUERY_NODE_PHYSICAL_PLAN_INSERT:
@@ -2512,6 +2575,10 @@ static int32_t jsonToSpecificNode(const SJson* pJson, void* pObj) {
       return jsonToPhysiIntervalNode(pJson, pObj);
     case QUERY_NODE_PHYSICAL_PLAN_SESSION_WINDOW:
       return jsonToPhysiSessionWindowNode(pJson, pObj);
+    case QUERY_NODE_PHYSICAL_PLAN_STATE_WINDOW:
+      return jsonToPhysiStateWindowNode(pJson, pObj);
+    case QUERY_NODE_PHYSICAL_PLAN_PARTITION:
+      return jsonToPhysiPartitionNode(pJson, pObj);
     case QUERY_NODE_PHYSICAL_PLAN_DISPATCH:
       return jsonToPhysiDispatchNode(pJson, pObj);
     case QUERY_NODE_PHYSICAL_SUBPLAN:
diff --git a/source/libs/nodes/src/nodesTraverseFuncs.c b/source/libs/nodes/src/nodesTraverseFuncs.c
index 15c078d8d2..e59a76ca3c 100644
--- a/source/libs/nodes/src/nodesTraverseFuncs.c
+++ b/source/libs/nodes/src/nodesTraverseFuncs.c
@@ -78,7 +78,7 @@ static EDealRes walkNode(SNode* pNode, ETraversalOrder order, FNodeWalker walker
       res = walkNode(((SOrderByExprNode*)pNode)->pExpr, order, walker, pContext);
       break;
     case QUERY_NODE_STATE_WINDOW:
-      res = walkNode(((SStateWindowNode*)pNode)->pCol, order, walker, pContext);
+      res = walkNode(((SStateWindowNode*)pNode)->pExpr, order, walker, pContext);
       break;
     case QUERY_NODE_SESSION_WINDOW: {
       SSessionWindowNode* pSession = (SSessionWindowNode*)pNode;
@@ -212,7 +212,7 @@ static EDealRes rewriteNode(SNode** pRawNode, ETraversalOrder order, FNodeRewrit
       res = rewriteNode(&(((SOrderByExprNode*)pNode)->pExpr), order, rewriter, pContext);
       break;
     case QUERY_NODE_STATE_WINDOW:
-      res = rewriteNode(&(((SStateWindowNode*)pNode)->pCol), order, rewriter, pContext);
+      res = rewriteNode(&(((SStateWindowNode*)pNode)->pExpr), order, rewriter, pContext);
       break;
     case QUERY_NODE_SESSION_WINDOW:
       res = rewriteNode(&(((SSessionWindowNode*)pNode)->pCol), order, rewriter, pContext);
@@ -301,10 +301,9 @@ void nodesWalkSelectStmt(SSelectStmt* pSelect, ESqlClause clause, FNodeWalker wa
     case SQL_CLAUSE_GROUP_BY:
       nodesWalkExpr(pSelect->pHaving, walker, pContext);
     case SQL_CLAUSE_HAVING:
-      nodesWalkExprs(pSelect->pOrderByList, walker, pContext);
-    case SQL_CLAUSE_ORDER_BY:
       nodesWalkExprs(pSelect->pProjectionList, walker, pContext);
     case SQL_CLAUSE_SELECT:
+      nodesWalkExprs(pSelect->pOrderByList, walker, pContext);
     default:
       break;
   }
diff --git a/source/libs/nodes/src/nodesUtilFuncs.c b/source/libs/nodes/src/nodesUtilFuncs.c
index 90f06c6c9c..9bb603e479 100644
--- a/source/libs/nodes/src/nodesUtilFuncs.c
+++ b/source/libs/nodes/src/nodesUtilFuncs.c
@@ -169,6 +169,8 @@ SNodeptr nodesMakeNode(ENodeType type) {
       return makeNode(type, sizeof(SWindowLogicNode));
     case QUERY_NODE_LOGIC_PLAN_SORT:
       return makeNode(type, sizeof(SSortLogicNode));
+    case QUERY_NODE_LOGIC_PLAN_PARTITION:
+      return makeNode(type, sizeof(SPartitionLogicNode));
     case QUERY_NODE_LOGIC_SUBPLAN:
       return makeNode(type, sizeof(SLogicSubplan));
     case QUERY_NODE_LOGIC_PLAN:
@@ -197,6 +199,10 @@ SNodeptr nodesMakeNode(ENodeType type) {
       return makeNode(type, sizeof(SIntervalPhysiNode));
     case QUERY_NODE_PHYSICAL_PLAN_SESSION_WINDOW:
       return makeNode(type, sizeof(SSessionWinodwPhysiNode));
+    case QUERY_NODE_PHYSICAL_PLAN_STATE_WINDOW:
+      return makeNode(type, sizeof(SStateWinodwPhysiNode));
+    case QUERY_NODE_PHYSICAL_PLAN_PARTITION:
+      return makeNode(type, sizeof(SPartitionPhysiNode));
     case QUERY_NODE_PHYSICAL_PLAN_DISPATCH:
       return makeNode(type, sizeof(SDataDispatcherNode));
     case QUERY_NODE_PHYSICAL_PLAN_INSERT:
@@ -302,7 +308,7 @@ void nodesDestroyNode(SNodeptr pNode) {
     case QUERY_NODE_LIMIT: // no pointer field
       break;
     case QUERY_NODE_STATE_WINDOW:
-      nodesDestroyNode(((SStateWindowNode*)pNode)->pCol);
+      nodesDestroyNode(((SStateWindowNode*)pNode)->pExpr);
       break;
     case QUERY_NODE_SESSION_WINDOW: {
       SSessionWindowNode* pSession = (SSessionWindowNode*)pNode;
diff --git a/source/libs/parser/inc/parAst.h b/source/libs/parser/inc/parAst.h
index 1fce11acdf..2fe325f1c5 100644
--- a/source/libs/parser/inc/parAst.h
+++ b/source/libs/parser/inc/parAst.h
@@ -103,7 +103,7 @@ SNode* createJoinTableNode(SAstCreateContext* pCxt, EJoinType type, SNode* pLeft
 SNode* createLimitNode(SAstCreateContext* pCxt, const SToken* pLimit, const SToken* pOffset);
 SNode* createOrderByExprNode(SAstCreateContext* pCxt, SNode* pExpr, EOrder order, ENullOrder nullOrder);
 SNode* createSessionWindowNode(SAstCreateContext* pCxt, SNode* pCol, SNode* pGap);
-SNode* createStateWindowNode(SAstCreateContext* pCxt, SNode* pCol);
+SNode* createStateWindowNode(SAstCreateContext* pCxt, SNode* pExpr);
 SNode* createIntervalWindowNode(SAstCreateContext* pCxt, SNode* pInterval, SNode* pOffset, SNode* pSliding, SNode* pFill);
 SNode* createFillNode(SAstCreateContext* pCxt, EFillMode mode, SNode* pValues);
 SNode* createGroupingSetNode(SAstCreateContext* pCxt, SNode* pNode);
diff --git a/source/libs/parser/inc/sql.y b/source/libs/parser/inc/sql.y
index 4ae33e5c5c..c30a75aa66 100644
--- a/source/libs/parser/inc/sql.y
+++ b/source/libs/parser/inc/sql.y
@@ -716,7 +716,7 @@ partition_by_clause_opt(A) ::= PARTITION BY expression_list(B).
 twindow_clause_opt(A) ::= .                                                       { A = NULL; }
 twindow_clause_opt(A) ::=
   SESSION NK_LP column_reference(B) NK_COMMA duration_literal(C) NK_RP.           { A = createSessionWindowNode(pCxt, releaseRawExprNode(pCxt, B), releaseRawExprNode(pCxt, C)); }
-twindow_clause_opt(A) ::= STATE_WINDOW NK_LP column_reference(B) NK_RP.           { A = createStateWindowNode(pCxt, releaseRawExprNode(pCxt, B)); }
+twindow_clause_opt(A) ::= STATE_WINDOW NK_LP expression(B) NK_RP.                 { A = createStateWindowNode(pCxt, releaseRawExprNode(pCxt, B)); }
 twindow_clause_opt(A) ::=
   INTERVAL NK_LP duration_literal(B) NK_RP sliding_opt(C) fill_opt(D).            { A = createIntervalWindowNode(pCxt, releaseRawExprNode(pCxt, B), NULL, C, D); }
 twindow_clause_opt(A) ::=
diff --git a/source/libs/parser/src/parAstCreater.c b/source/libs/parser/src/parAstCreater.c
index 07316ddd39..ebcecfe3cb 100644
--- a/source/libs/parser/src/parAstCreater.c
+++ b/source/libs/parser/src/parAstCreater.c
@@ -730,10 +730,10 @@ SNode* createSessionWindowNode(SAstCreateContext* pCxt, SNode* pCol, SNode* pGap
   return (SNode*)session;
 }
 
-SNode* createStateWindowNode(SAstCreateContext* pCxt, SNode* pCol) {
+SNode* createStateWindowNode(SAstCreateContext* pCxt, SNode* pExpr) {
   SStateWindowNode* state = (SStateWindowNode*)nodesMakeNode(QUERY_NODE_STATE_WINDOW);
   CHECK_OUT_OF_MEM(state);
-  state->pCol = pCol;
+  state->pExpr = pExpr;
   return (SNode*)state;
 }
 
diff --git a/source/libs/parser/src/sql.c b/source/libs/parser/src/sql.c
index 86e564adab..30c85e28fe 100644
--- a/source/libs/parser/src/sql.c
+++ b/source/libs/parser/src/sql.c
@@ -209,153 +209,155 @@ typedef union {
 **  yy_default[]       Default action for each state.
 **
 *********** Begin parsing tables **********************************************/
-#define YY_ACTTAB_COUNT (1448)
+#define YY_ACTTAB_COUNT (1461)
 static const YYACTIONTYPE yy_action[] = {
- /*     0 */  1250,  441,  441,  265,  441, 1246, 1413,   72,  304,  306,
- /*    10 */    72,  111,   30,   28,  348,  409,  252,  354,   20, 1263,
- /*    20 */   261, 1153,  990, 1261, 1261,  269, 1261, 1429,   31,   29,
- /*    30 */    27,   26,   25, 1399,  412,  441,  233,   98,  988,  117,
- /*    40 */   409,  305, 1399, 1014,  427, 1395, 1401,   11, 1386, 1429,
- /*    50 */  1301,   30,   28, 1094, 1395, 1401,  425, 1261,  799,  261,
+ /*     0 */  1250,  441,  441,  266,  441, 1246, 1413,   72,  305,  307,
+ /*    10 */    72,  111,   31,   29,  349,  409,  253,  355,   21, 1263,
+ /*    20 */   262, 1153,  990, 1261, 1261,  270, 1261, 1429,   32,   30,
+ /*    30 */    28,   27,   26, 1399,  412,  441,  234,   98,  988,  117,
+ /*    40 */   409,  306, 1399, 1014,  427, 1395, 1401,   11, 1386, 1429,
+ /*    50 */  1301,   31,   29, 1094, 1395, 1401,  425, 1261,  799,  262,
  /*    60 */   798,  990,   98,  409,   69, 1414, 1415, 1418, 1462,    1,
- /*    70 */    96,  276,  254, 1458,  120, 1413, 1308,  988,  800,  411,
- /*    80 */   121, 1469, 1470, 1308, 1474,   98,   11, 1340,  399,  251,
- /*    90 */    30,   28,  507, 1490, 1306,   96, 1429, 1252,  261,  325,
+ /*    70 */    96,  277,  255, 1458,  120, 1413, 1308,  988,  800,  411,
+ /*    80 */   121, 1469, 1470, 1308, 1474,   98,   11, 1340,  402,  252,
+ /*    90 */    31,   29,  507, 1490, 1306,   96, 1429, 1252,  262,  326,
  /*   100 */   990,  440, 1523,  425,  989,  122, 1469, 1470,    1, 1474,
- /*   110 */    22,   24,  191,  427, 1399,  125,  988, 1386,   96, 1521,
- /*   120 */    31,   29,   27,   26,   25,   11, 1395, 1402,  123, 1469,
+ /*   110 */    23,   25,  191,  427, 1399,  125,  988, 1386,   96, 1521,
+ /*   120 */    32,   30,   28,   27,   26,   11, 1395, 1402,  123, 1469,
  /*   130 */  1470,  507, 1474,   70, 1414, 1415, 1418, 1462,  991, 1386,
- /*   140 */   440, 1461, 1458,  989,  135,  134, 1413,    1,   31,   29,
- /*   150 */    27,   26,   25,  186,   65,  994,  995, 1038, 1039, 1040,
- /*   160 */  1041, 1042, 1043, 1044, 1045,   99,  232, 1429, 1010, 1406,
- /*   170 */   507,  299, 1253,  298,  425,  318,  126,  991,  330, 1010,
- /*   180 */   428, 1404,  989,  264,  427, 1429, 1348,  331, 1386,   12,
- /*   190 */  1413,  398,  425,  131,  994,  995, 1038, 1039, 1040, 1041,
+ /*   140 */   440, 1461, 1458,  989,  135,  134, 1413,    1,   32,   30,
+ /*   150 */    28,   27,   26,  186,   65,  994,  995, 1038, 1039, 1040,
+ /*   160 */  1041, 1042, 1043, 1044, 1045,   99,  233, 1429, 1010, 1406,
+ /*   170 */   507,  300, 1253,  299,  425,  319,  126,  991,  331, 1010,
+ /*   180 */   428, 1404,  989,  265,  427,  273, 1348,  332, 1386,   12,
+ /*   190 */  1413,  126, 1339, 1341,  994,  995, 1038, 1039, 1040, 1041,
  /*   200 */  1042, 1043, 1044, 1045,   69, 1414, 1415, 1418, 1462,  441,
- /*   210 */   126, 1429,  254, 1458, 1535,  311,  991,   47,  425, 1308,
- /*   220 */    46,  441, 1336, 1496,  402,  266,  440,  312,  427,  133,
+ /*   210 */   126, 1429,  255, 1458, 1535,  312,  991, 1239,  425, 1308,
+ /*   220 */   159,  441, 1336, 1496,  352,  267,  440,  313,  427,  133,
  /*   230 */  1306, 1261, 1386,  994,  995, 1038, 1039, 1040, 1041, 1042,
- /*   240 */  1043, 1044, 1045, 1261,  404,  400,   30,   28,   70, 1414,
- /*   250 */  1415, 1418, 1462,   64,  261,  329,  990, 1459,  324,  323,
- /*   260 */   322,  321,  320,   60,  317,  316,  315,  314,  310,  309,
- /*   270 */   308,  307,  988,   84, 1118,   12,   83,   82,   81,   80,
- /*   280 */    79,   78,   77,   76,   75,   30,   28,  426, 1156,  361,
- /*   290 */   111,  356,  300,  261,  360,  990,  126,  160, 1264,  357,
- /*   300 */   355,  126,  358,    7,  395, 1116, 1117, 1119, 1120,   84,
- /*   310 */     6,  988,   83,   82,   81,   80,   79,   78,   77,   76,
- /*   320 */    75,  441, 1308, 1523,   30,   28,  507,  338,  273,  503,
- /*   330 */   502,  441,  261, 1306,  990,  126,  125, 1258,  989,  114,
- /*   340 */  1521, 1225,    7, 1261,  272,   31,   29,   27,   26,   25,
- /*   350 */   988, 1339, 1341, 1261,  874,  464,  463,  462,  878,  461,
+ /*   240 */  1043, 1044, 1045, 1261, 1237,  354,   31,   29,   70, 1414,
+ /*   250 */  1415, 1418, 1462,  340,  262,  330,  990, 1459,  325,  324,
+ /*   260 */   323,  322,  321,  340,  318,  317,  316,  315,  311,  310,
+ /*   270 */   309,  308,  988,   84, 1118,   12,   83,   82,   81,   80,
+ /*   280 */    79,   78,   77,   76,   75,   31,   29,  426, 1156,  362,
+ /*   290 */   476,  357,  301,  262,  361,  990,  126,  160,  213,  358,
+ /*   300 */   356, 1291,  359,    7,  396, 1116, 1117, 1119, 1120,   84,
+ /*   310 */  1178,  988,   83,   82,   81,   80,   79,   78,   77,   76,
+ /*   320 */    75,  441, 1308, 1523,   31,   29,  507,  339,  274,  399,
+ /*   330 */   476,  441,  262, 1306,  990,  126,  125, 1258,  989,  114,
+ /*   340 */  1521, 1225,    7, 1261, 1012,   32,   30,   28,   27,   26,
+ /*   350 */   988,  131, 1386, 1261,  874,  464,  463,  462,  878,  461,
  /*   360 */   880,  881,  460,  883,  457,  507,  889,  454,  891,  892,
- /*   370 */   451,  448,  991,   27,   26,   25,  113,  989, 1167,  213,
- /*   380 */  1413,    7, 1291, 1375,   31,   29,   27,   26,   25,  994,
+ /*   370 */   451,  448,  991, 1177,  113,   48, 1167,  989,   47,  403,
+ /*   380 */  1413,    7,  404,  400,   32,   30,   28,   27,   26,  994,
  /*   390 */   995, 1038, 1039, 1040, 1041, 1042, 1043, 1044, 1045,  441,
- /*   400 */  1523, 1429, 1201, 1026,  507,  438,  339,  384,  425, 1239,
- /*   410 */   126,  991, 1178, 1522,  428, 1012,  989, 1521,  427,  283,
- /*   420 */  1349, 1261, 1386,   31,   29,   27,   26,   25,  994,  995,
+ /*   400 */  1523, 1429, 1201, 1108,  507,  438,  111,  385,  425, 1013,
+ /*   410 */   126,  991, 1176, 1522, 1264, 1386,  989, 1521,  427,  503,
+ /*   420 */   502, 1261, 1386,   32,   30,   28,   27,   26,  994,  995,
  /*   430 */  1038, 1039, 1040, 1041, 1042, 1043, 1044, 1045,  115, 1414,
- /*   440 */  1415, 1418, 1089,   31,   29,   27,   26,   25,  385,  271,
- /*   450 */   991,  361, 1015,  356, 1386,  339,  360,  111,  235,  160,
- /*   460 */   245,  357,  355, 1177,  358, 1263, 1413,  994,  995, 1038,
- /*   470 */  1039, 1040, 1041, 1042, 1043, 1044, 1045, 1537,  235, 1523,
- /*   480 */    30,   28,  159,    9,    8,  403,  351, 1429,  261, 1057,
- /*   490 */   990, 1155,  125,   49,  425,  291, 1521,  246,  375,  244,
- /*   500 */   243, 1026,  350, 1148,  427, 1386,  988,  353, 1386, 1057,
- /*   510 */   293, 1257, 1176,  413, 1175,   93,   92,   91,   90,   89,
+ /*   440 */  1415, 1418,    6,   32,   30,   28,   27,   26,  386,  272,
+ /*   450 */   991,  362, 1015,  357, 1386, 1248,  361,  111,  236,  160,
+ /*   460 */   246,  358,  356, 1175,  359, 1263, 1413,  994,  995, 1038,
+ /*   470 */  1039, 1040, 1041, 1042, 1043, 1044, 1045, 1537,  236, 1523,
+ /*   480 */    31,   29,  159,    9,    8, 1011,  352, 1429,  262, 1057,
+ /*   490 */   990, 1155,  125, 1059,  425,  292, 1521,  247,  376,  245,
+ /*   500 */   244, 1026,  351, 1148,  427, 1386,  988,  354, 1386, 1057,
+ /*   510 */   294, 1063, 1174,  413, 1173,   93,   92,   91,   90,   89,
  /*   520 */    88,   87,   86,   85,   68, 1414, 1415, 1418, 1462, 1523,
- /*   530 */    49,  274,  234, 1458,  159, 1058,  416,    1,  351,  111,
- /*   540 */    95, 1093,  125, 1174, 1523,  441, 1521, 1263, 1256, 1013,
- /*   550 */  1059,  439, 1070, 1062, 1386, 1058, 1386,  125,  476,  353,
- /*   560 */   507, 1521,  172, 1308,  441,  441, 1173, 1261, 1063, 1147,
- /*   570 */   205,  275,  989, 1062, 1307, 1172,   23,  259, 1052, 1053,
- /*   580 */  1054, 1055, 1056, 1060, 1061, 1386, 1261, 1261, 1011, 1476,
- /*   590 */  1194,   21, 1171, 1170, 1169, 1166,   23,  259, 1052, 1053,
- /*   600 */  1054, 1055, 1056, 1060, 1061, 1198,  991, 1473, 1386, 1101,
- /*   610 */  1165,  147,  362, 1248,  119, 1012,  112, 1386, 1481, 1089,
- /*   620 */   344,  219,  146,  994,  995, 1038, 1039, 1040, 1041, 1042,
- /*   630 */  1043, 1044, 1045,  217, 1386, 1386, 1386, 1386, 1164, 1163,
- /*   640 */  1162, 1161, 1160,  136, 1159, 1158,   50,    9,    8,  144,
- /*   650 */  1237, 1413, 1386,  420,  495,  494,  493,  492,  491,  490,
+ /*   530 */    50,  275,  235, 1458,   22, 1058,  416,    1, 1308,  111,
+ /*   540 */    95,   50,  125, 1172, 1523,  441, 1521, 1263, 1256, 1307,
+ /*   550 */   798,  439, 1070, 1062, 1386, 1058, 1386,  125,  467, 1257,
+ /*   560 */   507, 1521,  172,  417,  441,  441,  347, 1261, 1171, 1147,
+ /*   570 */   205,  276,  989, 1062, 1089, 1170,   24,  260, 1052, 1053,
+ /*   580 */  1054, 1055, 1056, 1060, 1061, 1386, 1261, 1261, 1244,   32,
+ /*   590 */    30,   28,   27,   26, 1169, 1166,   24,  260, 1052, 1053,
+ /*   600 */  1054, 1055, 1056, 1060, 1061, 1198,  991,   28,   27,   26,
+ /*   610 */  1386,  147,  428, 1101,  119, 1093,  112, 1386, 1349, 1012,
+ /*   620 */   345,  219,  146,  994,  995, 1038, 1039, 1040, 1041, 1042,
+ /*   630 */  1043, 1044, 1045,  217, 1165, 1476, 1386, 1386, 1164, 1163,
+ /*   640 */  1162, 1161, 1160,  136, 1159, 1158,   51, 1026,  420,  144,
+ /*   650 */   998, 1413, 1375, 1473,  495,  494,  493,  492,  491,  490,
  /*   660 */   489,  488,  207,  485,  484,  483,  482,  481,  480,  479,
- /*   670 */   478,  477, 1429, 1476, 1476,  486,   58,  467,  294,  425,
- /*   680 */  1386, 1386, 1386, 1386, 1386,  798, 1386, 1386,  417,  427,
- /*   690 */   424, 1472, 1471, 1386, 1254,  152,  476,  154,  150, 1413,
- /*   700 */   153,  346,  373,  143,   67,  138,  998,  140,  383,   69,
- /*   710 */  1414, 1415, 1418, 1462, 1092,  371, 1244,  254, 1458, 1535,
- /*   720 */  1429, 1150, 1151,  165,  137,  415,  466,  412, 1519,   45,
- /*   730 */    44,  303,  156,  130, 1189,  155,  104,  427,  297,  387,
- /*   740 */   158, 1386, 1413,  157, 1168, 1226,  241, 1187,  289,  396,
- /*   750 */   285,  281,  127,  188, 1302,  345,  364,   69, 1414, 1415,
- /*   760 */  1418, 1462,  181, 1429, 1492,  254, 1458,  120,   43,  367,
- /*   770 */   425, 1115, 1001,  175,  126,  997,  177, 1430, 1413,  187,
- /*   780 */   427,  410,  421,  190, 1386,  391, 1489,   32,   32,   32,
- /*   790 */  1064, 1023,  957,  194,    2, 1010,  196,  278, 1049, 1429,
- /*   800 */    69, 1414, 1415, 1418, 1462, 1413,  425,  282,  254, 1458,
- /*   810 */  1535,  101,  240,  242,  433,  418,  427,  841,  102, 1480,
- /*   820 */  1386,  202,  966,  211,  313,  413, 1429,  409,  510, 1338,
- /*   830 */   132,  104,  319,  425,  867,  327,  224, 1414, 1415, 1418,
- /*   840 */   326, 1000,  210,  427,  328,   94,  332, 1386, 1019,   98,
- /*   850 */   333,  499,   43,  209,  446,  862, 1523,  895,  102,  103,
- /*   860 */  1413,  899,  905,   70, 1414, 1415, 1418, 1462,  413,  125,
- /*   870 */   334,  423, 1458, 1521, 1018,  335,  139,   66, 1017,  336,
- /*   880 */   203, 1429,   96,  104,  102,  337,  904,  105,  425,  142,
- /*   890 */    48,  145,  184, 1469,  408,  340,  407, 1016,  427, 1523,
- /*   900 */   347,  349, 1386,   74, 1413, 1251,  376,  149, 1247, 1413,
- /*   910 */   437,  377,  125,  151,  106,  107, 1521, 1249,  115, 1414,
- /*   920 */  1415, 1418, 1245,  108,  109, 1429,  352,  359,  378,  250,
- /*   930 */  1429, 1015,  425,  382,  167,  390,  388,  425,  168,  379,
- /*   940 */   170,  397,  427, 1503,  431,  386, 1386,  427, 1493,  260,
- /*   950 */   389, 1386,  995,  974,  392,  164,  414, 1536, 1502,    5,
- /*   960 */  1483,  173,  228, 1414, 1415, 1418, 1413,  228, 1414, 1415,
- /*   970 */  1418, 1413,  406,  394,  393,  176,  253,    4,  401, 1089,
- /*   980 */    97, 1014,   33,  183,  180,  422,  255, 1429,  118, 1477,
- /*   990 */   419,  429, 1429,  182,  425,   17, 1347, 1346,  430,  425,
- /*  1000 */   263,  434, 1444,  435,  427,  198, 1538, 1413, 1386,  427,
- /*  1010 */   200,   57,  436, 1386, 1520, 1413,  258,   59, 1262,  212,
- /*  1020 */   473,  189,  444,  214,  227, 1414, 1415, 1418, 1429,  228,
- /*  1030 */  1414, 1415, 1418,  487,  208,  425, 1429,  506,  216,  220,
- /*  1040 */   221, 1413,   39,  425,  218,  427, 1380, 1379,  277, 1386,
- /*  1050 */  1413,  279,  262,  427, 1376,  280,  405, 1386,  984,  985,
- /*  1060 */   128,  284, 1429, 1374,  286,  228, 1414, 1415, 1418,  425,
- /*  1070 */   287, 1429,  288,  226, 1414, 1415, 1418, 1413,  425,  427,
- /*  1080 */  1373,  290, 1372, 1386, 1363, 1413,  292,  129,  427,  295,
- /*  1090 */   296,  969, 1386,  968, 1357, 1356,  302, 1355, 1429,  229,
- /*  1100 */  1414, 1415, 1418,  301, 1354,  425, 1429, 1331,  222, 1414,
- /*  1110 */  1415, 1418,  940,  425, 1330,  427, 1329, 1328, 1327, 1386,
- /*  1120 */  1326, 1413, 1325,  427, 1324, 1323, 1413, 1386, 1322, 1321,
- /*  1130 */  1320, 1319,  100, 1318, 1317,  230, 1414, 1415, 1418, 1316,
- /*  1140 */  1315, 1314, 1429,  223, 1414, 1415, 1418, 1429, 1313,  425,
- /*  1150 */  1312, 1311, 1310, 1309,  425, 1200, 1371,  942, 1365,  427,
- /*  1160 */  1353, 1344, 1413, 1386,  427, 1240,  141, 1413, 1386,  811,
- /*  1170 */  1199, 1197, 1186, 1413,  341,  343, 1185, 1182,  342,  231,
- /*  1180 */  1414, 1415, 1418, 1429, 1426, 1414, 1415, 1418, 1429, 1242,
- /*  1190 */   425,   73,  912,  486, 1429,  425,  910, 1241, 1195, 1190,
- /*  1200 */   427,  425, 1238,  148, 1386,  427,  840,  839,  247, 1386,
- /*  1210 */   248,  427,  365,  838, 1413, 1386,  837,  835,  834, 1188,
- /*  1220 */  1425, 1414, 1415, 1418, 1236, 1424, 1414, 1415, 1418, 1413,
- /*  1230 */   249,  238, 1414, 1415, 1418, 1429, 1181,  370,  268,  267,
- /*  1240 */  1413,  368,  425, 1180,  372,   71, 1370,  163, 1003,   42,
- /*  1250 */  1429,  976,  427, 1364,  990, 1413, 1386,  425,  110,  206,
- /*  1260 */   380, 1429, 1352,  472,  996,  166, 1351,  427,  425, 1343,
- /*  1270 */   988, 1386,  237, 1414, 1415, 1418, 1429,  381,  427,   36,
- /*  1280 */   366,  206, 1386,  425,  474,  472,   51,  239, 1414, 1415,
- /*  1290 */  1418,  169,  171,  427,    3,  374,   32, 1386,  236, 1414,
- /*  1300 */  1415, 1418,   37,  471,  470,  469,  474,  468,  116,  162,
- /*  1310 */   174, 1114,  369,  225, 1414, 1415, 1418,  363,  442,  178,
- /*  1320 */   161, 1108, 1107,   14,  507,  471,  470,  469,   52,  468,
- /*  1330 */   999,  179,   53,   34,   19, 1404,  989,   15, 1086,  185,
- /*  1340 */  1085,   35,  124, 1141,   41,   16,   10,   40,   54, 1136,
- /*  1350 */  1135,  256, 1140, 1139,  257,    8, 1050, 1024,  192,   13,
- /*  1360 */    18,  432,  193, 1112, 1004,  195,  197,   55, 1342,  199,
- /*  1370 */   991,   56,  201,   60, 1005,   38, 1403,  896,  445,  270,
- /*  1380 */   204, 1007,  995,  443,  449,  452,  447,  994,  995,  893,
- /*  1390 */   450,  873,  890,  453,  455,  884,  456,  458,  882,  459,
- /*  1400 */    61,  907,  888,  887,  886,  465,   62,   63,  906,  903,
- /*  1410 */   901,  475,  885,  809,  831,  830,  829,  828,  827,  826,
- /*  1420 */   825,  824,  823,  842,  821,  820, 1196,  819,  818,  817,
- /*  1430 */  1184,  816,  815,  814,  496,  497,  500, 1183,  501, 1179,
- /*  1440 */   498,  504,  505, 1154,  992,  215,  508,  509,
+ /*   670 */   478,  477, 1429, 1476, 1476,  424, 1386,  165,   58,  425,
+ /*   680 */  1386, 1386, 1386, 1386, 1386,  466, 1386, 1386,  284,  427,
+ /*   690 */   418, 1472, 1471, 1386, 1481, 1089, 1254,  486,  152, 1413,
+ /*   700 */   295,  150,  374,  143,   67,  138,  384,  140, 1168,   69,
+ /*   710 */  1414, 1415, 1418, 1462, 1226,  372, 1001,  255, 1458, 1535,
+ /*   720 */  1429, 1194, 1189,  154,  137,  415,  153,  412, 1519,   46,
+ /*   730 */    45,  304,  156,  130, 1187,  155,  158,  427,  298,  157,
+ /*   740 */  1302, 1386, 1413,  363,  365,   64,  242,  188,  290,  397,
+ /*   750 */   286,  282,  127,    9,    8,   60,  368,   69, 1414, 1415,
+ /*   760 */  1418, 1462,  181, 1429,  346,  255, 1458,  120,  104,   44,
+ /*   770 */   425,  388, 1115,  175,  126,  997,  177,  421, 1413,  187,
+ /*   780 */   427, 1150, 1151, 1049, 1386,  392, 1489,   33, 1092,  410,
+ /*   790 */  1064,   33,   33,  194, 1023,  957,  196, 1492, 1430, 1429,
+ /*   800 */    69, 1414, 1415, 1418, 1462, 1413,  425,  190,  255, 1458,
+ /*   810 */  1535,  101,  102,    2,  433,  202,  427, 1010,  104, 1480,
+ /*   820 */  1386,  867,  283,  279,  241,  413, 1429,  409,  510,  243,
+ /*   830 */   966,   44,  841,  425,  862,  211,  225, 1414, 1415, 1418,
+ /*   840 */   320, 1000,  210,  427,  314,   94, 1338, 1386,  132,   98,
+ /*   850 */   333,  499,  446,  209,  102,  895, 1523,  899,  103,  104,
+ /*   860 */  1413,  905,  904,   70, 1414, 1415, 1418, 1462,  413,  125,
+ /*   870 */   328,  423, 1458, 1521, 1019,  327,  329,   66,  334, 1018,
+ /*   880 */   203, 1429,   96,  102,  335,  336,  105,  139,  425,  337,
+ /*   890 */  1017,  338,  184, 1469,  408,  142,  407,   49,  427, 1523,
+ /*   900 */   341,  145, 1386, 1016, 1413,  348,   74,  350,  379, 1413,
+ /*   910 */   437, 1251,  125,  149,  378,  380, 1521, 1247,  115, 1414,
+ /*   920 */  1415, 1418,  151,  106,  353, 1429,  360,  377,  107, 1249,
+ /*   930 */  1429, 1015,  425, 1245,  108,  391,  109,  425,  168,  167,
+ /*   940 */   170,  251,  427,  387,  398,  389, 1386,  427, 1493,  261,
+ /*   950 */   390, 1386,  431,  974,  393,  164,  414, 1536,  383, 1503,
+ /*   960 */   995,  173,  229, 1414, 1415, 1418, 1413,  229, 1414, 1415,
+ /*   970 */  1418, 1413, 1502,  395,    5,  176, 1089,  394,  254,  406,
+ /*   980 */  1483,   97,  401,    4, 1014,   17,   34, 1429,  256, 1477,
+ /*   990 */   422,  183, 1429,  419,  425, 1347,  429,  435,  198,  425,
+ /*  1000 */   430,  264,  182,  180,  427,  118, 1346, 1413, 1386,  427,
+ /*  1010 */  1444,  434, 1262, 1386,  436, 1413,  259,  200,   59,   57,
+ /*  1020 */   212, 1520,  214,  208,  228, 1414, 1415, 1418, 1429,  229,
+ /*  1030 */  1414, 1415, 1418,  473,  444,  425, 1429,  189,  487,  506,
+ /*  1040 */   220, 1413, 1538,  425,  221,  427,   40,  216, 1380, 1386,
+ /*  1050 */  1413,  218,  263,  427, 1379,  278,  405, 1386, 1376,  280,
+ /*  1060 */   281,  984, 1429,  285,  985,  229, 1414, 1415, 1418,  425,
+ /*  1070 */   128, 1429, 1374,  222, 1414, 1415, 1418, 1413,  425,  427,
+ /*  1080 */   287,  288,  289, 1386, 1373, 1413, 1372,  291,  427, 1363,
+ /*  1090 */   293,  129, 1386,  296,  297,  969, 1357,  968, 1429,  227,
+ /*  1100 */  1414, 1415, 1418, 1356,  303,  425, 1429,  302,  230, 1414,
+ /*  1110 */  1415, 1418, 1355,  425, 1354,  427,  940, 1331, 1330, 1386,
+ /*  1120 */  1329, 1413, 1328,  427, 1327, 1326, 1413, 1386, 1325, 1324,
+ /*  1130 */  1323, 1322, 1321, 1320, 1319,  223, 1414, 1415, 1418,  100,
+ /*  1140 */  1318, 1317, 1429,  231, 1414, 1415, 1418, 1429, 1316,  425,
+ /*  1150 */  1315, 1314, 1313, 1312,  425, 1311, 1310,  942, 1309,  427,
+ /*  1160 */  1200, 1371, 1413, 1386,  427, 1365, 1353, 1413, 1386, 1344,
+ /*  1170 */   141, 1240, 1199, 1413,  811, 1197,  342, 1186,  343,  224,
+ /*  1180 */  1414, 1415, 1418, 1429,  232, 1414, 1415, 1418, 1429,  344,
+ /*  1190 */   425, 1185, 1182, 1242, 1429,  425,  910,   73, 1241,  148,
+ /*  1200 */   427,  425,  912,  486, 1386,  427,  840, 1195,  839, 1386,
+ /*  1210 */   248,  427, 1190, 1188, 1413, 1386,  838,  837,  835,  834,
+ /*  1220 */  1426, 1414, 1415, 1418,  249, 1425, 1414, 1415, 1418, 1413,
+ /*  1230 */   250, 1424, 1414, 1415, 1418, 1429,  366,  369,  269,  268,
+ /*  1240 */  1413,  371,  425, 1181, 1180,  373, 1370,   71, 1003, 1238,
+ /*  1250 */  1429,   43,  427,  163,  976, 1413, 1386,  425, 1364,  110,
+ /*  1260 */   381, 1429, 1352,  990,  996, 1236, 1351,  427,  425,  382,
+ /*  1270 */   166, 1386,  239, 1414, 1415, 1418, 1429,   37,  427,  988,
+ /*  1280 */  1343, 1413, 1386,  425,   52,  169,  116,  238, 1414, 1415,
+ /*  1290 */  1418,    3,  367,  427,  178,   33,   38, 1386,  240, 1414,
+ /*  1300 */  1415, 1418, 1429,  174, 1114,   19,  206,  375,  171,  425,
+ /*  1310 */   472, 1404,  179,  237, 1414, 1415, 1418,   14,  442,  427,
+ /*  1320 */  1107,  162,  206, 1386,  370,   53,  472,   20,  185,  364,
+ /*  1330 */   999,  474,  161,  507,   15,  124,   35, 1086, 1085,  226,
+ /*  1340 */  1414, 1415, 1418,   36, 1141,  989,   16,  474,   54, 1136,
+ /*  1350 */   471,  470,  469,   10,  468, 1135,   42,  257, 1140,   41,
+ /*  1360 */  1139,  258,    8,  192, 1004, 1050,  471,  470,  469, 1024,
+ /*  1370 */   468,   13,   18, 1342,  193, 1112,  195,  197,   55,  991,
+ /*  1380 */   432, 1007,  995,  199,  201,   56, 1005, 1403,  896,   60,
+ /*  1390 */    39,  445,  271,  204,  443,  447,  994,  995,  893,  449,
+ /*  1400 */   450,  452,  890,  453,  455,  884,  456,  458,  882,  888,
+ /*  1410 */   459,  873,  465,  907,   61,   62,   63,  906,  887,  886,
+ /*  1420 */   903,  885,  901,  831,  809,  475,  830,  823,  829,  828,
+ /*  1430 */   827,  826,  825,  824,  842,  821,  820,  819, 1196, 1184,
+ /*  1440 */   818,  817,  816, 1183,  815,  814,  496,  497,  500,  501,
+ /*  1450 */  1179,  504,  498,  505, 1154,  992,  215,  508, 1154, 1154,
+ /*  1460 */   509,
 };
 static const YYCODETYPE yy_lookahead[] = {
  /*     0 */   226,  210,  210,  217,  210,  226,  204,  216,  216,  210,
@@ -376,135 +378,135 @@ static const YYCODETYPE yy_lookahead[] = {
  /*   150 */    14,   15,   16,  130,  209,  155,  156,  157,  158,  159,
  /*   160 */   160,  161,  162,  163,  164,  220,   18,  225,   20,   69,
  /*   170 */    92,  137,  227,  139,  232,   27,  176,  138,   30,   20,
- /*   180 */   242,   81,  104,  245,  242,  225,  248,   39,  246,   69,
- /*   190 */   204,  128,  232,   44,  155,  156,  157,  158,  159,  160,
+ /*   180 */   242,   81,  104,  245,  242,  234,  248,   39,  246,   69,
+ /*   190 */   204,  176,  241,  242,  155,  156,  157,  158,  159,  160,
  /*   200 */   161,  162,  163,  164,  262,  263,  264,  265,  266,  210,
- /*   210 */   176,  225,  270,  271,  272,  216,  138,   68,  232,  225,
- /*   220 */    71,  210,  232,  281,  264,  231,   20,  216,  242,  239,
+ /*   210 */   176,  225,  270,  271,  272,  216,  138,    0,  232,  225,
+ /*   220 */    57,  210,  232,  281,   61,  231,   20,  216,  242,  239,
  /*   230 */   236,  232,  246,  155,  156,  157,  158,  159,  160,  161,
- /*   240 */   162,  163,  164,  232,  181,  182,   12,   13,  262,  263,
- /*   250 */   264,  265,  266,   69,   20,  107,   22,  271,  110,  111,
- /*   260 */   112,  113,  114,   79,  116,  117,  118,  119,  120,  121,
+ /*   240 */   162,  163,  164,  232,    0,   82,   12,   13,  262,  263,
+ /*   250 */   264,  265,  266,   46,   20,  107,   22,  271,  110,  111,
+ /*   260 */   112,  113,  114,   46,  116,  117,  118,  119,  120,  121,
  /*   270 */   122,  123,   38,   21,  155,   69,   24,   25,   26,   27,
  /*   280 */    28,   29,   30,   31,   32,   12,   13,   14,    0,   49,
- /*   290 */   225,   51,  251,   20,   54,   22,  176,   57,  233,   59,
- /*   300 */    60,  176,   62,   69,  185,  186,  187,  188,  189,   21,
- /*   310 */    43,   38,   24,   25,   26,   27,   28,   29,   30,   31,
- /*   320 */    32,  210,  225,  282,   12,   13,   92,  216,  231,  207,
- /*   330 */   208,  210,   20,  236,   22,  176,  295,  216,  104,  213,
- /*   340 */   299,  215,   69,  232,  234,   12,   13,   14,   15,   16,
- /*   350 */    38,  241,  242,  232,   83,   84,   85,   86,   87,   88,
+ /*   290 */    46,   51,  251,   20,   54,   22,  176,   57,  218,   59,
+ /*   300 */    60,  221,   62,   69,  185,  186,  187,  188,  189,   21,
+ /*   310 */   204,   38,   24,   25,   26,   27,   28,   29,   30,   31,
+ /*   320 */    32,  210,  225,  282,   12,   13,   92,  216,  231,  128,
+ /*   330 */    46,  210,   20,  236,   22,  176,  295,  216,  104,  213,
+ /*   340 */   299,  215,   69,  232,   20,   12,   13,   14,   15,   16,
+ /*   350 */    38,   44,  246,  232,   83,   84,   85,   86,   87,   88,
  /*   360 */    89,   90,   91,   92,   93,   92,   95,   96,   97,   98,
- /*   370 */    99,  100,  138,   14,   15,   16,  203,  104,  205,  218,
- /*   380 */   204,   69,  221,    0,   12,   13,   14,   15,   16,  155,
+ /*   370 */    99,  100,  138,  204,  203,   68,  205,  104,   71,   20,
+ /*   380 */   204,   69,  181,  182,   12,   13,   14,   15,   16,  155,
  /*   390 */   156,  157,  158,  159,  160,  161,  162,  163,  164,  210,
- /*   400 */   282,  225,    0,   70,   92,  216,   46,  210,  232,    0,
- /*   410 */   176,  138,  204,  295,  242,   20,  104,  299,  242,   36,
- /*   420 */   248,  232,  246,   12,   13,   14,   15,   16,  155,  156,
+ /*   400 */   282,  225,    0,   70,   92,  216,  225,  210,  232,   20,
+ /*   410 */   176,  138,  204,  295,  233,  246,  104,  299,  242,  207,
+ /*   420 */   208,  232,  246,   12,   13,   14,   15,   16,  155,  156,
  /*   430 */   157,  158,  159,  160,  161,  162,  163,  164,  262,  263,
- /*   440 */   264,  265,  175,   12,   13,   14,   15,   16,  251,  217,
- /*   450 */   138,   49,   20,   51,  246,   46,   54,  225,   47,   57,
+ /*   440 */   264,  265,   43,   12,   13,   14,   15,   16,  251,  217,
+ /*   450 */   138,   49,   20,   51,  246,  226,   54,  225,   47,   57,
  /*   460 */    35,   59,   60,  204,   62,  233,  204,  155,  156,  157,
  /*   470 */   158,  159,  160,  161,  162,  163,  164,  301,   47,  282,
  /*   480 */    12,   13,   57,    1,    2,   20,   61,  225,   20,   78,
- /*   490 */    22,    0,  295,  212,  232,  134,  299,   72,  251,   74,
+ /*   490 */    22,    0,  295,  124,  232,  134,  299,   72,  251,   74,
  /*   500 */    75,   70,   77,  131,  242,  246,   38,   82,  246,   78,
- /*   510 */   149,  230,  204,  251,  204,   24,   25,   26,   27,   28,
+ /*   510 */   149,  142,  204,  251,  204,   24,   25,   26,   27,   28,
  /*   520 */    29,   30,   31,   32,  262,  263,  264,  265,  266,  282,
- /*   530 */   212,  217,  270,  271,   57,  124,    3,   69,   61,  225,
- /*   540 */   222,    4,  295,  204,  282,  210,  299,  233,  230,   20,
- /*   550 */   124,  216,   70,  142,  246,  124,  246,  295,   46,   82,
- /*   560 */    92,  299,  130,  225,  210,  210,  204,  232,  142,  197,
- /*   570 */   216,  216,  104,  142,  236,  204,  165,  166,  167,  168,
- /*   580 */   169,  170,  171,  172,  173,  246,  232,  232,   20,  260,
- /*   590 */     0,  165,  204,  204,  204,  204,  165,  166,  167,  168,
- /*   600 */   169,  170,  171,  172,  173,    0,  138,  278,  246,   14,
- /*   610 */   204,   33,   22,  226,   36,   20,   18,  246,  174,  175,
+ /*   530 */   212,  217,  270,  271,  165,  124,    3,   69,  225,  225,
+ /*   540 */   222,  212,  295,  204,  282,  210,  299,  233,  230,  236,
+ /*   550 */    22,  216,   70,  142,  246,  124,  246,  295,   80,  230,
+ /*   560 */    92,  299,  130,   67,  210,  210,   38,  232,  204,  197,
+ /*   570 */   216,  216,  104,  142,  175,  204,  165,  166,  167,  168,
+ /*   580 */   169,  170,  171,  172,  173,  246,  232,  232,  226,   12,
+ /*   590 */    13,   14,   15,   16,  204,  204,  165,  166,  167,  168,
+ /*   600 */   169,  170,  171,  172,  173,    0,  138,   14,   15,   16,
+ /*   610 */   246,   33,  242,   14,   36,    4,   18,  246,  248,   20,
  /*   620 */    42,   23,   44,  155,  156,  157,  158,  159,  160,  161,
- /*   630 */   162,  163,  164,   35,  246,  246,  246,  246,  204,  204,
- /*   640 */   204,  204,  204,   45,  204,  204,   68,    1,    2,   71,
- /*   650 */     0,  204,  246,   67,   49,   50,   51,   52,   53,   54,
+ /*   630 */   162,  163,  164,   35,  204,  260,  246,  246,  204,  204,
+ /*   640 */   204,  204,  204,   45,  204,  204,   68,   70,   67,   71,
+ /*   650 */    38,  204,    0,  278,   49,   50,   51,   52,   53,   54,
  /*   660 */    55,   56,   57,   58,   59,   60,   61,   62,   63,   64,
- /*   670 */    65,   66,  225,  260,  260,   67,  209,   80,   70,  232,
- /*   680 */   246,  246,  246,  246,  246,   22,  246,  246,   67,  242,
- /*   690 */    47,  278,  278,  246,  227,   73,   46,   73,   76,  204,
- /*   700 */    76,   38,   21,  125,  106,  127,   38,  129,  254,  262,
- /*   710 */   263,  264,  265,  266,  177,   34,  226,  270,  271,  272,
- /*   720 */   225,  199,  200,  226,  146,  192,  226,  232,  281,  131,
- /*   730 */   132,  133,   73,  135,    0,   76,   67,  242,  140,   70,
- /*   740 */    73,  246,  204,   76,  205,  215,  148,    0,  150,  293,
- /*   750 */   152,  153,  154,  302,  235,  207,   22,  262,  263,  264,
- /*   760 */   265,  266,  287,  225,  261,  270,  271,  272,   67,   22,
- /*   770 */   232,   70,  104,   67,  176,   38,   70,  225,  204,  284,
- /*   780 */   242,  280,  196,  296,  246,  290,  291,   67,   67,   67,
- /*   790 */    70,   70,   70,   67,  283,   20,   70,  210,  155,  225,
- /*   800 */   262,  263,  264,  265,  266,  204,  232,   36,  270,  271,
- /*   810 */   272,   67,  257,  214,   70,  194,  242,   38,   67,  281,
- /*   820 */   246,   70,  136,  252,  210,  251,  225,  210,   19,  210,
- /*   830 */   115,   67,  240,  232,   70,  124,  262,  263,  264,  265,
- /*   840 */   238,  104,   33,  242,  238,   36,  210,  246,   20,  232,
- /*   850 */   256,   42,   67,   44,   67,   70,  282,   70,   67,   67,
+ /*   670 */    65,   66,  225,  260,  260,   47,  246,  226,  209,  232,
+ /*   680 */   246,  246,  246,  246,  246,  226,  246,  246,   36,  242,
+ /*   690 */   194,  278,  278,  246,  174,  175,  227,   67,   73,  204,
+ /*   700 */    70,   76,   21,  125,  106,  127,  254,  129,  205,  262,
+ /*   710 */   263,  264,  265,  266,  215,   34,  104,  270,  271,  272,
+ /*   720 */   225,    0,    0,   73,  146,  192,   76,  232,  281,  131,
+ /*   730 */   132,  133,   73,  135,    0,   76,   73,  242,  140,   76,
+ /*   740 */   235,  246,  204,   22,   22,   69,  148,  302,  150,  293,
+ /*   750 */   152,  153,  154,    1,    2,   79,   22,  262,  263,  264,
+ /*   760 */   265,  266,  287,  225,  207,  270,  271,  272,   67,   67,
+ /*   770 */   232,   70,   70,   67,  176,   38,   70,  196,  204,  284,
+ /*   780 */   242,  199,  200,  155,  246,  290,  291,   67,  177,  280,
+ /*   790 */    70,   67,   67,   67,   70,   70,   70,  261,  225,  225,
+ /*   800 */   262,  263,  264,  265,  266,  204,  232,  296,  270,  271,
+ /*   810 */   272,   67,   67,  283,   70,   70,  242,   20,   67,  281,
+ /*   820 */   246,   70,   36,  210,  257,  251,  225,  210,   19,  214,
+ /*   830 */   136,   67,   38,  232,   70,  252,  262,  263,  264,  265,
+ /*   840 */   240,  104,   33,  242,  210,   36,  210,  246,  115,  232,
+ /*   850 */   210,   42,   67,   44,   67,   70,  282,   70,   67,   67,
  /*   860 */   204,   70,   70,  262,  263,  264,  265,  266,  251,  295,
- /*   870 */   242,  270,  271,  299,   20,  250,  212,   68,   20,  232,
- /*   880 */    71,  225,  265,   67,   67,  243,   70,   70,  232,  212,
- /*   890 */   212,  212,  275,  276,  277,  210,  279,   20,  242,  282,
- /*   900 */   206,  225,  246,  210,  204,  225,  232,  225,  225,  204,
- /*   910 */   101,  256,  295,  225,  225,  225,  299,  225,  262,  263,
- /*   920 */   264,  265,  225,  225,  225,  225,  214,  214,  145,  206,
- /*   930 */   225,   20,  232,  242,  209,  126,  232,  232,  129,  255,
- /*   940 */   209,  184,  242,  292,  183,  250,  246,  242,  261,  249,
- /*   950 */   243,  246,  156,  144,  249,  146,  300,  301,  292,  191,
- /*   960 */   289,  247,  262,  263,  264,  265,  204,  262,  263,  264,
- /*   970 */   265,  204,  190,  246,  179,  247,  246,  178,  246,  175,
- /*   980 */   232,   20,  115,  273,  288,  195,  198,  225,  286,  260,
- /*   990 */   193,  246,  225,  285,  232,   69,  247,  247,  246,  232,
- /*  1000 */   246,  127,  269,  244,  242,  232,  303,  204,  246,  242,
- /*  1010 */   209,  209,  243,  246,  298,  204,  249,   69,  232,  221,
- /*  1020 */   214,  297,  228,  210,  262,  263,  264,  265,  225,  262,
- /*  1030 */   263,  264,  265,  214,  209,  232,  225,  206,  211,  219,
- /*  1040 */   219,  204,  253,  232,  202,  242,    0,    0,   60,  246,
- /*  1050 */   204,   38,  249,  242,    0,  151,  294,  246,   38,   38,
- /*  1060 */    38,  151,  225,    0,   38,  262,  263,  264,  265,  232,
- /*  1070 */    38,  225,  151,  262,  263,  264,  265,  204,  232,  242,
- /*  1080 */     0,   38,    0,  246,    0,  204,   38,   69,  242,  142,
- /*  1090 */   141,  104,  246,  138,    0,    0,  134,    0,  225,  262,
- /*  1100 */   263,  264,  265,   50,    0,  232,  225,    0,  262,  263,
- /*  1110 */   264,  265,   81,  232,    0,  242,    0,    0,    0,  246,
+ /*   870 */   124,  270,  271,  299,   20,  238,  238,   68,  256,   20,
+ /*   880 */    71,  225,  265,   67,  242,  250,   70,  212,  232,  232,
+ /*   890 */    20,  243,  275,  276,  277,  212,  279,  212,  242,  282,
+ /*   900 */   210,  212,  246,   20,  204,  206,  210,  225,  145,  204,
+ /*   910 */   101,  225,  295,  225,  256,  255,  299,  225,  262,  263,
+ /*   920 */   264,  265,  225,  225,  214,  225,  214,  232,  225,  225,
+ /*   930 */   225,   20,  232,  225,  225,  126,  225,  232,  129,  209,
+ /*   940 */   209,  206,  242,  250,  184,  232,  246,  242,  261,  249,
+ /*   950 */   243,  246,  183,  144,  249,  146,  300,  301,  242,  292,
+ /*   960 */   156,  247,  262,  263,  264,  265,  204,  262,  263,  264,
+ /*   970 */   265,  204,  292,  246,  191,  247,  175,  179,  246,  190,
+ /*   980 */   289,  232,  246,  178,   20,   69,  115,  225,  198,  260,
+ /*   990 */   195,  273,  225,  193,  232,  247,  246,  244,  232,  232,
+ /*  1000 */   246,  246,  285,  288,  242,  286,  247,  204,  246,  242,
+ /*  1010 */   269,  127,  232,  246,  243,  204,  249,  209,   69,  209,
+ /*  1020 */   221,  298,  210,  209,  262,  263,  264,  265,  225,  262,
+ /*  1030 */   263,  264,  265,  214,  228,  232,  225,  297,  214,  206,
+ /*  1040 */   219,  204,  303,  232,  219,  242,  253,  211,    0,  246,
+ /*  1050 */   204,  202,  249,  242,    0,   60,  294,  246,    0,   38,
+ /*  1060 */   151,   38,  225,  151,   38,  262,  263,  264,  265,  232,
+ /*  1070 */    38,  225,    0,  262,  263,  264,  265,  204,  232,  242,
+ /*  1080 */    38,   38,  151,  246,    0,  204,    0,   38,  242,    0,
+ /*  1090 */    38,   69,  246,  142,  141,  104,    0,  138,  225,  262,
+ /*  1100 */   263,  264,  265,    0,  134,  232,  225,   50,  262,  263,
+ /*  1110 */   264,  265,    0,  232,    0,  242,   81,    0,    0,  246,
  /*  1120 */     0,  204,    0,  242,    0,    0,  204,  246,    0,    0,
- /*  1130 */     0,    0,  115,    0,    0,  262,  263,  264,  265,    0,
+ /*  1130 */     0,    0,    0,    0,    0,  262,  263,  264,  265,  115,
  /*  1140 */     0,    0,  225,  262,  263,  264,  265,  225,    0,  232,
  /*  1150 */     0,    0,    0,    0,  232,    0,    0,   22,    0,  242,
- /*  1160 */     0,    0,  204,  246,  242,    0,   43,  204,  246,   48,
- /*  1170 */     0,    0,    0,  204,   38,   43,    0,    0,   36,  262,
- /*  1180 */   263,  264,  265,  225,  262,  263,  264,  265,  225,    0,
- /*  1190 */   232,   78,   38,   67,  225,  232,   22,    0,    0,    0,
- /*  1200 */   242,  232,    0,   76,  246,  242,   38,   38,   22,  246,
- /*  1210 */    22,  242,   39,   38,  204,  246,   38,   38,   38,    0,
- /*  1220 */   262,  263,  264,  265,    0,  262,  263,  264,  265,  204,
- /*  1230 */    22,  262,  263,  264,  265,  225,    0,   22,   12,   13,
- /*  1240 */   204,   38,  232,    0,   22,   20,    0,  147,   22,  130,
- /*  1250 */   225,   38,  242,    0,   22,  204,  246,  232,  143,   57,
- /*  1260 */    22,  225,    0,   61,   38,  127,    0,  242,  232,    0,
- /*  1270 */    38,  246,  262,  263,  264,  265,  225,  130,  242,  130,
- /*  1280 */     4,   57,  246,  232,   82,   61,   69,  262,  263,  264,
- /*  1290 */   265,   43,  125,  242,   67,   19,   67,  246,  262,  263,
- /*  1300 */   264,  265,   67,  101,  102,  103,   82,  105,   69,   33,
- /*  1310 */    70,   70,   36,  262,  263,  264,  265,   41,   92,   69,
- /*  1320 */    44,   70,   70,  180,   92,  101,  102,  103,   69,  105,
- /*  1330 */   104,   67,   69,  174,   67,   81,  104,  180,   70,   81,
- /*  1340 */    70,   67,   81,   70,   68,   67,  180,   71,    4,   38,
- /*  1350 */    38,   38,   38,   38,   38,    2,  155,   70,   81,   69,
- /*  1360 */    69,  128,   70,   70,  138,   69,   69,   69,    0,   43,
- /*  1370 */   138,   69,  125,   79,   22,   69,   81,   70,   38,   38,
- /*  1380 */    81,  155,  156,   80,   38,   38,   69,  155,  156,   70,
- /*  1390 */    69,   22,   70,   69,   38,   70,   69,   38,   70,   69,
- /*  1400 */    69,   38,   94,   94,   94,   82,   69,   69,  104,   38,
- /*  1410 */    22,   47,   94,   48,   22,   38,   38,   38,   38,   38,
- /*  1420 */    38,   38,   22,   38,   38,   38,    0,   38,   38,   38,
- /*  1430 */     0,   38,   38,   38,   38,   36,   38,    0,   37,    0,
- /*  1440 */    43,   22,   21,  304,   22,   22,   21,   20,  304,  304,
- /*  1450 */   304,  304,  304,  304,  304,  304,  304,  304,  304,  304,
- /*  1460 */   304,  304,  304,  304,  304,  304,  304,  304,  304,  304,
+ /*  1160 */     0,    0,  204,  246,  242,    0,    0,  204,  246,    0,
+ /*  1170 */    43,    0,    0,  204,   48,    0,   38,    0,   36,  262,
+ /*  1180 */   263,  264,  265,  225,  262,  263,  264,  265,  225,   43,
+ /*  1190 */   232,    0,    0,    0,  225,  232,   22,   78,    0,   76,
+ /*  1200 */   242,  232,   38,   67,  246,  242,   38,    0,   38,  246,
+ /*  1210 */    22,  242,    0,    0,  204,  246,   38,   38,   38,   38,
+ /*  1220 */   262,  263,  264,  265,   22,  262,  263,  264,  265,  204,
+ /*  1230 */    22,  262,  263,  264,  265,  225,   39,   38,   12,   13,
+ /*  1240 */   204,   22,  232,    0,    0,   22,    0,   20,   22,    0,
+ /*  1250 */   225,  130,  242,  147,   38,  204,  246,  232,    0,  143,
+ /*  1260 */    22,  225,    0,   22,   38,    0,    0,  242,  232,  130,
+ /*  1270 */   127,  246,  262,  263,  264,  265,  225,  130,  242,   38,
+ /*  1280 */     0,  204,  246,  232,   69,   43,   69,  262,  263,  264,
+ /*  1290 */   265,   67,    4,  242,   69,   67,   67,  246,  262,  263,
+ /*  1300 */   264,  265,  225,   70,   70,   69,   57,   19,  125,  232,
+ /*  1310 */    61,   81,   67,  262,  263,  264,  265,  180,   92,  242,
+ /*  1320 */    70,   33,   57,  246,   36,   69,   61,   67,   81,   41,
+ /*  1330 */   104,   82,   44,   92,  180,   81,  174,   70,   70,  262,
+ /*  1340 */   263,  264,  265,   67,   70,  104,   67,   82,    4,   38,
+ /*  1350 */   101,  102,  103,  180,  105,   38,   68,   38,   38,   71,
+ /*  1360 */    38,   38,    2,   81,  138,  155,  101,  102,  103,   70,
+ /*  1370 */   105,   69,   69,    0,   70,   70,   69,   69,   69,  138,
+ /*  1380 */   128,  155,  156,   43,  125,   69,   22,   81,   70,   79,
+ /*  1390 */    69,   38,   38,   81,   80,   69,  155,  156,   70,   38,
+ /*  1400 */    69,   38,   70,   69,   38,   70,   69,   38,   70,   94,
+ /*  1410 */    69,   22,   82,   38,   69,   69,   69,  104,   94,   94,
+ /*  1420 */    38,   94,   22,   22,   48,   47,   38,   22,   38,   38,
+ /*  1430 */    38,   38,   38,   38,   38,   38,   38,   38,    0,    0,
+ /*  1440 */    38,   38,   38,    0,   38,   38,   38,   36,   38,   37,
+ /*  1450 */     0,   22,   43,   21,  304,   22,   22,   21,  304,  304,
+ /*  1460 */    20,  304,  304,  304,  304,  304,  304,  304,  304,  304,
  /*  1470 */   304,  304,  304,  304,  304,  304,  304,  304,  304,  304,
  /*  1480 */   304,  304,  304,  304,  304,  304,  304,  304,  304,  304,
  /*  1490 */   304,  304,  304,  304,  304,  304,  304,  304,  304,  304,
@@ -522,91 +524,93 @@ static const YYCODETYPE yy_lookahead[] = {
  /*  1610 */   304,  304,  304,  304,  304,  304,  304,  304,  304,  304,
  /*  1620 */   304,  304,  304,  304,  304,  304,  304,  304,  304,  304,
  /*  1630 */   304,  304,  304,  304,  304,  304,  304,  304,  304,  304,
- /*  1640 */   304,  304,  304,  304,  304,  304,  304,  304,  304,
+ /*  1640 */   304,  304,  304,  304,  304,  304,  304,  304,  304,  304,
+ /*  1650 */   304,  304,  304,  304,  304,  304,  304,  304,  304,  304,
+ /*  1660 */   304,  304,
 };
 #define YY_SHIFT_COUNT    (510)
 #define YY_SHIFT_MIN      (0)
-#define YY_SHIFT_MAX      (1439)
+#define YY_SHIFT_MAX      (1450)
 static const unsigned short int yy_shift_ofst[] = {
  /*     0 */   598,    0,   39,   78,   78,   78,   78,  234,   78,   78,
  /*    10 */   312,  468,  120,  273,  312,  312,  312,  312,  312,  312,
  /*    20 */   312,  312,  312,  312,  312,  312,  312,  312,  312,  312,
- /*    30 */   312,  312,  312,  206,  206,  206,  159, 1226, 1226,   34,
- /*    40 */    81,   81,  125, 1226,   81,   81,   81,   81,   81,   81,
- /*    50 */   360,  395,  465,  465,  125,  529,  395,   81,   81,  395,
- /*    60 */    81,  395,  529,  395,  395,   81,  512,  148,  431,  411,
- /*    70 */   411,  252,  425, 1232,  240, 1232, 1232, 1232, 1232, 1232,
- /*    80 */  1232, 1232, 1232, 1232, 1232, 1232, 1232, 1232, 1232, 1232,
- /*    90 */  1232, 1232, 1232, 1232,   38,  409,   23,   23,   23,  650,
- /*   100 */   568,  529,  395,  395,  395,  597,  271,  271,  271,  271,
- /*   110 */   271,  271,  809,  288,  402,  372,  119,  477,   63,  663,
- /*   120 */   432,  444,  267,  444,  595,  533,  537,  775,  771,  779,
- /*   130 */   686,  775,  775,  715,  711,  711,  775,  828,  529,  854,
- /*   140 */   360,  568,  858,  360,  360,  775,  360,  877,  395,  395,
- /*   150 */   395,  395,  395,  395,  395,  395,  395,  395,  395,  779,
- /*   160 */   779,  775,  877,  568,  828,  783,  529,  854,  512,  568,
- /*   170 */   858,  512,  911,  757,  761,  796,  757,  761,  796,  796,
- /*   180 */   768,  782,  795,  799,  804,  568,  961,  867,  788,  790,
- /*   190 */   797,  926,  395,  761,  796,  796,  761,  796,  874,  568,
- /*   200 */   858,  512,  597,  512,  568,  948,  779,  779,  775,  512,
- /*   210 */   877, 1448, 1448, 1448, 1448, 1448,  605,  578,  491, 1276,
- /*   220 */  1202, 1224,   16,  108,  333,  136,  136,  136,  136,  136,
- /*   230 */   136,  136,  149,   36,  482,  426,  359,  359,  359,  359,
- /*   240 */   383,  361,  608,  622,  624,  659,  667,  590,  734,  747,
- /*   250 */   681,  669,  701,  706,  646,  522,  621,  586,  720,  643,
- /*   260 */   721,  100,  722,  726,  744,  751,  764,  668,  737,  785,
- /*   270 */   787,  791,  792,  816,  817,  184, 1046, 1047,  988, 1054,
- /*   280 */  1013,  904, 1020, 1021, 1022,  910, 1063, 1026, 1032,  921,
- /*   290 */  1080, 1043, 1082, 1048, 1084, 1018,  947,  949,  987,  955,
- /*   300 */  1094, 1095, 1053,  962, 1097, 1104, 1031, 1107, 1114, 1116,
- /*   310 */  1117, 1118, 1120, 1122, 1124, 1125, 1128, 1129, 1130, 1131,
- /*   320 */  1017, 1133, 1134, 1139, 1140, 1141, 1148, 1135, 1150, 1151,
- /*   330 */  1152, 1153, 1155, 1156, 1158, 1160, 1161, 1123, 1165, 1121,
- /*   340 */  1170, 1171, 1136, 1142, 1132, 1172, 1176, 1177, 1189, 1113,
- /*   350 */  1127, 1154, 1126, 1174, 1197, 1168, 1169, 1175, 1178, 1126,
- /*   360 */  1179, 1180, 1198, 1186, 1199, 1188, 1173, 1219, 1208, 1203,
- /*   370 */  1236, 1215, 1243, 1222, 1225, 1246, 1119, 1100, 1213, 1253,
- /*   380 */  1115, 1238, 1147, 1138, 1262, 1266, 1149, 1269, 1217, 1248,
- /*   390 */  1167, 1227, 1229, 1143, 1240, 1235, 1241, 1239, 1250, 1251,
- /*   400 */  1259, 1252, 1264, 1254, 1263, 1267, 1157, 1268, 1270, 1258,
- /*   410 */  1159, 1274, 1261, 1273, 1278, 1166, 1344, 1311, 1312, 1313,
- /*   420 */  1314, 1315, 1316, 1353, 1201, 1277, 1287, 1290, 1291, 1292,
- /*   430 */  1293, 1296, 1297, 1233, 1298, 1368, 1326, 1247, 1302, 1294,
- /*   440 */  1295, 1299, 1352, 1306, 1303, 1307, 1340, 1341, 1317, 1319,
- /*   450 */  1346, 1321, 1322, 1347, 1324, 1325, 1356, 1327, 1328, 1359,
- /*   460 */  1330, 1308, 1309, 1310, 1318, 1369, 1323, 1331, 1363, 1304,
- /*   470 */  1337, 1338, 1371, 1126, 1388, 1365, 1364, 1392, 1377, 1378,
- /*   480 */  1379, 1380, 1381, 1382, 1383, 1400, 1385, 1126, 1386, 1387,
- /*   490 */  1389, 1390, 1391, 1393, 1394, 1395, 1426, 1396, 1399, 1397,
- /*   500 */  1430, 1398, 1401, 1437, 1439, 1419, 1421, 1422, 1423, 1425,
- /*   510 */  1427,
+ /*    30 */   312,  312,  312,  312,  206,  206,  206,  159, 1226, 1226,
+ /*    40 */    34,   81,   81,   15, 1226,   81,   81,   81,   81,   81,
+ /*    50 */    81,  207,  324,  359,   15,  389,  324,   81,   81,  324,
+ /*    60 */    81,  324,  389,  324,  324,   81,  284,  148,  431,  411,
+ /*    70 */   411,  252,  425, 1241,  240, 1241, 1241, 1241, 1241, 1241,
+ /*    80 */  1241, 1241, 1241, 1241, 1241, 1241, 1241, 1241, 1241, 1241,
+ /*    90 */  1241, 1241, 1241, 1241,   38,  217,   23,   23,   23,  244,
+ /*   100 */   465,  389,  324,  324,  324,  478,  271,  271,  271,  271,
+ /*   110 */   271,  271,  809,  288,  402,  372,  119,  163,  201,  528,
+ /*   120 */   432,  520,  399,  520,  599,  533,  611,  797,  786,  794,
+ /*   130 */   694,  797,  797,  733,  746,  746,  797,  854,  389,  859,
+ /*   140 */   207,  465,  870,  207,  207,  797,  207,  883,  324,  324,
+ /*   150 */   324,  324,  324,  324,  324,  324,  324,  324,  324,  794,
+ /*   160 */   794,  797,  883,  465,  854,  763,  389,  859,  284,  465,
+ /*   170 */   870,  284,  911,  760,  769,  804,  760,  769,  804,  804,
+ /*   180 */   783,  789,  798,  805,  801,  465,  964,  871,  790,  795,
+ /*   190 */   800,  916,  324,  769,  804,  804,  769,  804,  884,  465,
+ /*   200 */   870,  284,  478,  284,  465,  949,  794,  794,  797,  284,
+ /*   210 */   883, 1461, 1461, 1461, 1461, 1461,  605,  578,  491, 1288,
+ /*   220 */  1249, 1265,  333,   16,  108,  577,  136,  136,  136,  136,
+ /*   230 */   136,  136,  136,  307,   36,  482,  369,  593,  593,  593,
+ /*   240 */   593,  652,  361,  630,  625,  650,  659,  663,  721,  722,
+ /*   250 */   734,  681,  701,  702,  706,  752,  582,  496,  581,  720,
+ /*   260 */   628,  724,  100,  725,  726,  744,  745,  751,  612,  737,
+ /*   270 */   764,  785,  787,  791,  792,  816,  676, 1048, 1054,  995,
+ /*   280 */  1058, 1021,  909, 1023, 1026, 1032,  912, 1072, 1042, 1043,
+ /*   290 */   931, 1084, 1049, 1086, 1052, 1089, 1022,  951,  953,  991,
+ /*   300 */   959, 1096, 1103, 1057,  970, 1112, 1114, 1035, 1117, 1118,
+ /*   310 */  1120, 1122, 1124, 1125, 1128, 1129, 1130, 1131, 1132, 1133,
+ /*   320 */  1134, 1024, 1140, 1141, 1148, 1150, 1151, 1152, 1135, 1153,
+ /*   330 */  1155, 1156, 1158, 1160, 1161, 1165, 1166, 1169, 1127, 1171,
+ /*   340 */  1126, 1172, 1175, 1138, 1142, 1146, 1177, 1191, 1192, 1193,
+ /*   350 */  1119, 1123, 1164, 1136, 1174, 1198, 1168, 1170, 1178, 1179,
+ /*   360 */  1136, 1180, 1181, 1207, 1188, 1212, 1202, 1197, 1213, 1208,
+ /*   370 */  1199, 1243, 1219, 1244, 1223, 1227, 1246, 1121, 1106, 1216,
+ /*   380 */  1258, 1116, 1238, 1139, 1143, 1262, 1266, 1147, 1280, 1215,
+ /*   390 */  1242, 1183, 1224, 1228, 1137, 1233, 1229, 1234, 1217, 1225,
+ /*   400 */  1236, 1250, 1245, 1230, 1256, 1260, 1154, 1267, 1268, 1247,
+ /*   410 */  1162, 1276, 1254, 1274, 1279, 1173, 1344, 1311, 1317, 1319,
+ /*   420 */  1320, 1322, 1323, 1360, 1210, 1282, 1299, 1302, 1303, 1304,
+ /*   430 */  1305, 1307, 1308, 1252, 1309, 1373, 1340, 1259, 1316, 1310,
+ /*   440 */  1306, 1312, 1364, 1321, 1314, 1318, 1353, 1354, 1326, 1328,
+ /*   450 */  1361, 1331, 1332, 1363, 1334, 1335, 1366, 1337, 1338, 1369,
+ /*   460 */  1341, 1315, 1324, 1325, 1327, 1389, 1330, 1345, 1375, 1313,
+ /*   470 */  1346, 1347, 1382, 1136, 1400, 1376, 1378, 1401, 1388, 1390,
+ /*   480 */  1391, 1392, 1393, 1394, 1395, 1405, 1396, 1136, 1397, 1398,
+ /*   490 */  1399, 1402, 1403, 1404, 1406, 1407, 1438, 1408, 1411, 1409,
+ /*   500 */  1439, 1410, 1412, 1443, 1450, 1429, 1432, 1433, 1434, 1436,
+ /*   510 */  1440,
 };
 #define YY_REDUCE_COUNT (215)
 #define YY_REDUCE_MIN   (-226)
-#define YY_REDUCE_MAX   (1051)
+#define YY_REDUCE_MAX   (1077)
 static const short yy_reduce_ofst[] = {
  /*     0 */  -180,  262,  495, -198,  -58,  447,  538,  574, -129,  601,
  /*    10 */   656,  -14,  617,  700,  705,  762,  176,  767,  803,  811,
  /*    20 */   837,  846,  873,  881,  917,  922,  958,  963,  969, 1010,
- /*    30 */  1025, 1036, 1051, -195, -170, -147,  197, -213, -204,   41,
- /*    40 */  -209, -206,  247, -132, -208, -175,   -1,   11,  111,  121,
- /*    50 */   318, -142, -176,  -40,  118,  -62, -214,  189,  335,   -6,
- /*    60 */   354,  232,  110,   97,  314,  355,  -55, -201, -156, -156,
- /*    70 */  -156,  173, -185, -107,  126,  208,  259,  308,  310,  339,
- /*    80 */   362,  371,  388,  389,  390,  391,  406,  434,  435,  436,
- /*    90 */   437,  438,  440,  441,  122,  281,  329,  413,  414,  467,
- /*   100 */   -10,  172,   65, -149,  338,  161, -226, -221,  387,  490,
- /*   110 */   497,  500,  454,  539,  530,  451,  456,  519,  475,  548,
- /*   120 */   503,  501,  501,  501,  552,  487,  511,  587,  555,  599,
- /*   130 */   571,  614,  619,  592,  602,  606,  636,  594,  628,  625,
- /*   140 */   664,  647,  642,  677,  678,  685,  679,  694,  676,  680,
- /*   150 */   682,  683,  688,  689,  690,  692,  697,  698,  699,  712,
- /*   160 */   713,  693,  723,  674,  655,  684,  691,  695,  725,  704,
- /*   170 */   707,  731,  687,  651,  714,  727,  666,  728,  730,  732,
- /*   180 */   671,  696,  702,  708,  501,  748,  729,  710,  703,  716,
- /*   190 */   724,  733,  552,  749,  745,  752,  750,  754,  759,  773,
- /*   200 */   769,  801,  798,  802,  786,  794,  806,  819,  813,  825,
- /*   210 */   831,  789,  820,  821,  827,  842,
+ /*    30 */  1025, 1036, 1051, 1077, -195, -170, -147,  197, -213, -204,
+ /*    40 */    41, -209, -206,  247, -132, -208, -175,   -1,   11,  111,
+ /*    50 */   121,  318, -142, -176,  118,  -62, -214,  189,  335,   -6,
+ /*    60 */   354,  232,  -49,   97,  314,  355,  -55, -201, -156, -156,
+ /*    70 */  -156,  171, -185, -107,  126,  106,  169,  208,  259,  308,
+ /*    80 */   310,  339,  364,  371,  390,  391,  430,  434,  435,  436,
+ /*    90 */   437,  438,  440,  441,  212,  329,  375,  413,  414,  469,
+ /*   100 */   -10,  370,  181, -149,  313,   80, -226, -221,  229,  362,
+ /*   110 */   451,  459,  452,  503,  499,  445,  456,  505,  475,  557,
+ /*   120 */   536,  509,  509,  509,  573,  511,  530,  613,  567,  615,
+ /*   130 */   583,  634,  636,  600,  637,  638,  640,  622,  642,  635,
+ /*   140 */   675,  657,  648,  683,  685,  690,  689,  699,  682,  686,
+ /*   150 */   688,  692,  697,  698,  703,  704,  708,  709,  711,  710,
+ /*   160 */   712,  696,  735,  695,  658,  660,  716,  693,  730,  713,
+ /*   170 */   707,  731,  687,  667,  714,  727,  680,  728,  732,  736,
+ /*   180 */   691,  715,  719,  717,  509,  749,  729,  718,  739,  723,
+ /*   190 */   740,  741,  573,  748,  750,  754,  759,  755,  753,  766,
+ /*   200 */   771,  808,  799,  810,  780,  806,  819,  824,  812,  814,
+ /*   210 */   833,  793,  821,  825,  836,  849,
 };
 static const YYACTIONTYPE yy_default[] = {
  /*     0 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
@@ -614,7 +618,7 @@ static const YYACTIONTYPE yy_default[] = {
  /*    20 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
  /*    30 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
  /*    40 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
- /*    50 */  1205, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
+ /*    50 */  1152, 1205, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
  /*    60 */  1152, 1152, 1152, 1152, 1152, 1152, 1203, 1332, 1152, 1464,
  /*    70 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
  /*    80 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
@@ -631,11 +635,11 @@ static const YYACTIONTYPE yy_default[] = {
  /*   190 */  1526, 1152, 1152, 1504, 1152, 1152, 1504, 1152, 1345, 1152,
  /*   200 */  1152, 1203, 1152, 1203, 1152, 1259, 1152, 1152, 1152, 1203,
  /*   210 */  1152, 1360, 1293, 1293, 1206, 1157, 1152, 1152, 1152, 1152,
- /*   220 */  1152, 1152, 1152, 1152, 1152, 1428, 1517, 1516, 1427, 1441,
- /*   230 */  1440, 1439, 1152, 1152, 1152, 1152, 1422, 1423, 1421, 1420,
- /*   240 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
- /*   250 */  1152, 1152, 1152, 1152, 1465, 1152, 1527, 1531, 1152, 1152,
- /*   260 */  1152, 1405, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
+ /*   220 */  1152, 1152, 1152, 1152, 1152, 1152, 1428, 1517, 1516, 1427,
+ /*   230 */  1441, 1440, 1439, 1152, 1152, 1152, 1152, 1422, 1423, 1421,
+ /*   240 */  1420, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
+ /*   250 */  1152, 1152, 1152, 1152, 1152, 1465, 1152, 1527, 1531, 1152,
+ /*   260 */  1152, 1152, 1405, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
  /*   270 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
  /*   280 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
  /*   290 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
@@ -644,11 +648,11 @@ static const YYACTIONTYPE yy_default[] = {
  /*   320 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
  /*   330 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
  /*   340 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
- /*   350 */  1152, 1152, 1304, 1152, 1152, 1152, 1152, 1152, 1152, 1229,
- /*   360 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
+ /*   350 */  1152, 1152, 1152, 1304, 1152, 1152, 1152, 1152, 1152, 1152,
+ /*   360 */  1229, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
  /*   370 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
  /*   380 */  1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
- /*   390 */  1152, 1488, 1498, 1152, 1152, 1152, 1152, 1152, 1152, 1152,
+ /*   390 */  1152, 1152, 1488, 1498, 1152, 1152, 1152, 1152, 1152, 1152,
  /*   400 */  1152, 1152, 1152, 1405, 1152, 1515, 1152, 1474, 1470, 1152,
  /*   410 */  1152, 1466, 1152, 1152, 1525, 1152, 1152, 1152, 1152, 1152,
  /*   420 */  1152, 1152, 1152, 1460, 1152, 1152, 1152, 1152, 1152, 1152,
@@ -1424,7 +1428,7 @@ static const char *const yyRuleName[] = {
  /* 343 */ "partition_by_clause_opt ::= PARTITION BY expression_list",
  /* 344 */ "twindow_clause_opt ::=",
  /* 345 */ "twindow_clause_opt ::= SESSION NK_LP column_reference NK_COMMA duration_literal NK_RP",
- /* 346 */ "twindow_clause_opt ::= STATE_WINDOW NK_LP column_reference NK_RP",
+ /* 346 */ "twindow_clause_opt ::= STATE_WINDOW NK_LP expression NK_RP",
  /* 347 */ "twindow_clause_opt ::= INTERVAL NK_LP duration_literal NK_RP sliding_opt fill_opt",
  /* 348 */ "twindow_clause_opt ::= INTERVAL NK_LP duration_literal NK_COMMA duration_literal NK_RP sliding_opt fill_opt",
  /* 349 */ "sliding_opt ::=",
@@ -2385,7 +2389,7 @@ static const struct {
   {  286,   -3 }, /* (343) partition_by_clause_opt ::= PARTITION BY expression_list */
   {  287,    0 }, /* (344) twindow_clause_opt ::= */
   {  287,   -6 }, /* (345) twindow_clause_opt ::= SESSION NK_LP column_reference NK_COMMA duration_literal NK_RP */
-  {  287,   -4 }, /* (346) twindow_clause_opt ::= STATE_WINDOW NK_LP column_reference NK_RP */
+  {  287,   -4 }, /* (346) twindow_clause_opt ::= STATE_WINDOW NK_LP expression NK_RP */
   {  287,   -6 }, /* (347) twindow_clause_opt ::= INTERVAL NK_LP duration_literal NK_RP sliding_opt fill_opt */
   {  287,   -8 }, /* (348) twindow_clause_opt ::= INTERVAL NK_LP duration_literal NK_COMMA duration_literal NK_RP sliding_opt fill_opt */
   {  247,    0 }, /* (349) sliding_opt ::= */
@@ -3590,7 +3594,7 @@ static YYACTIONTYPE yy_reduce(
       case 345: /* twindow_clause_opt ::= SESSION NK_LP column_reference NK_COMMA duration_literal NK_RP */
 { yymsp[-5].minor.yy104 = createSessionWindowNode(pCxt, releaseRawExprNode(pCxt, yymsp[-3].minor.yy104), releaseRawExprNode(pCxt, yymsp[-1].minor.yy104)); }
         break;
-      case 346: /* twindow_clause_opt ::= STATE_WINDOW NK_LP column_reference NK_RP */
+      case 346: /* twindow_clause_opt ::= STATE_WINDOW NK_LP expression NK_RP */
 { yymsp[-3].minor.yy104 = createStateWindowNode(pCxt, releaseRawExprNode(pCxt, yymsp[-1].minor.yy104)); }
         break;
       case 347: /* twindow_clause_opt ::= INTERVAL NK_LP duration_literal NK_RP sliding_opt fill_opt */
diff --git a/source/libs/planner/src/planLogicCreater.c b/source/libs/planner/src/planLogicCreater.c
index 149071bad2..819eb6d699 100644
--- a/source/libs/planner/src/planLogicCreater.c
+++ b/source/libs/planner/src/planLogicCreater.c
@@ -477,6 +477,18 @@ static int32_t createWindowLogicNodeFinalize(SLogicPlanContext* pCxt, SSelectStm
   return code;
 }
 
+static int32_t createWindowLogicNodeByState(SLogicPlanContext* pCxt, SStateWindowNode* pState, SSelectStmt* pSelect, SLogicNode** pLogicNode) {
+  SWindowLogicNode* pWindow = nodesMakeNode(QUERY_NODE_LOGIC_PLAN_WINDOW);
+  if (NULL == pWindow) {
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
+
+  pWindow->winType = WINDOW_TYPE_STATE;
+  pWindow->pStateExpr = nodesCloneNode(pState->pExpr);
+
+  return createWindowLogicNodeFinalize(pCxt, pSelect, pWindow, pLogicNode);
+}
+
 static int32_t createWindowLogicNodeBySession(SLogicPlanContext* pCxt, SSessionWindowNode* pSession, SSelectStmt* pSelect, SLogicNode** pLogicNode) {
   SWindowLogicNode* pWindow = nodesMakeNode(QUERY_NODE_LOGIC_PLAN_WINDOW);
   if (NULL == pWindow) {
@@ -525,6 +537,8 @@ static int32_t createWindowLogicNode(SLogicPlanContext* pCxt, SSelectStmt* pSele
   }
 
   switch (nodeType(pSelect->pWindow)) {
+    case QUERY_NODE_STATE_WINDOW:
+      return createWindowLogicNodeByState(pCxt, (SStateWindowNode*)pSelect->pWindow, pSelect, pLogicNode);
     case QUERY_NODE_SESSION_WINDOW:
       return createWindowLogicNodeBySession(pCxt, (SSessionWindowNode*)pSelect->pWindow, pSelect, pLogicNode);
     case QUERY_NODE_INTERVAL_WINDOW:
@@ -642,6 +656,29 @@ static int32_t createPartitionLogicNode(SLogicPlanContext* pCxt, SSelectStmt* pS
     return TSDB_CODE_OUT_OF_MEMORY;
   }
 
+  SNodeList* pCols = NULL;
+  int32_t code = nodesCollectColumns(pSelect, SQL_CLAUSE_PARTITION_BY, NULL, &pCols);
+  if (TSDB_CODE_SUCCESS == code && NULL != pCols) {
+    pPartition->node.pTargets = nodesCloneList(pCols);
+    if (NULL == pPartition->node.pTargets) {
+      code = TSDB_CODE_OUT_OF_MEMORY;
+    }
+  }
+
+  if (TSDB_CODE_SUCCESS == code) {
+    pPartition->pPartitionKeys = nodesCloneList(pSelect->pPartitionByList);
+    if (NULL == pPartition->pPartitionKeys) {
+      code = TSDB_CODE_OUT_OF_MEMORY;
+    }
+  }
+
+  if (TSDB_CODE_SUCCESS == code) {
+    *pLogicNode = (SLogicNode*)pPartition;
+  } else {
+    nodesDestroyNode(pPartition);
+  }
+
+  return code;
   return TSDB_CODE_SUCCESS;
 }
 
@@ -650,7 +687,35 @@ static int32_t createDistinctLogicNode(SLogicPlanContext* pCxt, SSelectStmt* pSe
     return TSDB_CODE_SUCCESS;
   }
 
-  return TSDB_CODE_SUCCESS;
+  SAggLogicNode* pAgg = (SAggLogicNode*)nodesMakeNode(QUERY_NODE_LOGIC_PLAN_AGG);
+  if (NULL == pAgg) {
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
+
+  int32_t code = TSDB_CODE_SUCCESS;
+  // set grouyp keys, agg funcs and having conditions
+  pAgg->pGroupKeys = nodesCloneList(pSelect->pProjectionList);
+  if (NULL == pAgg->pGroupKeys) {
+    code = TSDB_CODE_OUT_OF_MEMORY;
+  }
+
+  // rewrite the expression in subsequent clauses
+  if (TSDB_CODE_SUCCESS == code) {
+    code = rewriteExpr(pAgg->pGroupKeys, pSelect, SQL_CLAUSE_SELECT);
+  }
+
+  // set the output
+  if (TSDB_CODE_SUCCESS == code) {
+    code = createColumnByRewriteExps(pCxt, pAgg->pGroupKeys, &pAgg->node.pTargets);
+  }
+
+  if (TSDB_CODE_SUCCESS == code) {
+    *pLogicNode = (SLogicNode*)pAgg;
+  } else {
+    nodesDestroyNode(pAgg);
+  }
+
+  return code;
 }
 
 static int32_t createSelectLogicNode(SLogicPlanContext* pCxt, SSelectStmt* pSelect, SLogicNode** pLogicNode) {
diff --git a/source/libs/planner/src/planPhysiCreater.c b/source/libs/planner/src/planPhysiCreater.c
index 04fe0d7da4..965c5c9579 100644
--- a/source/libs/planner/src/planPhysiCreater.c
+++ b/source/libs/planner/src/planPhysiCreater.c
@@ -176,6 +176,10 @@ static int16_t getUnsetSlotId(const SArray* pSlotIdsInfo) {
 }
 
 static int32_t addDataBlockSlotsImpl(SPhysiPlanContext* pCxt, SNodeList* pList, SDataBlockDescNode* pDataBlockDesc, const char* pStmtName, bool output) {
+  if (NULL == pList) {
+    return TSDB_CODE_SUCCESS;
+  }
+
   int32_t code = TSDB_CODE_SUCCESS;
   SHashObj* pHash = taosArrayGetP(pCxt->pLocationHelper, pDataBlockDesc->dataBlockId);
   int16_t nextSlotId = taosHashGetSize(pHash), slotId = 0;
@@ -219,6 +223,23 @@ static int32_t addDataBlockSlots(SPhysiPlanContext* pCxt, SNodeList* pList, SDat
   return addDataBlockSlotsImpl(pCxt, pList, pDataBlockDesc, NULL, false);
 }
 
+static int32_t addDataBlockSlot(SPhysiPlanContext* pCxt, SNode** pNode, SDataBlockDescNode* pDataBlockDesc) {
+  if (NULL == pNode || NULL == *pNode) {
+    return TSDB_CODE_SUCCESS;
+  }
+
+  SNodeList* pList = NULL;
+  int32_t code = nodesListMakeAppend(&pList, *pNode);
+  if (TSDB_CODE_SUCCESS == code) {
+    code = addDataBlockSlots(pCxt, pList, pDataBlockDesc);
+  }
+  if (TSDB_CODE_SUCCESS == code) {
+    *pNode = nodesListGetNode(pList, 0);
+  }
+  nodesClearList(pList);
+  return code;
+}
+
 static int32_t addDataBlockSlotsForProject(SPhysiPlanContext* pCxt, const char* pStmtName, SNodeList* pList, SDataBlockDescNode* pDataBlockDesc) {
   return addDataBlockSlotsImpl(pCxt, pList, pDataBlockDesc, pStmtName, true);
 }
@@ -244,6 +265,7 @@ static EDealRes doSetSlotId(SNode* pNode, void* pContext) {
     }
     // pIndex is definitely not NULL, otherwise it is a bug
     if (NULL == pIndex) {
+      pCxt->errCode = TSDB_CODE_PLAN_INTERNAL_ERROR;
       return DEAL_RES_ERROR;
     }
     ((SColumnNode*)pNode)->dataBlockId = pIndex->dataBlockId;
@@ -614,6 +636,25 @@ static int32_t rewritePrecalcExprs(SPhysiPlanContext* pCxt, SNodeList* pList, SN
   return cxt.errCode;
 }
 
+static int32_t rewritePrecalcExpr(SPhysiPlanContext* pCxt, SNode* pNode, SNodeList** pPrecalcExprs, SNode** pRewritten) {
+  if (NULL == pNode) {
+    return TSDB_CODE_SUCCESS;
+  }
+
+  SNodeList* pList = NULL;
+  int32_t code = nodesListMakeAppend(&pList, pNode);
+  SNodeList* pRewrittenList = NULL;
+  if (TSDB_CODE_SUCCESS == code) {
+    code = rewritePrecalcExprs(pCxt, pList, pPrecalcExprs, &pRewrittenList);
+  }
+  if (TSDB_CODE_SUCCESS == code) {
+    *pRewritten = nodesListGetNode(pRewrittenList, 0);
+  }
+  nodesClearList(pList);
+  nodesClearList(pRewrittenList);
+  return code;
+}
+
 static int32_t createAggPhysiNode(SPhysiPlanContext* pCxt, SNodeList* pChildren, SAggLogicNode* pAggLogicNode, SPhysiNode** pPhyNode) {
   SAggPhysiNode* pAgg = (SAggPhysiNode*)makePhysiNode(pCxt, (SLogicNode*)pAggLogicNode, QUERY_NODE_PHYSICAL_PLAN_AGG);
   if (NULL == pAgg) {
@@ -818,6 +859,40 @@ static int32_t createSessionWindowPhysiNode(SPhysiPlanContext* pCxt, SNodeList*
   return createWindowPhysiNodeFinalize(pCxt, pChildren, &pSession->window, pWindowLogicNode, pPhyNode);
 }
 
+static int32_t createStateWindowPhysiNode(SPhysiPlanContext* pCxt, SNodeList* pChildren, SWindowLogicNode* pWindowLogicNode, SPhysiNode** pPhyNode) {
+  SStateWinodwPhysiNode* pState = (SStateWinodwPhysiNode*)makePhysiNode(pCxt, (SLogicNode*)pWindowLogicNode, QUERY_NODE_PHYSICAL_PLAN_STATE_WINDOW);
+  if (NULL == pState) {
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
+
+  SNodeList* pPrecalcExprs = NULL;
+  SNode* pStateKey = NULL;
+  int32_t code = rewritePrecalcExpr(pCxt, pWindowLogicNode->pStateExpr, &pPrecalcExprs, &pStateKey);
+
+  SDataBlockDescNode* pChildTupe = (((SPhysiNode*)nodesListGetNode(pChildren, 0))->pOutputDataBlockDesc);
+  // push down expression to pOutputDataBlockDesc of child node
+  if (TSDB_CODE_SUCCESS == code && NULL != pPrecalcExprs) {
+    code = setListSlotId(pCxt, pChildTupe->dataBlockId, -1, pPrecalcExprs, &pState->window.pExprs);
+    if (TSDB_CODE_SUCCESS == code) {
+      code = addDataBlockSlots(pCxt, pState->window.pExprs, pChildTupe);
+    }
+  }
+
+  if (TSDB_CODE_SUCCESS == code) {
+    code = setNodeSlotId(pCxt, pChildTupe->dataBlockId, -1, pStateKey, &pState->pStateKey);
+    if (TSDB_CODE_SUCCESS == code) {
+      code = addDataBlockSlot(pCxt, &pState->pStateKey, pState->window.node.pOutputDataBlockDesc);
+    }
+  }
+
+  if (TSDB_CODE_SUCCESS != code) {
+    nodesDestroyNode(pState);
+    return code;
+  }
+
+  return createWindowPhysiNodeFinalize(pCxt, pChildren, &pState->window, pWindowLogicNode, pPhyNode);
+}
+
 static int32_t createWindowPhysiNode(SPhysiPlanContext* pCxt, SNodeList* pChildren, SWindowLogicNode* pWindowLogicNode, SPhysiNode** pPhyNode) {
   switch (pWindowLogicNode->winType) {
     case WINDOW_TYPE_INTERVAL:
@@ -825,7 +900,7 @@ static int32_t createWindowPhysiNode(SPhysiPlanContext* pCxt, SNodeList* pChildr
     case WINDOW_TYPE_SESSION:
       return createSessionWindowPhysiNode(pCxt, pChildren, pWindowLogicNode, pPhyNode);
     case WINDOW_TYPE_STATE:
-      break;
+      return createStateWindowPhysiNode(pCxt, pChildren, pWindowLogicNode, pPhyNode);
     default:
       break;
   }
@@ -867,6 +942,41 @@ static int32_t createSortPhysiNode(SPhysiPlanContext* pCxt, SNodeList* pChildren
   return code;
 }
 
+static int32_t createPartitionPhysiNode(SPhysiPlanContext* pCxt, SNodeList* pChildren, SPartitionLogicNode* pPartLogicNode, SPhysiNode** pPhyNode) {
+  SPartitionPhysiNode* pPart = (SPartitionPhysiNode*)makePhysiNode(pCxt, (SLogicNode*)pPartLogicNode, QUERY_NODE_PHYSICAL_PLAN_PARTITION);
+  if (NULL == pPart) {
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
+
+  SNodeList* pPrecalcExprs = NULL;
+  SNodeList* pPartitionKeys = NULL;
+  int32_t code = rewritePrecalcExprs(pCxt, pPartLogicNode->pPartitionKeys, &pPrecalcExprs, &pPartitionKeys);
+
+  SDataBlockDescNode* pChildTupe = (((SPhysiNode*)nodesListGetNode(pChildren, 0))->pOutputDataBlockDesc);
+  // push down expression to pOutputDataBlockDesc of child node
+  if (TSDB_CODE_SUCCESS == code && NULL != pPrecalcExprs) {
+    code = setListSlotId(pCxt, pChildTupe->dataBlockId, -1, pPrecalcExprs, &pPart->pExprs);
+    if (TSDB_CODE_SUCCESS == code) {
+      code = addDataBlockSlots(pCxt, pPart->pExprs, pChildTupe);
+    }
+  }
+
+  if (TSDB_CODE_SUCCESS == code) {
+    code = setListSlotId(pCxt, pChildTupe->dataBlockId, -1, pPartitionKeys, &pPart->pPartitionKeys);
+    if (TSDB_CODE_SUCCESS == code) {
+      code = addDataBlockSlots(pCxt, pPart->pPartitionKeys, pPart->node.pOutputDataBlockDesc);
+    }
+  }
+
+  if (TSDB_CODE_SUCCESS == code) {
+    *pPhyNode = (SPhysiNode*)pPart;
+  } else {
+    nodesDestroyNode(pPart);
+  }
+
+  return code;
+}
+
 static int32_t doCreatePhysiNode(SPhysiPlanContext* pCxt, SLogicNode* pLogicNode, SSubplan* pSubplan, SNodeList* pChildren, SPhysiNode** pPhyNode) {
   switch (nodeType(pLogicNode)) {
     case QUERY_NODE_LOGIC_PLAN_SCAN:
@@ -883,6 +993,8 @@ static int32_t doCreatePhysiNode(SPhysiPlanContext* pCxt, SLogicNode* pLogicNode
       return createWindowPhysiNode(pCxt, pChildren, (SWindowLogicNode*)pLogicNode, pPhyNode);
     case QUERY_NODE_LOGIC_PLAN_SORT:
       return createSortPhysiNode(pCxt, pChildren, (SSortLogicNode*)pLogicNode, pPhyNode);
+    case QUERY_NODE_LOGIC_PLAN_PARTITION:
+      return createPartitionPhysiNode(pCxt, pChildren, (SPartitionLogicNode*)pLogicNode, pPhyNode);
     default:
       break;
   }
diff --git a/source/libs/planner/test/plannerTest.cpp b/source/libs/planner/test/plannerTest.cpp
index 32b221c4fb..7f79fc6a5a 100644
--- a/source/libs/planner/test/plannerTest.cpp
+++ b/source/libs/planner/test/plannerTest.cpp
@@ -217,6 +217,32 @@ TEST_F(PlannerTest, sessionWindow) {
   ASSERT_TRUE(run());
 }
 
+TEST_F(PlannerTest, stateWindow) {
+  setDatabase("root", "test");
+
+  bind("SELECT count(*) FROM t1 state_window(c1)");
+  ASSERT_TRUE(run());
+
+  bind("SELECT count(*) FROM t1 state_window(c1 + 10)");
+  ASSERT_TRUE(run());
+}
+
+TEST_F(PlannerTest, partitionBy) {
+  setDatabase("root", "test");
+
+  bind("SELECT * FROM t1 partition by c1");
+  ASSERT_TRUE(run());
+
+  bind("SELECT count(*) FROM t1 partition by c1");
+  ASSERT_TRUE(run());
+
+  bind("SELECT count(*) FROM t1 partition by c1 group by c2");
+  ASSERT_TRUE(run());
+
+  bind("SELECT count(*) FROM st1 partition by tag1, tag2 interval(10s)");
+  ASSERT_TRUE(run());
+}
+
 TEST_F(PlannerTest, orderBy) {
   setDatabase("root", "test");
 
@@ -230,6 +256,19 @@ TEST_F(PlannerTest, orderBy) {
   ASSERT_TRUE(run());
 }
 
+TEST_F(PlannerTest, distinct) {
+  setDatabase("root", "test");
+
+  bind("SELECT distinct c1 FROM t1");
+  ASSERT_TRUE(run());
+
+  bind("SELECT distinct c1, c2 + 10 FROM t1");
+  ASSERT_TRUE(run());
+
+  bind("SELECT distinct c1 + 10 a FROM t1 order by a");
+  ASSERT_TRUE(run());
+}
+
 TEST_F(PlannerTest, limit) {
   setDatabase("root", "test");
 
diff --git a/source/util/src/terror.c b/source/util/src/terror.c
index 90fa624a8d..38cd5fad4b 100644
--- a/source/util/src/terror.c
+++ b/source/util/src/terror.c
@@ -432,6 +432,9 @@ TAOS_DEFINE_ERROR(TSDB_CODE_SCH_STATUS_ERROR,             "scheduler status erro
 TAOS_DEFINE_ERROR(TSDB_CODE_SCH_INTERNAL_ERROR,           "scheduler internal error")
 TAOS_DEFINE_ERROR(TSDB_CODE_QW_MSG_ERROR,                 "Invalid msg order")
 
+//planner
+TAOS_DEFINE_ERROR(TSDB_CODE_PLAN_INTERNAL_ERROR,          "planner internal error")
+
 #ifdef TAOS_ERROR_C
 };
 #endif

From cf727c9db30104bfc86ffcc8fab4488d25708336 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 07:54:52 +0000
Subject: [PATCH 16/25] fix show tables problem

---
 source/libs/executor/src/executorimpl.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/source/libs/executor/src/executorimpl.c b/source/libs/executor/src/executorimpl.c
index 2ca06cb15f..148d16aaf8 100644
--- a/source/libs/executor/src/executorimpl.c
+++ b/source/libs/executor/src/executorimpl.c
@@ -310,6 +310,10 @@ SSDataBlock* createOutputBuf_rv1(SDataBlockDescNode* pNode) {
     idata.info.slotId = pDescNode->slotId;
     idata.info.precision = pDescNode->dataType.precision;
 
+    if (IS_VAR_DATA_TYPE(idata.info.type)) {
+      pBlock->info.hasVarCol = true;
+    }
+
     taosArrayPush(pBlock->pDataBlock, &idata);
   }
 

From e40dc9b9291b193886ec2757b0c8a2380ba9c022 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 08:06:28 +0000
Subject: [PATCH 17/25] commit meta when close vnode

---
 source/dnode/vnode/src/meta/metaTDBImpl.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/source/dnode/vnode/src/meta/metaTDBImpl.c b/source/dnode/vnode/src/meta/metaTDBImpl.c
index 282e0a51f6..c273f584ff 100644
--- a/source/dnode/vnode/src/meta/metaTDBImpl.c
+++ b/source/dnode/vnode/src/meta/metaTDBImpl.c
@@ -190,6 +190,9 @@ int metaOpenDB(SMeta *pMeta) {
 
 void metaCloseDB(SMeta *pMeta) {
   if (pMeta->pDB) {
+    tdbCommit(pMeta->pDB->pEnv, &pMeta->pDB->txn);
+    tdbTxnClose(&pMeta->pDB->txn);
+    clearPool(pMeta->pDB->pPool);
     tdbDbClose(pMeta->pDB->pCtbIdx);
     tdbDbClose(pMeta->pDB->pNtbIdx);
     tdbDbClose(pMeta->pDB->pStbIdx);

From 610832f0b47d2d535bfaf12c60368708d7b74098 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 08:41:53 +0000
Subject: [PATCH 18/25] keep committed data in TDB

---
 source/libs/tdb/src/db/tdbBtree.c |  3 ++-
 source/libs/tdb/src/db/tdbPager.c | 36 +++++++++++++++++++------------
 source/libs/tdb/src/db/tdbUtil.c  | 13 +++++++++++
 source/libs/tdb/src/inc/tdbOs.h   | 20 +++++++++--------
 source/libs/tdb/src/inc/tdbUtil.h |  1 +
 5 files changed, 49 insertions(+), 24 deletions(-)

diff --git a/source/libs/tdb/src/db/tdbBtree.c b/source/libs/tdb/src/db/tdbBtree.c
index 285434b402..f760598714 100644
--- a/source/libs/tdb/src/db/tdbBtree.c
+++ b/source/libs/tdb/src/db/tdbBtree.c
@@ -297,7 +297,8 @@ static int tdbBtreeOpenImpl(SBTree *pBt) {
 
   {
     // 1. TODO: Search the main DB to check if the DB exists
-    pgno = 0;
+    ret = tdbPagerOpenDB(pBt->pPager, &pgno, true);
+    ASSERT(ret == 0);
   }
 
   if (pgno != 0) {
diff --git a/source/libs/tdb/src/db/tdbPager.c b/source/libs/tdb/src/db/tdbPager.c
index f78568b51c..7672ccf34e 100644
--- a/source/libs/tdb/src/db/tdbPager.c
+++ b/source/libs/tdb/src/db/tdbPager.c
@@ -76,6 +76,8 @@ int tdbPagerOpen(SPCache *pCache, const char *fileName, SPager **ppPager) {
 
   // pPager->jfd = -1;
   pPager->pageSize = tdbPCacheGetPageSize(pCache);
+  // pPager->dbOrigSize
+  ret = tdbGetFileSize(pPager->fd, pPager->pageSize, &(pPager->dbOrigSize));
 
   *ppPager = pPager;
   return 0;
@@ -91,26 +93,32 @@ int tdbPagerOpenDB(SPager *pPager, SPgno *ppgno, bool toCreate) {
   SPage *pPage;
   int    ret;
 
-  {
-    // TODO: try to search the main DB to get the page number
+  if (pPager->dbOrigSize > 0) {
+    pgno = 1;
+  } else {
     pgno = 0;
   }
 
-  // if (pgno == 0 && toCreate) {
-  //   ret = tdbPagerAllocPage(pPager, &pPage, &pgno);
-  //   if (ret < 0) {
-  //     return -1;
-  //   }
+  {
+      // TODO: try to search the main DB to get the page number
+      // pgno = 0;
+  }
 
-  //   // TODO: Need to zero the page
+      // if (pgno == 0 && toCreate) {
+      //   ret = tdbPagerAllocPage(pPager, &pPage, &pgno);
+      //   if (ret < 0) {
+      //     return -1;
+      //   }
 
-  //   ret = tdbPagerWrite(pPager, pPage);
-  //   if (ret < 0) {
-  //     return -1;
-  //   }
-  // }
+      //   // TODO: Need to zero the page
 
-  *ppgno = pgno;
+      //   ret = tdbPagerWrite(pPager, pPage);
+      //   if (ret < 0) {
+      //     return -1;
+      //   }
+      // }
+
+      *ppgno = pgno;
   return 0;
 }
 
diff --git a/source/libs/tdb/src/db/tdbUtil.c b/source/libs/tdb/src/db/tdbUtil.c
index fc299b3fc1..09e126772b 100644
--- a/source/libs/tdb/src/db/tdbUtil.c
+++ b/source/libs/tdb/src/db/tdbUtil.c
@@ -30,5 +30,18 @@ int tdbGnrtFileID(const char *fname, uint8_t *fileid, bool unique) {
     ((uint64_t *)fileid)[2] = taosRand();
   }
 
+  return 0;
+}
+
+int tdbGetFileSize(tdb_fd_t fd, int szPage, SPgno *size) {
+  int     ret;
+  int64_t szBytes;
+
+  ret = tdbOsFileSize(fd, &szBytes);
+  if (ret < 0) {
+    return -1;
+  }
+
+  *size = szBytes / szPage;
   return 0;
 }
\ No newline at end of file
diff --git a/source/libs/tdb/src/inc/tdbOs.h b/source/libs/tdb/src/inc/tdbOs.h
index ae389708f4..1d87285091 100644
--- a/source/libs/tdb/src/inc/tdbOs.h
+++ b/source/libs/tdb/src/inc/tdbOs.h
@@ -46,13 +46,14 @@ typedef TdFilePtr tdb_fd_t;
 
 #define tdbOsOpen(PATH, OPTION, MODE) taosOpenFile((PATH), (OPTION))
 
-#define tdbOsClose(FD) taosCloseFile(&(FD))
-#define tdbOsRead      taosReadFile
-#define tdbOsPRead     taosPReadFile
-#define tdbOsWrite     taosWriteFile
-#define tdbOsFSync     taosFsyncFile
-#define tdbOsLSeek     taosLSeekFile
-#define tdbOsRemove    remove
+#define tdbOsClose(FD)           taosCloseFile(&(FD))
+#define tdbOsRead                taosReadFile
+#define tdbOsPRead               taosPReadFile
+#define tdbOsWrite               taosWriteFile
+#define tdbOsFSync               taosFsyncFile
+#define tdbOsLSeek               taosLSeekFile
+#define tdbOsRemove              remove
+#define tdbOsFileSize(FD, PSIZE) taosFStatFile(FD, PSIZE, NULL)
 
 /* directory */
 #define tdbOsMkdir taosMkDir
@@ -110,10 +111,11 @@ i64 tdbOsWrite(tdb_fd_t fd, const void *pData, i64 nBytes);
 #define tdbOsFSync  fsync
 #define tdbOsLSeek  lseek
 #define tdbOsRemove remove
+#define tdbOsFileSize(FD, PSIZE)
 
 /* directory */
-#define tdbOsMkdir  mkdir
-#define tdbOsRmdir  rmdir
+#define tdbOsMkdir mkdir
+#define tdbOsRmdir rmdir
 
 // For threads and lock -----------------
 /* spin lock */
diff --git a/source/libs/tdb/src/inc/tdbUtil.h b/source/libs/tdb/src/inc/tdbUtil.h
index 6abddb5b22..042026d0a4 100644
--- a/source/libs/tdb/src/inc/tdbUtil.h
+++ b/source/libs/tdb/src/inc/tdbUtil.h
@@ -29,6 +29,7 @@ extern "C" {
 #define TDB_ROUND8(x) (((x) + 7) & ~7)
 
 int tdbGnrtFileID(const char *fname, uint8_t *fileid, bool unique);
+int tdbGetFileSize(tdb_fd_t fd, int szPage, SPgno *size);
 
 #define TDB_REALLOC(PTR, SIZE)                                                               \
   ({                                                                                         \

From e707047d6c9654ba68ebea2a3460b83d06413d28 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 09:39:05 +0000
Subject: [PATCH 19/25] fix a small bug

---
 source/dnode/vnode/src/meta/metaTDBImpl.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/source/dnode/vnode/src/meta/metaTDBImpl.c b/source/dnode/vnode/src/meta/metaTDBImpl.c
index c273f584ff..3ab26c3d4d 100644
--- a/source/dnode/vnode/src/meta/metaTDBImpl.c
+++ b/source/dnode/vnode/src/meta/metaTDBImpl.c
@@ -547,7 +547,7 @@ tb_uid_t metaCtbCursorNext(SMCtbCursor *pCtbCur) {
     return 0;
   }
 
-  pCtbIdxKey = pCtbCur->pVal;
+  pCtbIdxKey = pCtbCur->pKey;
 
   return pCtbIdxKey->uid;
 }

From 6863999ec5f5974dcd807d6eaa51e6ee48dd26a7 Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Wed, 6 Apr 2022 17:59:08 +0800
Subject: [PATCH 20/25] [td-14493] support having in group by

---
 source/dnode/vnode/inc/tsdb.h            |   4 +-
 source/dnode/vnode/src/tsdb/tsdbRead.c   |   4 +-
 source/libs/executor/inc/executorimpl.h  |  12 +-
 source/libs/executor/src/executorimpl.c  | 210 ++++++-----------------
 source/libs/executor/src/groupoperator.c | 197 +++++++++++++++++++--
 source/libs/executor/src/scanoperator.c  |  35 +---
 source/libs/executor/src/tsimplehash.c   |   6 +-
 source/libs/function/src/builtinsimpl.c  |   2 +-
 source/util/src/thash.c                  |  16 +-
 9 files changed, 262 insertions(+), 224 deletions(-)

diff --git a/source/dnode/vnode/inc/tsdb.h b/source/dnode/vnode/inc/tsdb.h
index 3cabb6ce2e..38d6d879fa 100644
--- a/source/dnode/vnode/inc/tsdb.h
+++ b/source/dnode/vnode/inc/tsdb.h
@@ -172,9 +172,9 @@ tsdbReaderT *tsdbQueryTables(STsdb *tsdb, STsdbQueryCond *pCond, STableGroupInfo
 
 tsdbReaderT tsdbQueryCacheLast(STsdb *tsdb, STsdbQueryCond *pCond, STableGroupInfo *groupList, uint64_t qId, void* pMemRef);
 
-int32_t tsdbGetFileBlocksDistInfo(tsdbReaderT* queryHandle, STableBlockDistInfo* pTableBlockInfo);
+int32_t tsdbGetFileBlocksDistInfo(tsdbReaderT* pReader, STableBlockDistInfo* pTableBlockInfo);
 
-bool isTsdbCacheLastRow(tsdbReaderT* pTsdbReadHandle);
+bool isTsdbCacheLastRow(tsdbReaderT* pReader);
 
 /**
  *
diff --git a/source/dnode/vnode/src/tsdb/tsdbRead.c b/source/dnode/vnode/src/tsdb/tsdbRead.c
index 4a3e764b13..144aa7fcdc 100644
--- a/source/dnode/vnode/src/tsdb/tsdbRead.c
+++ b/source/dnode/vnode/src/tsdb/tsdbRead.c
@@ -3046,8 +3046,8 @@ bool tsdbGetExternalRow(tsdbReaderT pHandle) {
 //  return code;
 //}
 
-bool isTsdbCacheLastRow(tsdbReaderT* pTsdbReadHandle) {
-  return ((STsdbReadHandle *)pTsdbReadHandle)->cachelastrow > TSDB_CACHED_TYPE_NONE;
+bool isTsdbCacheLastRow(tsdbReaderT* pReader) {
+  return ((STsdbReadHandle *)pReader)->cachelastrow > TSDB_CACHED_TYPE_NONE;
 }
 
 int32_t checkForCachedLastRow(STsdbReadHandle* pTsdbReadHandle, STableGroupInfo *groupList) {
diff --git a/source/libs/executor/inc/executorimpl.h b/source/libs/executor/inc/executorimpl.h
index 3cbcbc004b..bf9377c83d 100644
--- a/source/libs/executor/inc/executorimpl.h
+++ b/source/libs/executor/inc/executorimpl.h
@@ -551,6 +551,7 @@ typedef struct SGroupbyOperatorInfo {
   SOptrBasicInfo binfo;
   SArray*        pGroupCols;
   SArray*        pGroupColVals; // current group column values, SArray<SGroupKeys>
+  SNode*         pCondition;
   bool           isInit;       // denote if current val is initialized or not
   char*          keyBuf;       // group by keys for hash
   int32_t        groupKeyLen;  // total group by column width
@@ -630,9 +631,10 @@ typedef struct SDistinctOperatorInfo {
   SHashObj*    pSet;
   SSDataBlock* pRes;
   bool         recordNullVal;  // has already record the null value, no need to try again
-  int64_t      threshold;  // todo remove it
-  int64_t      outputCapacity;// todo remove it
-  int32_t      totalBytes; // todo remove it
+//  int64_t      threshold;  // todo remove it
+//  int64_t      outputCapacity;// todo remove it
+//  int32_t      totalBytes; // todo remove it
+  SResultInfo  resInfo;
   char*        buf;
   SArray*      pDistinctDataInfo;
 } SDistinctOperatorInfo;
@@ -651,6 +653,8 @@ void doDestroyBasicInfo(SOptrBasicInfo* pInfo, int32_t numOfOutput);
 int32_t setSDataBlockFromFetchRsp(SSDataBlock* pRes, SLoadRemoteDataInfo* pLoadInfo, int32_t numOfRows,
                                          char* pData, int32_t compLen, int32_t numOfOutput, int64_t startTs,
                                          uint64_t* total, SArray* pColList);
+void doSetOperatorCompleted(SOperatorInfo* pOperator);
+void doFilter(const SNode* pFilterNode, SSDataBlock* pBlock);
 
 SOperatorInfo* createExchangeOperatorInfo(const SNodeList* pSources, SSDataBlock* pBlock, SExecTaskInfo* pTaskInfo);
 SOperatorInfo* createTableScanOperatorInfo(void* pTsdbReadHandle, int32_t order, int32_t numOfCols, int32_t repeatTime,
@@ -667,7 +671,7 @@ SOperatorInfo* createIntervalOperatorInfo(SOperatorInfo* downstream, SExprInfo*
                                           const STableGroupInfo* pTableGroupInfo, SExecTaskInfo* pTaskInfo);
 SOperatorInfo* createSessionAggOperatorInfo(SOperatorInfo* downstream, SExprInfo* pExprInfo, int32_t numOfCols, SSDataBlock* pResBlock, int64_t gap, SExecTaskInfo* pTaskInfo);
 SOperatorInfo* createGroupOperatorInfo(SOperatorInfo* downstream, SExprInfo* pExprInfo, int32_t numOfCols, SSDataBlock* pResultBlock,
-                                       SArray* pGroupColList, SExecTaskInfo* pTaskInfo, const STableGroupInfo* pTableGroupInfo);
+                                       SArray* pGroupColList, SNode* pCondition, SExecTaskInfo* pTaskInfo, const STableGroupInfo* pTableGroupInfo);
 SOperatorInfo* createDataBlockInfoScanOperator(void* dataReader, SExecTaskInfo* pTaskInfo);
 SOperatorInfo* createStreamScanOperatorInfo(void* streamReadHandle, SSDataBlock* pResBlock, SArray* pColList, SArray* pTableIdList, SExecTaskInfo* pTaskInfo);
 
diff --git a/source/libs/executor/src/executorimpl.c b/source/libs/executor/src/executorimpl.c
index 3e47356223..2323291fc1 100644
--- a/source/libs/executor/src/executorimpl.c
+++ b/source/libs/executor/src/executorimpl.c
@@ -51,8 +51,6 @@
 
 #define GET_FORWARD_DIRECTION_FACTOR(ord) (((ord) == TSDB_ORDER_ASC) ? QUERY_ASC_FORWARD_STEP : QUERY_DESC_FORWARD_STEP)
 
-#define MULTI_KEY_DELIM "-"
-
 enum {
   TS_JOIN_TS_EQUAL = 0,
   TS_JOIN_TS_NOT_EQUALS = 1,
@@ -221,12 +219,13 @@ static void destroyConditionOperatorInfo(void* param, int32_t numOfOutput);
 static void destroyOperatorInfo(SOperatorInfo* pOperator);
 static void destroySysTableScannerOperatorInfo(void* param, int32_t numOfOutput);
 
-static void doSetOperatorCompleted(SOperatorInfo* pOperator) {
+void doSetOperatorCompleted(SOperatorInfo* pOperator) {
   pOperator->status = OP_EXEC_DONE;
   if (pOperator->pTaskInfo != NULL) {
     setTaskStatus(pOperator->pTaskInfo, TASK_COMPLETED);
   }
 }
+
 #define OPTR_IS_OPENED(_optr)  (((_optr)->status & OP_OPENED) == OP_OPENED)
 #define OPTR_SET_OPENED(_optr) ((_optr)->status |= OP_OPENED)
 
@@ -1245,11 +1244,7 @@ static void doAggregateImpl(SOperatorInfo* pOperator, TSKEY startTs, SqlFunction
 }
 
 static void setPseudoOutputColInfo(SSDataBlock* pResult, SqlFunctionCtx* pCtx, SArray* pPseudoList) {
-  size_t num = 0;
-  if (pPseudoList != NULL) {
-    num = taosArrayGetSize(pPseudoList);
-  }
-
+  size_t num = (pPseudoList != NULL)? taosArrayGetSize(pPseudoList):0;
   for (int32_t i = 0; i < num; ++i) {
     pCtx[i].pOutput = taosArrayGet(pResult->pDataBlock, i);
   }
@@ -1291,8 +1286,10 @@ static void projectApplyFunctions(SExprInfo* pExpr, SSDataBlock* pResult, SSData
         pCtx[k].pOutput = taosArrayGet(pResult->pDataBlock, k);
         pCtx[k].offset  = pResult->info.rows;  // set the start offset
 
-        int32_t* outputColIndex = taosArrayGet(pPseudoList, 0);
-        pCtx[k].pTsOutput = (SColumnInfoData*)pCtx[*outputColIndex].pOutput;
+        if (taosArrayGetSize(pPseudoList) > 0) {
+          int32_t* outputColIndex = taosArrayGet(pPseudoList, 0);
+          pCtx[k].pTsOutput = (SColumnInfoData*)pCtx[*outputColIndex].pOutput;
+        }
 
         int32_t numOfRows = pCtx[k].fpSet.process(&pCtx[k]);
         pResult->info.rows += numOfRows;
@@ -3272,6 +3269,44 @@ void setResultRowOutputBufInitCtx_rv(SResultRow* pResult, SqlFunctionCtx* pCtx,
   }
 }
 
+void doFilter(const SNode* pFilterNode, SSDataBlock* pBlock) {
+  if (pFilterNode == NULL) {
+    return;
+  }
+
+  SFilterInfo* filter = NULL;
+  int32_t      code = filterInitFromNode((SNode*)pFilterNode, &filter, 0);
+
+  SFilterColumnParam param1 = {.numOfCols = pBlock->info.numOfCols, .pDataBlock = pBlock->pDataBlock};
+  code = filterSetDataFromSlotId(filter, &param1);
+
+  int8_t* rowRes = NULL;
+  bool    keep = filterExecute(filter, pBlock, &rowRes, NULL, param1.numOfCols);
+
+  SSDataBlock* px = createOneDataBlock(pBlock);
+  blockDataEnsureCapacity(px, pBlock->info.rows);
+
+  int32_t numOfRow = 0;
+  for (int32_t i = 0; i < pBlock->info.numOfCols; ++i) {
+    SColumnInfoData* pDst = taosArrayGet(px->pDataBlock, i);
+    SColumnInfoData* pSrc = taosArrayGet(pBlock->pDataBlock, i);
+
+    numOfRow = 0;
+    for (int32_t j = 0; j < pBlock->info.rows; ++j) {
+      if (rowRes[j] == 0) {
+        continue;
+      }
+
+      colDataAppend(pDst, numOfRow, colDataGetData(pSrc, j), false);
+      numOfRow += 1;
+    }
+
+    *pSrc = *pDst;
+  }
+
+  pBlock->info.rows = numOfRow;
+}
+
 void doSetTableGroupOutputBuf(SAggOperatorInfo* pAggInfo, int32_t numOfOutput, int32_t tableGroupId, SExecTaskInfo* pTaskInfo) {
   // for simple group by query without interval, all the tables belong to one group result.
   int64_t uid = 0;
@@ -6206,14 +6241,6 @@ static void destroyOrderOperatorInfo(void* param, int32_t numOfOutput) {
   taosArrayDestroy(pInfo->pSortInfo);
 }
 
-static void destroyDistinctOperatorInfo(void* param, int32_t numOfOutput) {
-  SDistinctOperatorInfo* pInfo = (SDistinctOperatorInfo*)param;
-  taosHashCleanup(pInfo->pSet);
-  taosMemoryFreeClear(pInfo->buf);
-  taosArrayDestroy(pInfo->pDistinctDataInfo);
-  pInfo->pRes = blockDataDestroy(pInfo->pRes);
-}
-
 void destroyExchangeOperatorInfo(void* param, int32_t numOfOutput) {
   SExchangeInfo* pExInfo = (SExchangeInfo*)param;
   taosArrayDestroy(pExInfo->pSources);
@@ -6785,146 +6812,7 @@ SOperatorInfo* createTagScanOperatorInfo(STaskRuntimeEnv* pRuntimeEnv, SExprInfo
 
   return pOperator;
 }
-static bool initMultiDistinctInfo(SDistinctOperatorInfo* pInfo, SOperatorInfo* pOperator, SSDataBlock* pBlock) {
-  if (taosArrayGetSize(pInfo->pDistinctDataInfo) == pOperator->numOfOutput) {
-    // distinct info already inited
-    return true;
-  }
-  for (int i = 0; i < pOperator->numOfOutput; i++) {
-    //    pInfo->totalBytes += pOperator->pExpr[i].base.colBytes;
-  }
-  for (int i = 0; i < pOperator->numOfOutput; i++) {
-    int numOfBlock = (int)(taosArrayGetSize(pBlock->pDataBlock));
-    assert(i < numOfBlock);
-    for (int j = 0; j < numOfBlock; j++) {
-      SColumnInfoData* pColDataInfo = taosArrayGet(pBlock->pDataBlock, j);
-      if (pColDataInfo->info.colId == pOperator->pExpr[i].base.resSchema.colId) {
-        SDistinctDataInfo item = {.index = j, .type = pColDataInfo->info.type, .bytes = pColDataInfo->info.bytes};
-        taosArrayInsert(pInfo->pDistinctDataInfo, i, &item);
-      }
-    }
-  }
-  pInfo->totalBytes += (int32_t)strlen(MULTI_KEY_DELIM) * (pOperator->numOfOutput);
-  pInfo->buf = taosMemoryCalloc(1, pInfo->totalBytes);
-  return taosArrayGetSize(pInfo->pDistinctDataInfo) == pOperator->numOfOutput ? true : false;
-}
 
-static void buildMultiDistinctKey(SDistinctOperatorInfo* pInfo, SSDataBlock* pBlock, int32_t rowId) {
-  char* p = pInfo->buf;
-  memset(p, 0, pInfo->totalBytes);
-
-  for (int i = 0; i < taosArrayGetSize(pInfo->pDistinctDataInfo); i++) {
-    SDistinctDataInfo* pDistDataInfo = (SDistinctDataInfo*)taosArrayGet(pInfo->pDistinctDataInfo, i);
-    SColumnInfoData*   pColDataInfo = taosArrayGet(pBlock->pDataBlock, pDistDataInfo->index);
-    char*              val = ((char*)pColDataInfo->pData) + pColDataInfo->info.bytes * rowId;
-    if (isNull(val, pDistDataInfo->type)) {
-      p += pDistDataInfo->bytes;
-      continue;
-    }
-    if (IS_VAR_DATA_TYPE(pDistDataInfo->type)) {
-      memcpy(p, varDataVal(val), varDataLen(val));
-      p += varDataLen(val);
-    } else {
-      memcpy(p, val, pDistDataInfo->bytes);
-      p += pDistDataInfo->bytes;
-    }
-    memcpy(p, MULTI_KEY_DELIM, strlen(MULTI_KEY_DELIM));
-    p += strlen(MULTI_KEY_DELIM);
-  }
-}
-
-static SSDataBlock* hashDistinct(SOperatorInfo* pOperator, bool* newgroup) {
-  if (pOperator->status == OP_EXEC_DONE) {
-    return NULL;
-  }
-
-  SDistinctOperatorInfo* pInfo = pOperator->info;
-  SSDataBlock*           pRes = pInfo->pRes;
-
-  pRes->info.rows = 0;
-  SSDataBlock* pBlock = NULL;
-
-  SOperatorInfo* pDownstream = pOperator->pDownstream[0];
-  while (1) {
-    publishOperatorProfEvent(pDownstream, QUERY_PROF_BEFORE_OPERATOR_EXEC);
-    pBlock = pDownstream->getNextFn(pDownstream, newgroup);
-    publishOperatorProfEvent(pDownstream, QUERY_PROF_AFTER_OPERATOR_EXEC);
-
-    if (pBlock == NULL) {
-      doSetOperatorCompleted(pOperator);
-      break;
-    }
-    if (!initMultiDistinctInfo(pInfo, pOperator, pBlock)) {
-      doSetOperatorCompleted(pOperator);
-      break;
-    }
-
-    // ensure result output buf
-    if (pRes->info.rows + pBlock->info.rows > pInfo->outputCapacity) {
-      int32_t newSize = pRes->info.rows + pBlock->info.rows;
-      for (int i = 0; i < taosArrayGetSize(pRes->pDataBlock); i++) {
-        SColumnInfoData*   pResultColInfoData = taosArrayGet(pRes->pDataBlock, i);
-        SDistinctDataInfo* pDistDataInfo = taosArrayGet(pInfo->pDistinctDataInfo, i);
-        char*              tmp = taosMemoryRealloc(pResultColInfoData->pData, newSize * pDistDataInfo->bytes);
-        if (tmp == NULL) {
-          return NULL;
-        } else {
-          pResultColInfoData->pData = tmp;
-        }
-      }
-      pInfo->outputCapacity = newSize;
-    }
-
-    for (int32_t i = 0; i < pBlock->info.rows; i++) {
-      buildMultiDistinctKey(pInfo, pBlock, i);
-      if (taosHashGet(pInfo->pSet, pInfo->buf, pInfo->totalBytes) == NULL) {
-        int32_t dummy;
-        taosHashPut(pInfo->pSet, pInfo->buf, pInfo->totalBytes, &dummy, sizeof(dummy));
-        for (int j = 0; j < taosArrayGetSize(pRes->pDataBlock); j++) {
-          SDistinctDataInfo* pDistDataInfo = taosArrayGet(pInfo->pDistinctDataInfo, j);  // distinct meta info
-          SColumnInfoData*   pColInfoData = taosArrayGet(pBlock->pDataBlock, pDistDataInfo->index);  // src
-          SColumnInfoData*   pResultColInfoData = taosArrayGet(pRes->pDataBlock, j);                 // dist
-
-          char* val = ((char*)pColInfoData->pData) + pDistDataInfo->bytes * i;
-          char* start = pResultColInfoData->pData + pDistDataInfo->bytes * pInfo->pRes->info.rows;
-          memcpy(start, val, pDistDataInfo->bytes);
-        }
-        pRes->info.rows += 1;
-      }
-    }
-
-    if (pRes->info.rows >= pInfo->threshold) {
-      break;
-    }
-  }
-  return (pInfo->pRes->info.rows > 0) ? pInfo->pRes : NULL;
-}
-
-SOperatorInfo* createDistinctOperatorInfo(SOperatorInfo* downstream, SExprInfo* pExpr, int32_t numOfCols, SSDataBlock* pResBlock, SExecTaskInfo* pTaskInfo) {
-  SDistinctOperatorInfo* pInfo = taosMemoryCalloc(1, sizeof(SDistinctOperatorInfo));
-  SOperatorInfo* pOperator = taosMemoryCalloc(1, sizeof(SOperatorInfo));
-
-  pOperator->resultInfo.capacity = 4096;  // todo extract function.
-
-  pInfo->totalBytes        = 0;
-  pInfo->buf               = NULL;
-
-  pInfo->pDistinctDataInfo = taosArrayInit(numOfCols, sizeof(SDistinctDataInfo));
-  pInfo->pSet = taosHashInit(64, taosGetDefaultHashFunction(TSDB_DATA_TYPE_BINARY), false, HASH_NO_LOCK);
-
-  pOperator->name         = "DistinctOperator";
-  pOperator->blockingOptr = true;
-  pOperator->status       = OP_NOT_OPENED;
-//  pOperator->operatorType = DISTINCT;
-  pOperator->pExpr        = pExpr;
-  pOperator->numOfOutput  = numOfCols;
-  pOperator->info         = pInfo;
-  pOperator->getNextFn    = hashDistinct;
-  pOperator->closeFn      = destroyDistinctOperatorInfo;
-
-  int32_t code = appendDownstream(pOperator, &downstream, 1);
-  return pOperator;
-}
 
 static int32_t getColumnIndexInSource(SQueriedTableInfo* pTableInfo, SExprBasicInfo* pExpr, SColumnInfo* pTagCols) {
   int32_t j = 0;
@@ -7147,7 +7035,6 @@ SOperatorInfo* createOperatorTree(SPhysiNode* pPhyNode, SExecTaskInfo* pTaskInfo
       int32_t     numOfCols = 0;
       tsdbReaderT pDataReader = doCreateDataReader((STableScanPhysiNode*)pPhyNode, pHandle, pTableGroupInfo, (uint64_t)queryId, taskId);
       SArray* pColList = extractColMatchInfo(pScanPhyNode->pScanCols, pScanPhyNode->node.pOutputDataBlockDesc, &numOfCols);
-
       return createTableScanOperatorInfo(pDataReader, pScanPhyNode->order, numOfCols, pScanPhyNode->count,
                                          pScanPhyNode->reverse, pColList, pScanPhyNode->node.pConditions, pTaskInfo);
     } else if (QUERY_NODE_PHYSICAL_PLAN_EXCHANGE == nodeType(pPhyNode)) {
@@ -7157,8 +7044,7 @@ SOperatorInfo* createOperatorTree(SPhysiNode* pPhyNode, SExecTaskInfo* pTaskInfo
     } else if (QUERY_NODE_PHYSICAL_PLAN_STREAM_SCAN == nodeType(pPhyNode)) {
       SScanPhysiNode* pScanPhyNode = (SScanPhysiNode*)pPhyNode;  // simple child table.
 
-      int32_t code = doCreateTableGroup(pHandle->meta, pScanPhyNode->tableType, pScanPhyNode->uid, pTableGroupInfo,
-                                        queryId, taskId);
+      int32_t code = doCreateTableGroup(pHandle->meta, pScanPhyNode->tableType, pScanPhyNode->uid, pTableGroupInfo, queryId, taskId);
       SArray* tableIdList = extractTableIdList(pTableGroupInfo);
 
       SSDataBlock* pResBlock = createOutputBuf_rv1(pScanPhyNode->node.pOutputDataBlockDesc);
@@ -7215,7 +7101,7 @@ SOperatorInfo* createOperatorTree(SPhysiNode* pPhyNode, SExecTaskInfo* pTaskInfo
 
       if (pAggNode->pGroupKeys != NULL) {
         SArray* pColList = extractColumnInfo(pAggNode->pGroupKeys);
-        return createGroupOperatorInfo(op, pExprInfo, num, pResBlock, pColList, pTaskInfo, NULL);
+        return createGroupOperatorInfo(op, pExprInfo, num, pResBlock, pColList, pAggNode->node.pConditions, pTaskInfo, NULL);
       } else {
         return createAggregateOperatorInfo(op, pExprInfo, num, pResBlock, pTaskInfo, pTableGroupInfo);
       }
@@ -7306,7 +7192,7 @@ static tsdbReaderT createDataReaderImpl(STableScanPhysiNode* pTableScanNode, STa
       continue;
     }
 
-    cond.colList[j].type = pColNode->node.resType.type;
+    cond.colList[j].type  = pColNode->node.resType.type;
     cond.colList[j].bytes = pColNode->node.resType.bytes;
     cond.colList[j].colId = pColNode->colId;
     j += 1;
diff --git a/source/libs/executor/src/groupoperator.c b/source/libs/executor/src/groupoperator.c
index 90cb97f6c7..c22b58a0d8 100644
--- a/source/libs/executor/src/groupoperator.c
+++ b/source/libs/executor/src/groupoperator.c
@@ -245,13 +245,14 @@ static SSDataBlock* hashGroupbyAggregate(SOperatorInfo* pOperator, bool* newgrou
   }
 
   SGroupbyOperatorInfo* pInfo = pOperator->info;
+  SSDataBlock* pRes = pInfo->binfo.pRes;
+
   if (pOperator->status == OP_RES_TO_RETURN) {
-    toSDatablock(&pInfo->groupResInfo, pInfo->aggSup.pResultBuf, pInfo->binfo.pRes, pInfo->binfo.capacity,
-                 pInfo->binfo.rowCellInfoOffset);
-    if (pInfo->binfo.pRes->info.rows == 0 || !hasRemainDataInCurrentGroup(&pInfo->groupResInfo)) {
+    toSDatablock(&pInfo->groupResInfo, pInfo->aggSup.pResultBuf, pRes, pInfo->binfo.capacity, pInfo->binfo.rowCellInfoOffset);
+    if (pRes->info.rows == 0 || !hasRemainDataInCurrentGroup(&pInfo->groupResInfo)) {
       pOperator->status = OP_EXEC_DONE;
     }
-    return pInfo->binfo.pRes;
+    return pRes;
   }
 
   int32_t        order = TSDB_ORDER_ASC;
@@ -283,18 +284,29 @@ static SSDataBlock* hashGroupbyAggregate(SOperatorInfo* pOperator, bool* newgrou
   //    pInfo->binfo.rowCellInfoOffset);
   //  }
 
-  blockDataEnsureCapacity(pInfo->binfo.pRes, pInfo->binfo.capacity);
+  blockDataEnsureCapacity(pRes, pInfo->binfo.capacity);
   initGroupResInfo(&pInfo->groupResInfo, &pInfo->binfo.resultRowInfo);
-  toSDatablock(&pInfo->groupResInfo, pInfo->aggSup.pResultBuf, pInfo->binfo.pRes, pInfo->binfo.capacity,
-               pInfo->binfo.rowCellInfoOffset);
-  if (pInfo->binfo.pRes->info.rows == 0 || !hasRemainDataInCurrentGroup(&pInfo->groupResInfo)) {
-    pOperator->status = OP_EXEC_DONE;
+
+  while(1) {
+    toSDatablock(&pInfo->groupResInfo, pInfo->aggSup.pResultBuf, pRes, pInfo->binfo.capacity,
+                 pInfo->binfo.rowCellInfoOffset);
+    doFilter(pInfo->pCondition, pRes);
+
+    bool hasRemain = hasRemainDataInCurrentGroup(&pInfo->groupResInfo);
+    if (!hasRemain) {
+      pOperator->status = OP_EXEC_DONE;
+      break;
+    }
+
+    if (pRes->info.rows > 0) {
+      break;
+    }
   }
 
   return pInfo->binfo.pRes;
 }
 
-SOperatorInfo* createGroupOperatorInfo(SOperatorInfo* downstream, SExprInfo* pExprInfo, int32_t numOfCols, SSDataBlock* pResultBlock, SArray* pGroupColList, SExecTaskInfo* pTaskInfo,
+SOperatorInfo* createGroupOperatorInfo(SOperatorInfo* downstream, SExprInfo* pExprInfo, int32_t numOfCols, SSDataBlock* pResultBlock, SArray* pGroupColList, SNode* pCondition, SExecTaskInfo* pTaskInfo,
                                        const STableGroupInfo* pTableGroupInfo) {
   SGroupbyOperatorInfo* pInfo = taosMemoryCalloc(1, sizeof(SGroupbyOperatorInfo));
   SOperatorInfo*        pOperator = taosMemoryCalloc(1, sizeof(SOperatorInfo));
@@ -303,6 +315,7 @@ SOperatorInfo* createGroupOperatorInfo(SOperatorInfo* downstream, SExprInfo* pEx
   }
 
   pInfo->pGroupCols = pGroupColList;
+  pInfo->pCondition = pCondition;
   initAggInfo(&pInfo->binfo, &pInfo->aggSup, pExprInfo, numOfCols, 4096, pResultBlock, pTaskInfo->id.str);
   initResultRowInfo(&pInfo->binfo.resultRowInfo, 8);
 
@@ -329,4 +342,166 @@ SOperatorInfo* createGroupOperatorInfo(SOperatorInfo* downstream, SExprInfo* pEx
   taosMemoryFreeClear(pInfo);
   taosMemoryFreeClear(pOperator);
   return NULL;
-}
\ No newline at end of file
+}
+
+#define MULTI_KEY_DELIM "-"
+
+static void destroyDistinctOperatorInfo(void* param, int32_t numOfOutput) {
+  SDistinctOperatorInfo* pInfo = (SDistinctOperatorInfo*)param;
+  taosHashCleanup(pInfo->pSet);
+  taosMemoryFreeClear(pInfo->buf);
+  taosArrayDestroy(pInfo->pDistinctDataInfo);
+  pInfo->pRes = blockDataDestroy(pInfo->pRes);
+}
+
+static void buildMultiDistinctKey(SDistinctOperatorInfo* pInfo, SSDataBlock* pBlock, int32_t rowId) {
+  char* p = pInfo->buf;
+//  memset(p, 0, pInfo->totalBytes);
+
+  for (int i = 0; i < taosArrayGetSize(pInfo->pDistinctDataInfo); i++) {
+    SDistinctDataInfo* pDistDataInfo = (SDistinctDataInfo*)taosArrayGet(pInfo->pDistinctDataInfo, i);
+    SColumnInfoData*   pColDataInfo = taosArrayGet(pBlock->pDataBlock, pDistDataInfo->index);
+
+    char* val = ((char*)pColDataInfo->pData) + pColDataInfo->info.bytes * rowId;
+    if (isNull(val, pDistDataInfo->type)) {
+      p += pDistDataInfo->bytes;
+      continue;
+    }
+    if (IS_VAR_DATA_TYPE(pDistDataInfo->type)) {
+      memcpy(p, varDataVal(val), varDataLen(val));
+      p += varDataLen(val);
+    } else {
+      memcpy(p, val, pDistDataInfo->bytes);
+      p += pDistDataInfo->bytes;
+    }
+    memcpy(p, MULTI_KEY_DELIM, strlen(MULTI_KEY_DELIM));
+    p += strlen(MULTI_KEY_DELIM);
+  }
+}
+
+static bool initMultiDistinctInfo(SDistinctOperatorInfo* pInfo, SOperatorInfo* pOperator) {
+  for (int i = 0; i < pOperator->numOfOutput; i++) {
+    //    pInfo->totalBytes += pOperator->pExpr[i].base.colBytes;
+  }
+#if 0
+  for (int i = 0; i < pOperator->numOfOutput; i++) {
+    int numOfCols = (int)(taosArrayGetSize(pBlock->pDataBlock));
+    assert(i < numOfCols);
+    
+    for (int j = 0; j < numOfCols; j++) {
+      SColumnInfoData* pColDataInfo = taosArrayGet(pBlock->pDataBlock, j);
+      if (pColDataInfo->info.colId == pOperator->pExpr[i].base.resSchema.colId) {
+        SDistinctDataInfo item = {.index = j, .type = pColDataInfo->info.type, .bytes = pColDataInfo->info.bytes};
+        taosArrayInsert(pInfo->pDistinctDataInfo, i, &item);
+      }
+    }
+  }
+#endif
+
+//  pInfo->totalBytes += (int32_t)strlen(MULTI_KEY_DELIM) * (pOperator->numOfOutput);
+//  pInfo->buf = taosMemoryCalloc(1, pInfo->totalBytes);
+  return taosArrayGetSize(pInfo->pDistinctDataInfo) == pOperator->numOfOutput ? true : false;
+}
+
+static SSDataBlock* hashDistinct(SOperatorInfo* pOperator, bool* newgroup) {
+  if (pOperator->status == OP_EXEC_DONE) {
+    return NULL;
+  }
+
+  SDistinctOperatorInfo* pInfo = pOperator->info;
+  SSDataBlock*           pRes = pInfo->pRes;
+
+  pRes->info.rows = 0;
+  SSDataBlock* pBlock = NULL;
+
+  SOperatorInfo* pDownstream = pOperator->pDownstream[0];
+  while (1) {
+    publishOperatorProfEvent(pDownstream, QUERY_PROF_BEFORE_OPERATOR_EXEC);
+    pBlock = pDownstream->getNextFn(pDownstream, newgroup);
+    publishOperatorProfEvent(pDownstream, QUERY_PROF_AFTER_OPERATOR_EXEC);
+
+    if (pBlock == NULL) {
+      doSetOperatorCompleted(pOperator);
+      break;
+    }
+
+    // ensure result output buf
+    if (pRes->info.rows + pBlock->info.rows > pInfo->resInfo.capacity) {
+      int32_t newSize = pRes->info.rows + pBlock->info.rows;
+      for (int i = 0; i < taosArrayGetSize(pRes->pDataBlock); i++) {
+        SColumnInfoData*   pResultColInfoData = taosArrayGet(pRes->pDataBlock, i);
+        SDistinctDataInfo* pDistDataInfo = taosArrayGet(pInfo->pDistinctDataInfo, i);
+
+//        char* tmp = taosMemoryRealloc(pResultColInfoData->pData, newSize * pDistDataInfo->bytes);
+//        if (tmp == NULL) {
+//          return NULL;
+//        } else {
+//          pResultColInfoData->pData = tmp;
+//        }
+      }
+      pInfo->resInfo.capacity = newSize;
+    }
+
+    for (int32_t i = 0; i < pBlock->info.rows; i++) {
+      buildMultiDistinctKey(pInfo, pBlock, i);
+      if (taosHashGet(pInfo->pSet, pInfo->buf, 0) == NULL) {
+        taosHashPut(pInfo->pSet, pInfo->buf, 0, NULL, 0);
+
+        for (int j = 0; j < taosArrayGetSize(pRes->pDataBlock); j++) {
+          SDistinctDataInfo* pDistDataInfo = taosArrayGet(pInfo->pDistinctDataInfo, j);  // distinct meta info
+          SColumnInfoData*   pColInfoData = taosArrayGet(pBlock->pDataBlock, pDistDataInfo->index);  // src
+          SColumnInfoData*   pResultColInfoData = taosArrayGet(pRes->pDataBlock, j);                 // dist
+
+          char* val = ((char*)pColInfoData->pData) + pDistDataInfo->bytes * i;
+          char* start = pResultColInfoData->pData + pDistDataInfo->bytes * pInfo->pRes->info.rows;
+          memcpy(start, val, pDistDataInfo->bytes);
+        }
+
+        pRes->info.rows += 1;
+      }
+    }
+
+    if (pRes->info.rows >= pInfo->resInfo.threshold) {
+      break;
+    }
+  }
+
+  return (pInfo->pRes->info.rows > 0) ? pInfo->pRes : NULL;
+}
+
+SOperatorInfo* createDistinctOperatorInfo(SOperatorInfo* downstream, SExprInfo* pExpr, int32_t numOfCols, SSDataBlock* pResBlock, SExecTaskInfo* pTaskInfo) {
+  SDistinctOperatorInfo* pInfo = taosMemoryCalloc(1, sizeof(SDistinctOperatorInfo));
+  SOperatorInfo* pOperator = taosMemoryCalloc(1, sizeof(SOperatorInfo));
+  if (pInfo == NULL || pOperator == NULL) {
+    goto _error;
+  }
+
+  pOperator->resultInfo.capacity = 4096;  // todo extract function.
+
+//  pInfo->totalBytes        = 0;
+  pInfo->buf               = NULL;
+
+  pInfo->pDistinctDataInfo = taosArrayInit(numOfCols, sizeof(SDistinctDataInfo));
+  initMultiDistinctInfo(pInfo, pOperator);
+  
+  pInfo->pSet = taosHashInit(64, taosGetDefaultHashFunction(TSDB_DATA_TYPE_BINARY), false, HASH_NO_LOCK);
+
+  pOperator->name         = "DistinctOperator";
+  pOperator->blockingOptr = true;
+  pOperator->status       = OP_NOT_OPENED;
+//  pOperator->operatorType = DISTINCT;
+  pOperator->pExpr        = pExpr;
+  pOperator->numOfOutput  = numOfCols;
+  pOperator->info         = pInfo;
+  pOperator->getNextFn    = hashDistinct;
+  pOperator->closeFn      = destroyDistinctOperatorInfo;
+
+  int32_t code = appendDownstream(pOperator, &downstream, 1);
+  return pOperator;
+
+  _error:
+  pTaskInfo->code = TSDB_CODE_OUT_OF_MEMORY;
+  taosMemoryFree(pInfo);
+  taosMemoryFree(pOperator);
+  return NULL;
+}
diff --git a/source/libs/executor/src/scanoperator.c b/source/libs/executor/src/scanoperator.c
index ace84c5f9a..f39a18da68 100644
--- a/source/libs/executor/src/scanoperator.c
+++ b/source/libs/executor/src/scanoperator.c
@@ -34,6 +34,7 @@
 #define SET_REVERSE_SCAN_FLAG(_info) ((_info)->scanFlag = REVERSE_SCAN)
 #define SWITCH_ORDER(n)           (((n) = ((n) == TSDB_ORDER_ASC) ? TSDB_ORDER_DESC : TSDB_ORDER_ASC))
 
+
 void switchCtxOrder(SqlFunctionCtx* pCtx, int32_t numOfOutput) {
   for (int32_t i = 0; i < numOfOutput; ++i) {
     SWITCH_ORDER(pCtx[i].order);
@@ -91,39 +92,7 @@ int32_t loadDataBlock(SExecTaskInfo* pTaskInfo, STableScanInfo* pTableScanInfo,
     taosArraySet(pBlock->pDataBlock, pColMatchInfo->targetSlotId, p);
   }
 
-  if (pTableScanInfo->pFilterNode != NULL) {
-    SFilterInfo* filter = NULL;
-    int32_t      code = filterInitFromNode((SNode*)pTableScanInfo->pFilterNode, &filter, 0);
-
-    SFilterColumnParam param1 = {.numOfCols = pBlock->info.numOfCols, .pDataBlock = pBlock->pDataBlock};
-    code = filterSetDataFromSlotId(filter, &param1);
-
-    int8_t* rowRes = NULL;
-    bool keep = filterExecute(filter, pBlock, &rowRes, NULL, param1.numOfCols);
-
-    SSDataBlock* px = createOneDataBlock(pBlock);
-    blockDataEnsureCapacity(px, pBlock->info.rows);
-
-    int32_t numOfRow = 0;
-    for (int32_t i = 0; i < pBlock->info.numOfCols; ++i) {
-      SColumnInfoData* pDst = taosArrayGet(px->pDataBlock, i);
-      SColumnInfoData* pSrc = taosArrayGet(pBlock->pDataBlock, i);
-
-      numOfRow = 0;
-      for (int32_t j = 0; j < pBlock->info.rows; ++j) {
-        if (rowRes[j] == 0) {
-          continue;
-        }
-
-        colDataAppend(pDst, numOfRow, colDataGetData(pSrc, j), false);
-        numOfRow += 1;
-      }
-      *pSrc = *pDst;
-    }
-
-    pBlock->info.rows = numOfRow;
-  }
-
+  doFilter(pTableScanInfo->pFilterNode, pBlock);
   return TSDB_CODE_SUCCESS;
 }
 
diff --git a/source/libs/executor/src/tsimplehash.c b/source/libs/executor/src/tsimplehash.c
index 981da0415e..5c10216a96 100644
--- a/source/libs/executor/src/tsimplehash.c
+++ b/source/libs/executor/src/tsimplehash.c
@@ -37,7 +37,7 @@ typedef struct SHNode {
   char             data[];
 } SHNode;
 
-typedef struct SSHashObj {
+struct SSHashObj {
   SHNode         **hashList;
   size_t           capacity;     // number of slots
   int64_t          size;         // number of elements in hash table
@@ -45,7 +45,7 @@ typedef struct SSHashObj {
   _equal_fn_t      equalFp;      // equal function
   int32_t          keyLen;
   int32_t          dataLen;
-} SSHashObj;
+};
 
 static FORCE_INLINE int32_t taosHashCapacity(int32_t length) {
   int32_t len = MIN(length, HASH_MAX_CAPACITY);
@@ -107,7 +107,7 @@ static SHNode *doCreateHashNode(const void *key, size_t keyLen, const void *pDat
   return pNewNode;
 }
 
-void taosHashTableResize(SSHashObj *pHashObj) {
+static void taosHashTableResize(SSHashObj *pHashObj) {
   if (!HASH_NEED_RESIZE(pHashObj)) {
     return;
   }
diff --git a/source/libs/function/src/builtinsimpl.c b/source/libs/function/src/builtinsimpl.c
index 6f7fcd37be..fb6abe3c5d 100644
--- a/source/libs/function/src/builtinsimpl.c
+++ b/source/libs/function/src/builtinsimpl.c
@@ -917,7 +917,7 @@ int32_t diffFunction(SqlFunctionCtx *pCtx) {
             colDataAppendInt32(pOutput, pos, &delta);
           }
 
-          if (tsList != NULL) {
+          if (pTsOutput != NULL) {
             colDataAppendInt64(pTsOutput, pos, &tsList[i]);
           }
         }
diff --git a/source/util/src/thash.c b/source/util/src/thash.c
index 88e5c12770..809d008aa7 100644
--- a/source/util/src/thash.c
+++ b/source/util/src/thash.c
@@ -311,10 +311,6 @@ int32_t taosHashPut(SHashObj *pHashObj, const void *key, size_t keyLen, const vo
   }
 
   uint32_t   hashVal = (*pHashObj->hashFp)(key, (uint32_t)keyLen);
-  SHashNode *pNewNode = doCreateHashNode(key, keyLen, data, size, hashVal);
-  if (pNewNode == NULL) {
-    return -1;
-  }
 
   // need the resize process, write lock applied
   if (HASH_NEED_RESIZE(pHashObj)) {
@@ -355,6 +351,11 @@ int32_t taosHashPut(SHashObj *pHashObj, const void *key, size_t keyLen, const vo
 
   if (pNode == NULL) {
     // no data in hash table with the specified key, add it into hash table
+    SHashNode *pNewNode = doCreateHashNode(key, keyLen, data, size, hashVal);
+    if (pNewNode == NULL) {
+      return -1;
+    }
+
     pushfrontNodeInEntryList(pe, pNewNode);
     assert(pe->next != NULL);
 
@@ -368,9 +369,12 @@ int32_t taosHashPut(SHashObj *pHashObj, const void *key, size_t keyLen, const vo
   } else {
     // not support the update operation, return error
     if (pHashObj->enableUpdate) {
+      SHashNode *pNewNode = doCreateHashNode(key, keyLen, data, size, hashVal);
+      if (pNewNode == NULL) {
+        return -1;
+      }
+
       doUpdateHashNode(pHashObj, pe, prev, pNode, pNewNode);
-    } else {
-      FREE_HASH_NODE(pNewNode);
     }
 
     taosHashEntryWUnlock(pHashObj, pe);

From f136211bad2d0b78e31ce032282c04c1249e10f3 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 10:21:36 +0000
Subject: [PATCH 21/25] fix TDB encode and decode bug

---
 source/libs/tdb/src/inc/tdbUtil.h    |  7 +++++--
 source/libs/tdb/test/CMakeLists.txt  |  6 +++++-
 source/libs/tdb/test/tdbUtilTest.cpp | 20 ++++++++++++++++++++
 3 files changed, 30 insertions(+), 3 deletions(-)
 create mode 100644 source/libs/tdb/test/tdbUtilTest.cpp

diff --git a/source/libs/tdb/src/inc/tdbUtil.h b/source/libs/tdb/src/inc/tdbUtil.h
index 042026d0a4..7085e71567 100644
--- a/source/libs/tdb/src/inc/tdbUtil.h
+++ b/source/libs/tdb/src/inc/tdbUtil.h
@@ -84,15 +84,18 @@ static inline int tdbPutVarInt(u8 *p, int v) {
 static inline int tdbGetVarInt(const u8 *p, int *v) {
   int n = 0;
   int tv = 0;
+  int t;
 
   for (;;) {
     if (p[n] <= 0x7f) {
-      tv = (tv << 7) | p[n];
+      t = p[n];
+      tv |= (t << (7 * n));
       n++;
       break;
     }
 
-    tv = (tv << 7) | (p[n] & 0x7f);
+    t = p[n] & 0x7f;
+    tv |= (t << (7 * n));
     n++;
   }
 
diff --git a/source/libs/tdb/test/CMakeLists.txt b/source/libs/tdb/test/CMakeLists.txt
index 2d77c1f4e9..b2c8aaf9bc 100644
--- a/source/libs/tdb/test/CMakeLists.txt
+++ b/source/libs/tdb/test/CMakeLists.txt
@@ -1,3 +1,7 @@
 # tdbTest
 add_executable(tdbTest "tdbTest.cpp")
-target_link_libraries(tdbTest tdb gtest gtest_main)
\ No newline at end of file
+target_link_libraries(tdbTest tdb gtest gtest_main)
+
+# tdbUtilTest
+add_executable(tdbUtilTest "tdbUtilTest.cpp")
+target_link_libraries(tdbUtilTest tdb gtest gtest_main)
\ No newline at end of file
diff --git a/source/libs/tdb/test/tdbUtilTest.cpp b/source/libs/tdb/test/tdbUtilTest.cpp
new file mode 100644
index 0000000000..5fd45daa3e
--- /dev/null
+++ b/source/libs/tdb/test/tdbUtilTest.cpp
@@ -0,0 +1,20 @@
+#include <gtest/gtest.h>
+
+#include "tdbInt.h"
+
+#include <string>
+
+TEST(tdb_util_test, simple_test) {
+  int vEncode = 5000;
+  int vDecode;
+  int nEncode;
+  int nDecode;
+  u8  buffer[128];
+
+  nEncode = tdbPutVarInt(buffer, vEncode);
+
+  nDecode = tdbGetVarInt(buffer, &vDecode);
+
+  GTEST_ASSERT_EQ(nEncode, nDecode);
+  GTEST_ASSERT_EQ(vEncode, vDecode);
+}
\ No newline at end of file

From 3c7927d2b4dfc046e00643a567b57427da3ed438 Mon Sep 17 00:00:00 2001
From: Xiaoyu Wang <xiaoyuwang@taosdata.com>
Date: Wed, 6 Apr 2022 06:50:38 -0400
Subject: [PATCH 22/25] normal table rollup translate, and condition rewrite

---
 include/libs/nodes/cmdnodes.h             |  1 +
 source/libs/parser/src/parAstCreater.c    |  1 +
 source/libs/parser/src/parAstParser.c     |  2 +-
 source/libs/parser/src/parCalcConst.c     | 83 ++++++++++++++++++++++-
 source/libs/parser/src/parTranslater.c    | 72 ++++++++++++++------
 source/libs/parser/test/parserAstTest.cpp | 22 +++++-
 6 files changed, 157 insertions(+), 24 deletions(-)

diff --git a/include/libs/nodes/cmdnodes.h b/include/libs/nodes/cmdnodes.h
index 069aec14b5..921d3cc07b 100644
--- a/include/libs/nodes/cmdnodes.h
+++ b/include/libs/nodes/cmdnodes.h
@@ -94,6 +94,7 @@ typedef struct SColumnDefNode {
   char colName[TSDB_COL_NAME_LEN];
   SDataType dataType;
   char comments[TSDB_STB_COMMENT_LEN];
+  bool sma;
 } SColumnDefNode;
 
 typedef struct SCreateTableStmt {
diff --git a/source/libs/parser/src/parAstCreater.c b/source/libs/parser/src/parAstCreater.c
index ebcecfe3cb..3f7c22f4c5 100644
--- a/source/libs/parser/src/parAstCreater.c
+++ b/source/libs/parser/src/parAstCreater.c
@@ -1050,6 +1050,7 @@ SNode* createColumnDefNode(SAstCreateContext* pCxt, const SToken* pColName, SDat
   if (NULL != pComment) {
     trimString(pComment->z, pComment->n, pCol->comments, sizeof(pCol->comments));
   }
+  pCol->sma = true;
   return (SNode*)pCol;
 }
 
diff --git a/source/libs/parser/src/parAstParser.c b/source/libs/parser/src/parAstParser.c
index 76af6b7ac7..8d0ddafd26 100644
--- a/source/libs/parser/src/parAstParser.c
+++ b/source/libs/parser/src/parAstParser.c
@@ -65,7 +65,7 @@ int32_t parse(SParseContext* pParseCxt, SQuery** pQuery) {
       }
       default:
         Parse(pParser, t0.type, t0, &cxt);
-        // ParseTrace(stdout, "");
+        ParseTrace(stdout, "");
         if (!cxt.valid) {
           goto abort_parse;
         }
diff --git a/source/libs/parser/src/parCalcConst.c b/source/libs/parser/src/parCalcConst.c
index 46cdcc4243..d21628956a 100644
--- a/source/libs/parser/src/parCalcConst.c
+++ b/source/libs/parser/src/parCalcConst.c
@@ -90,14 +90,91 @@ static EDealRes calcConst(SNode** pNode, void* pContext) {
   return DEAL_RES_CONTINUE;
 }
 
+static bool isCondition(const SNode* pNode) {
+  if (QUERY_NODE_OPERATOR == nodeType(pNode)) {
+    return nodesIsComparisonOp((const SOperatorNode*)pNode);
+  }
+  return (QUERY_NODE_LOGIC_CONDITION == nodeType(pNode));
+}
+
+static int32_t rewriteIsTrue(SNode* pSrc, SNode** pIsTrue) {
+  SOperatorNode* pOp = nodesMakeNode(QUERY_NODE_OPERATOR);
+  if (NULL == pOp) {
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
+  pOp->opType = OP_TYPE_IS_TRUE;
+  pOp->pLeft = pSrc;
+  *pIsTrue = (SNode*)pOp;
+  return TSDB_CODE_SUCCESS;
+}
+
+static EDealRes doRewriteCondition(SNode** pNode, void* pContext) {
+  if (QUERY_NODE_LOGIC_CONDITION == nodeType(*pNode)) {
+    SNode* pParam = NULL;
+    FOREACH(pParam, ((SLogicConditionNode*)*pNode)->pParameterList) {
+      if (!isCondition(pParam)) {
+        SNode* pIsTrue = NULL;
+        if (TSDB_CODE_SUCCESS != rewriteIsTrue(pParam, &pIsTrue)) {
+          ((SCalcConstContext*)pContext)->code = TSDB_CODE_OUT_OF_MEMORY;
+          return DEAL_RES_ERROR;
+        }
+        REPLACE_NODE(pIsTrue);
+      }
+    }
+  }
+  return DEAL_RES_CONTINUE;
+}
+
+static int32_t rewriteCondition(SCalcConstContext* pCxt, SNode** pNode) {
+  if (!isCondition(*pNode)) {
+    return rewriteIsTrue(*pNode, pNode);
+  }
+  nodesRewriteExprPostOrder(pNode, doRewriteCondition, pCxt);
+  return pCxt->code;
+}
+
+static int32_t rewriteConditionForFromTable(SCalcConstContext* pCxt, SNode* pTable) {
+  if (QUERY_NODE_JOIN_TABLE == nodeType(pTable)) {
+    SJoinTableNode* pJoin = (SJoinTableNode*)pTable;
+    pCxt->code = rewriteConditionForFromTable(pCxt, pJoin->pLeft);
+    if (TSDB_CODE_SUCCESS == pCxt->code) {
+      pCxt->code = rewriteConditionForFromTable(pCxt, pJoin->pRight);
+    }
+    if (TSDB_CODE_SUCCESS == pCxt->code) {
+      pCxt->code = rewriteCondition(pCxt, &pJoin->pOnCond);
+    }
+  }
+  return pCxt->code;
+}
+
+static int32_t calcConstFromTable(SCalcConstContext* pCxt, SSelectStmt* pSelect) {
+  nodesRewriteExprPostOrder(&pSelect->pFromTable, calcConst, pCxt);
+  if (TSDB_CODE_SUCCESS == pCxt->code) {
+    pCxt->code = rewriteConditionForFromTable(pCxt, pSelect->pFromTable);
+  }
+  return pCxt->code;
+}
+
+static int32_t calcConstCondition(SCalcConstContext* pCxt, SNode** pCond) {
+  if (NULL == *pCond) {
+    return TSDB_CODE_SUCCESS;
+  }
+
+  nodesRewriteExprPostOrder(pCond, calcConst, pCxt);
+  if (TSDB_CODE_SUCCESS == pCxt->code) {
+    pCxt->code = rewriteCondition(pCxt, pCond);
+  }
+  return pCxt->code;
+}
+
 static int32_t calcConstSelect(SSelectStmt* pSelect) {
   SCalcConstContext cxt = { .code = TSDB_CODE_SUCCESS };
   nodesRewriteExprsPostOrder(pSelect->pProjectionList, calcConst, &cxt);
   if (TSDB_CODE_SUCCESS == cxt.code) {
-    nodesRewriteExprPostOrder(&pSelect->pFromTable, calcConst, &cxt);
+    cxt.code = calcConstFromTable(&cxt, pSelect);
   }
   if (TSDB_CODE_SUCCESS == cxt.code) {
-    nodesRewriteExprPostOrder(&pSelect->pWhere, calcConst, &cxt);
+    cxt.code = calcConstCondition(&cxt, &pSelect->pWhere);
   }
   if (TSDB_CODE_SUCCESS == cxt.code) {
     nodesRewriteExprsPostOrder(pSelect->pPartitionByList, calcConst, &cxt);
@@ -109,7 +186,7 @@ static int32_t calcConstSelect(SSelectStmt* pSelect) {
     nodesRewriteExprsPostOrder(pSelect->pGroupByList, calcConst, &cxt);
   }
   if (TSDB_CODE_SUCCESS == cxt.code) {
-    nodesRewriteExprPostOrder(&pSelect->pHaving, calcConst, &cxt);
+    cxt.code = calcConstCondition(&cxt, &pSelect->pHaving);
   }
   if (TSDB_CODE_SUCCESS == cxt.code) {
     nodesRewriteExprsPostOrder(pSelect->pOrderByList, calcConst, &cxt);
diff --git a/source/libs/parser/src/parTranslater.c b/source/libs/parser/src/parTranslater.c
index bf909a0eaf..813cf3477d 100644
--- a/source/libs/parser/src/parTranslater.c
+++ b/source/libs/parser/src/parTranslater.c
@@ -1115,7 +1115,7 @@ static int32_t columnNodeToField(SNodeList* pList, SArray** pArray) {
   return TSDB_CODE_SUCCESS;
 }
 
-static const SColumnDefNode* findColDef(const SNodeList* pCols, const SColumnNode* pCol) {
+static SColumnDefNode* findColDef(SNodeList* pCols, const SColumnNode* pCol) {
   SNode* pColDef = NULL;
   FOREACH(pColDef, pCols) {
     if (0 == strcmp(pCol->colName, ((SColumnDefNode*)pColDef)->colName)) {
@@ -1125,16 +1125,20 @@ static const SColumnDefNode* findColDef(const SNodeList* pCols, const SColumnNod
   return NULL;
 }
 
-static int32_t checkCreateSuperTable(STranslateContext* pCxt, SCreateTableStmt* pStmt) {
+static int32_t checkCreateTable(STranslateContext* pCxt, SCreateTableStmt* pStmt) {
   if (NULL != pStmt->pOptions->pSma) {
     SNode* pNode = NULL;
+    FOREACH(pNode, pStmt->pCols) {
+      ((SColumnDefNode*)pNode)->sma = false;
+    }
     FOREACH(pNode, pStmt->pOptions->pSma) {
       SColumnNode* pSmaCol = (SColumnNode*)pNode;
-      const SColumnDefNode* pColDef = findColDef(pStmt->pCols, pSmaCol);
+      SColumnDefNode* pColDef = findColDef(pStmt->pCols, pSmaCol);
       if (NULL == pColDef) {
         return generateSyntaxErrMsg(&pCxt->msgBuf, TSDB_CODE_PAR_INVALID_COLUMN, pSmaCol->colName);
       }
       pSmaCol->node.resType = pColDef->dataType;
+      pColDef->sma = true;
     }
   }
   if (NULL != pStmt->pOptions->pFuncs) {
@@ -1154,7 +1158,7 @@ static int32_t getAggregationMethod(SNodeList* pFuncs) {
 }
 
 static int32_t translateCreateSuperTable(STranslateContext* pCxt, SCreateTableStmt* pStmt) {
-  int32_t code = checkCreateSuperTable(pCxt, pStmt);
+  int32_t code = checkCreateTable(pCxt, pStmt);
   if (TSDB_CODE_SUCCESS != code) {
     return code;
   }
@@ -1485,7 +1489,7 @@ static int32_t nodeTypeToShowType(ENodeType nt) {
     case QUERY_NODE_SHOW_CONNECTIONS_STMT:
       return TSDB_MGMT_TABLE_CONNS;
     case QUERY_NODE_SHOW_LICENCE_STMT:
-      return 0; // todo
+      return TSDB_MGMT_TABLE_GRANTS;
     case QUERY_NODE_SHOW_QUERIES_STMT:
       return TSDB_MGMT_TABLE_QUERIES;
     case QUERY_NODE_SHOW_SCORES_STMT:
@@ -2157,11 +2161,11 @@ typedef struct SVgroupTablesBatch {
   char               dbName[TSDB_DB_NAME_LEN];
 } SVgroupTablesBatch;
 
-static void toSchema(const SColumnDefNode* pCol, col_id_t colId, SSchemaEx* pSchema) {
+static void toSchemaEx(const SColumnDefNode* pCol, col_id_t colId, SSchemaEx* pSchema) {
   pSchema->colId = colId;
   pSchema->type = pCol->dataType.type;
   pSchema->bytes = calcTypeBytes(pCol->dataType);
-  pSchema->sma = TSDB_BSMA_TYPE_LATEST;  // TODO: use default value currently, and use the real value later.
+  pSchema->sma = pCol->sma ? TSDB_BSMA_TYPE_LATEST : TSDB_BSMA_TYPE_NONE;
   strcpy(pSchema->name, pCol->colName);
 }
 
@@ -2171,33 +2175,60 @@ static void destroyCreateTbReq(SVCreateTbReq* pReq) {
   taosMemoryFreeClear(pReq->ntbCfg.pSchema);
 }
 
-static int32_t buildNormalTableBatchReq(int32_t acctId, const char* pDbName, const char* pTableName,
-    const SNodeList* pColumns, const SVgroupInfo* pVgroupInfo, SVgroupTablesBatch* pBatch) {
+static int32_t buildSmaParam(STableOptions* pOptions, SVCreateTbReq* pReq) {
+  if (0 == LIST_LENGTH(pOptions->pFuncs)) {
+    return TSDB_CODE_SUCCESS;
+  }
+
+  pReq->ntbCfg.pRSmaParam = taosMemoryCalloc(1, sizeof(SRSmaParam));
+  if (NULL == pReq->ntbCfg.pRSmaParam) {
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
+  pReq->ntbCfg.pRSmaParam->delay = pOptions->delay;
+  pReq->ntbCfg.pRSmaParam->xFilesFactor = pOptions->filesFactor;
+  pReq->ntbCfg.pRSmaParam->nFuncIds = LIST_LENGTH(pOptions->pFuncs);
+  pReq->ntbCfg.pRSmaParam->pFuncIds = taosMemoryCalloc(pReq->ntbCfg.pRSmaParam->nFuncIds, sizeof(func_id_t));
+  if (NULL == pReq->ntbCfg.pRSmaParam->pFuncIds) {
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
+  int32_t index = 0;
+  SNode* pFunc = NULL;
+  FOREACH(pFunc, pOptions->pFuncs) {
+    pReq->ntbCfg.pRSmaParam->pFuncIds[index++] = ((SFunctionNode*)pFunc)->funcId;
+  }
+
+  return TSDB_CODE_SUCCESS;
+}
+
+static int32_t buildNormalTableBatchReq(int32_t acctId, const SCreateTableStmt* pStmt, const SVgroupInfo* pVgroupInfo, SVgroupTablesBatch* pBatch) {
   char dbFName[TSDB_DB_FNAME_LEN] = {0};
   SName name = { .type = TSDB_DB_NAME_T, .acctId = acctId };
-  strcpy(name.dbname, pDbName);
+  strcpy(name.dbname, pStmt->dbName);
   tNameGetFullDbName(&name, dbFName);
 
   SVCreateTbReq req = {0};
   req.type = TD_NORMAL_TABLE;
   req.dbFName = strdup(dbFName);
-  req.name = strdup(pTableName);
-  req.ntbCfg.nCols = LIST_LENGTH(pColumns);
+  req.name = strdup(pStmt->tableName);
+  req.ntbCfg.nCols = LIST_LENGTH(pStmt->pCols);
   req.ntbCfg.pSchema = taosMemoryCalloc(req.ntbCfg.nCols, sizeof(SSchemaEx));
   if (NULL == req.name || NULL == req.ntbCfg.pSchema) {
     destroyCreateTbReq(&req);
     return TSDB_CODE_OUT_OF_MEMORY;
   }
   SNode* pCol;
-  int32_t index = 0;
-  FOREACH(pCol, pColumns) {
-    toSchema((SColumnDefNode*)pCol, index + 1, req.ntbCfg.pSchema + index);
+  col_id_t index = 0;
+  FOREACH(pCol, pStmt->pCols) {
+    toSchemaEx((SColumnDefNode*)pCol, index + 1, req.ntbCfg.pSchema + index);
     ++index;
   }
-  // TODO: use the real sma for normal table.
+  if (TSDB_CODE_SUCCESS != buildSmaParam(pStmt->pOptions, &req)) {
+    destroyCreateTbReq(&req);
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
 
   pBatch->info = *pVgroupInfo;
-  strcpy(pBatch->dbName, pDbName);
+  strcpy(pBatch->dbName, pStmt->dbName);
   pBatch->req.pArray = taosArrayInit(1, sizeof(struct SVCreateTbReq));
   if (NULL == pBatch->req.pArray) {
     destroyCreateTbReq(&req);
@@ -2278,7 +2309,7 @@ static int32_t buildCreateTableDataBlock(int32_t acctId, const SCreateTableStmt*
   }
 
   SVgroupTablesBatch tbatch = {0};
-  int32_t code = buildNormalTableBatchReq(acctId, pStmt->dbName, pStmt->tableName, pStmt->pCols, pInfo, &tbatch);
+  int32_t code = buildNormalTableBatchReq(acctId, pStmt, pInfo, &tbatch);
   if (TSDB_CODE_SUCCESS == code) {
     code = serializeVgroupTablesBatch(&tbatch, *pBufArray);
   }
@@ -2293,8 +2324,11 @@ static int32_t buildCreateTableDataBlock(int32_t acctId, const SCreateTableStmt*
 static int32_t rewriteCreateTable(STranslateContext* pCxt, SQuery* pQuery) {
   SCreateTableStmt* pStmt = (SCreateTableStmt*)pQuery->pRoot;
 
+  int32_t code = checkCreateTable(pCxt, pStmt);
   SVgroupInfo info = {0};
-  int32_t code = getTableHashVgroup(pCxt, pStmt->dbName, pStmt->tableName, &info);
+  if (TSDB_CODE_SUCCESS == code) {
+    code = getTableHashVgroup(pCxt, pStmt->dbName, pStmt->tableName, &info);
+  }
   SArray* pBufArray = NULL;
   if (TSDB_CODE_SUCCESS == code) {
     code = buildCreateTableDataBlock(pCxt->pParseCxt->acctId, pStmt, &info, &pBufArray);
diff --git a/source/libs/parser/test/parserAstTest.cpp b/source/libs/parser/test/parserAstTest.cpp
index 72670bb79a..8b19701c52 100644
--- a/source/libs/parser/test/parserAstTest.cpp
+++ b/source/libs/parser/test/parserAstTest.cpp
@@ -44,7 +44,7 @@ protected:
     query_ = nullptr;
     bool res = runImpl(parseCode, translateCode);
     qDestroyQuery(query_);
-    if (!res) {
+    if (1/*!res*/) {
       dump();
     }
     return res;
@@ -69,6 +69,12 @@ private:
       return (terrno == translateCode);
     }
     translatedAstStr_ = toString(query_->pRoot);
+    code = calculateConstant(&cxt_, query_);
+    if (code != TSDB_CODE_SUCCESS) {
+      calcConstErrStr_ = string("code:") + tstrerror(code) + string(", msg:") + errMagBuf_;
+      return false;
+    }
+    calcConstAstStr_ = toString(query_->pRoot);
     return (TSDB_CODE_SUCCESS == translateCode);
   }
 
@@ -88,6 +94,13 @@ private:
       cout << "translate output: " << endl;
       cout << translatedAstStr_ << endl;
     }
+    if (!calcConstErrStr_.empty()) {
+      cout << "calculateConstant error: " << calcConstErrStr_ << endl;
+    }
+    if (!calcConstAstStr_.empty()) {
+      cout << "calculateConstant output: " << endl;
+      cout << calcConstAstStr_ << endl;
+    }
   }
 
   string toString(const SNode* pRoot, bool format = false) {
@@ -112,6 +125,8 @@ private:
     parsedAstStr_.clear();
     translateErrStr_.clear();
     translatedAstStr_.clear();
+    calcConstErrStr_.clear();
+    calcConstAstStr_.clear();
   }
 
   string acctId_;
@@ -124,6 +139,8 @@ private:
   string parsedAstStr_;
   string translateErrStr_;
   string translatedAstStr_;
+  string calcConstErrStr_;
+  string calcConstAstStr_;
 };
 
 TEST_F(ParserTest, createAccount) {
@@ -191,6 +208,9 @@ TEST_F(ParserTest, selectConstant) {
 
   bind("SELECT 1234567890123456789012345678901234567890, 20.1234567890123456789012345678901234567890, 'abc', \"wxy\", TIMESTAMP '2022-02-09 17:30:20', true, false, 15s FROM t1");
   ASSERT_TRUE(run());
+
+  bind("SELECT 123 + 45 FROM t1 where 2 - 1");
+  ASSERT_TRUE(run());
 }
 
 TEST_F(ParserTest, selectExpression) {

From eafc0f88bc5f6655a3906ede85ffcc9f15d43f29 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 11:00:52 +0000
Subject: [PATCH 23/25] remove tid from tsdb

---
 source/dnode/vnode/src/inc/tsdbReadImpl.h  | 32 +++++++++++-----------
 source/dnode/vnode/src/tsdb/tsdbCommit.c   |  1 -
 source/dnode/vnode/src/tsdb/tsdbReadImpl.c | 31 +++++++++++----------
 3 files changed, 32 insertions(+), 32 deletions(-)

diff --git a/source/dnode/vnode/src/inc/tsdbReadImpl.h b/source/dnode/vnode/src/inc/tsdbReadImpl.h
index 90a877bb83..682306221d 100644
--- a/source/dnode/vnode/src/inc/tsdbReadImpl.h
+++ b/source/dnode/vnode/src/inc/tsdbReadImpl.h
@@ -17,12 +17,12 @@
 #define _TD_TSDB_READ_IMPL_H_
 
 #include "os.h"
+#include "tcommon.h"
 #include "tfs.h"
 #include "tsdb.h"
 #include "tsdbFile.h"
-#include "tskiplist.h"
 #include "tsdbMemory.h"
-#include "tcommon.h"
+#include "tskiplist.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -31,7 +31,6 @@ extern "C" {
 typedef struct SReadH SReadH;
 
 typedef struct {
-  int32_t  tid;
   uint32_t len;
   uint32_t offset;
   uint32_t hasLast : 2;
@@ -81,7 +80,7 @@ typedef struct {
   TSKEY    keyLast;
 } SBlockV0;
 
-#define SBlock          SBlockV0  // latest SBlock definition
+#define SBlock SBlockV0  // latest SBlock definition
 
 #endif
 
@@ -165,19 +164,19 @@ typedef struct {
 typedef void SAggrBlkData;  // SBlockCol cols[];
 
 struct SReadH {
-  STsdb *       pRepo;
+  STsdb        *pRepo;
   SDFileSet     rSet;     // FSET to read
-  SArray *      aBlkIdx;  // SBlockIdx array
-  STable *      pTable;   // table to read
-  SBlockIdx *   pBlkIdx;  // current reading table SBlockIdx
+  SArray       *aBlkIdx;  // SBlockIdx array
+  STable       *pTable;   // table to read
+  SBlockIdx    *pBlkIdx;  // current reading table SBlockIdx
   int           cidx;
-  SBlockInfo *  pBlkInfo;
-  SBlockData *  pBlkData;      // Block info
+  SBlockInfo   *pBlkInfo;
+  SBlockData   *pBlkData;      // Block info
   SAggrBlkData *pAggrBlkData;  // Aggregate Block info
-  SDataCols *   pDCols[2];
-  void *        pBuf;    // buffer
-  void *        pCBuf;   // compression buffer
-  void *        pExBuf;  // extra buffer
+  SDataCols    *pDCols[2];
+  void         *pBuf;    // buffer
+  void         *pCBuf;   // compression buffer
+  void         *pExBuf;  // extra buffer
 };
 
 #define TSDB_READ_REPO(rh)      ((rh)->pRepo)
@@ -222,14 +221,15 @@ int   tsdbLoadBlockIdx(SReadH *pReadh);
 int   tsdbSetReadTable(SReadH *pReadh, STable *pTable);
 int   tsdbLoadBlockInfo(SReadH *pReadh, void *pTarget);
 int   tsdbLoadBlockData(SReadH *pReadh, SBlock *pBlock, SBlockInfo *pBlockInfo);
-int   tsdbLoadBlockDataCols(SReadH *pReadh, SBlock *pBlock, SBlockInfo *pBlkInfo, const int16_t *colIds, int numOfColsIds);
+int   tsdbLoadBlockDataCols(SReadH *pReadh, SBlock *pBlock, SBlockInfo *pBlkInfo, const int16_t *colIds,
+                            int numOfColsIds);
 int   tsdbLoadBlockStatis(SReadH *pReadh, SBlock *pBlock);
 int   tsdbEncodeSBlockIdx(void **buf, SBlockIdx *pIdx);
 void *tsdbDecodeSBlockIdx(void *buf, SBlockIdx *pIdx);
 void  tsdbGetBlockStatis(SReadH *pReadh, SDataStatis *pStatis, int numOfCols, SBlock *pBlock);
 
 static FORCE_INLINE int tsdbMakeRoom(void **ppBuf, size_t size) {
-  void * pBuf = *ppBuf;
+  void  *pBuf = *ppBuf;
   size_t tsize = taosTSizeof(pBuf);
 
   if (tsize < size) {
diff --git a/source/dnode/vnode/src/tsdb/tsdbCommit.c b/source/dnode/vnode/src/tsdb/tsdbCommit.c
index 2a8875fefe..18d3de94cc 100644
--- a/source/dnode/vnode/src/tsdb/tsdbCommit.c
+++ b/source/dnode/vnode/src/tsdb/tsdbCommit.c
@@ -701,7 +701,6 @@ int tsdbWriteBlockInfoImpl(SDFile *pHeadf, STable *pTable, SArray *pSupA, SArray
   // Set pIdx
   pBlock = taosArrayGetLast(pSupA);
 
-  pIdx->tid = TABLE_TID(pTable);
   pIdx->uid = TABLE_UID(pTable);
   pIdx->hasLast = pBlock->last ? 1 : 0;
   pIdx->maxKey = pBlock->keyLast;
diff --git a/source/dnode/vnode/src/tsdb/tsdbReadImpl.c b/source/dnode/vnode/src/tsdb/tsdbReadImpl.c
index 8cb90d76a6..687e2ca776 100644
--- a/source/dnode/vnode/src/tsdb/tsdbReadImpl.c
+++ b/source/dnode/vnode/src/tsdb/tsdbReadImpl.c
@@ -98,7 +98,7 @@ int tsdbSetAndOpenReadFSet(SReadH *pReadh, SDFileSet *pSet) {
 void tsdbCloseAndUnsetFSet(SReadH *pReadh) { tsdbResetReadFile(pReadh); }
 
 int tsdbLoadBlockIdx(SReadH *pReadh) {
-  SDFile *  pHeadf = TSDB_READ_HEAD_FILE(pReadh);
+  SDFile   *pHeadf = TSDB_READ_HEAD_FILE(pReadh);
   SBlockIdx blkIdx;
 
   ASSERT(taosArrayGetSize(pReadh->aBlkIdx) == 0);
@@ -149,8 +149,8 @@ int tsdbLoadBlockIdx(SReadH *pReadh) {
     }
 
     tsize++;
-    ASSERT(tsize == 1 || ((SBlockIdx *)taosArrayGet(pReadh->aBlkIdx, tsize - 2))->tid <
-                             ((SBlockIdx *)taosArrayGet(pReadh->aBlkIdx, tsize - 1))->tid);
+    // ASSERT(tsize == 1 || ((SBlockIdx *)taosArrayGet(pReadh->aBlkIdx, tsize - 2))->tid <
+    //                          ((SBlockIdx *)taosArrayGet(pReadh->aBlkIdx, tsize - 1))->tid);
   }
 
   return 0;
@@ -180,7 +180,7 @@ int tsdbSetReadTable(SReadH *pReadh, STable *pTable) {
       }
 
       SBlockIdx *pBlkIdx = taosArrayGet(pReadh->aBlkIdx, pReadh->cidx);
-      if (pBlkIdx->tid == TABLE_TID(pTable)) {
+      if (pBlkIdx->uid == TABLE_TID(pTable)) {
         if (pBlkIdx->uid == TABLE_UID(pTable)) {
           pReadh->pBlkIdx = pBlkIdx;
         } else {
@@ -188,7 +188,7 @@ int tsdbSetReadTable(SReadH *pReadh, STable *pTable) {
         }
         pReadh->cidx++;
         break;
-      } else if (pBlkIdx->tid > TABLE_TID(pTable)) {
+      } else if (pBlkIdx->uid > TABLE_TID(pTable)) {
         pReadh->pBlkIdx = NULL;
         break;
       } else {
@@ -205,7 +205,7 @@ int tsdbSetReadTable(SReadH *pReadh, STable *pTable) {
 int tsdbLoadBlockInfo(SReadH *pReadh, void *pTarget) {
   ASSERT(pReadh->pBlkIdx != NULL);
 
-  SDFile *   pHeadf = TSDB_READ_HEAD_FILE(pReadh);
+  SDFile    *pHeadf = TSDB_READ_HEAD_FILE(pReadh);
   SBlockIdx *pBlkIdx = pReadh->pBlkIdx;
 
   if (tsdbSeekDFile(pHeadf, pBlkIdx->offset, SEEK_SET) < 0) {
@@ -237,7 +237,7 @@ int tsdbLoadBlockInfo(SReadH *pReadh, void *pTarget) {
     return -1;
   }
 
-  ASSERT(pBlkIdx->tid == pReadh->pBlkInfo->tid && pBlkIdx->uid == pReadh->pBlkInfo->uid);
+  // ASSERT(pBlkIdx->tid == pReadh->pBlkInfo->tid && pBlkIdx->uid == pReadh->pBlkInfo->uid);
 
   if (pTarget) {
     memcpy(pTarget, (void *)(pReadh->pBlkInfo), pBlkIdx->len);
@@ -275,7 +275,8 @@ int tsdbLoadBlockData(SReadH *pReadh, SBlock *pBlock, SBlockInfo *pBlkInfo) {
   return 0;
 }
 
-int tsdbLoadBlockDataCols(SReadH *pReadh, SBlock *pBlock, SBlockInfo *pBlkInfo, const int16_t *colIds, int numOfColsIds) {
+int tsdbLoadBlockDataCols(SReadH *pReadh, SBlock *pBlock, SBlockInfo *pBlkInfo, const int16_t *colIds,
+                          int numOfColsIds) {
   ASSERT(pBlock->numOfSubBlocks > 0);
   int8_t update = pReadh->pRepo->config.update;
 
@@ -388,7 +389,7 @@ static int tsdbLoadBlockOffset(SReadH *pReadh, SBlock *pBlock) {
 int tsdbEncodeSBlockIdx(void **buf, SBlockIdx *pIdx) {
   int tlen = 0;
 
-  tlen += taosEncodeVariantI32(buf, pIdx->tid);
+  // tlen += taosEncodeVariantI32(buf, pIdx->tid);
   tlen += taosEncodeVariantU32(buf, pIdx->len);
   tlen += taosEncodeVariantU32(buf, pIdx->offset);
   tlen += taosEncodeFixedU8(buf, pIdx->hasLast);
@@ -404,7 +405,7 @@ void *tsdbDecodeSBlockIdx(void *buf, SBlockIdx *pIdx) {
   uint32_t numOfBlocks = 0;
   uint64_t value = 0;
 
-  if ((buf = taosDecodeVariantI32(buf, &(pIdx->tid))) == NULL) return NULL;
+  // if ((buf = taosDecodeVariantI32(buf, &(pIdx->tid))) == NULL) return NULL;
   if ((buf = taosDecodeVariantU32(buf, &(pIdx->len))) == NULL) return NULL;
   if ((buf = taosDecodeVariantU32(buf, &(pIdx->offset))) == NULL) return NULL;
   if ((buf = taosDecodeFixedU8(buf, &(hasLast))) == NULL) return NULL;
@@ -538,9 +539,9 @@ static int tsdbLoadBlockDataImpl(SReadH *pReadh, SBlock *pBlock, SDataCols *pDat
   pDataCols->numOfRows = pBlock->numOfRows;
 
   // Recover the data
-  int ccol = 0;  // loop iter for SBlockCol object
-  int dcol = 0;  // loop iter for SDataCols object
-  int nBitmaps = (int)TD_BITMAP_BYTES(pBlock->numOfRows);
+  int        ccol = 0;  // loop iter for SBlockCol object
+  int        dcol = 0;  // loop iter for SDataCols object
+  int        nBitmaps = (int)TD_BITMAP_BYTES(pBlock->numOfRows);
   SBlockCol *pBlockCol = NULL;
   while (dcol < pDataCols->numOfCols) {
     SDataCol *pDataCol = &(pDataCols->cols[dcol]);
@@ -686,7 +687,7 @@ static int tsdbLoadBlockDataColsImpl(SReadH *pReadh, SBlock *pBlock, SDataCols *
   ASSERT(pBlock->numOfSubBlocks == 0 || pBlock->numOfSubBlocks == 1);
   ASSERT(colIds[0] == PRIMARYKEY_TIMESTAMP_COL_ID);
 
-  SDFile *  pDFile = (pBlock->last) ? TSDB_READ_LAST_FILE(pReadh) : TSDB_READ_DATA_FILE(pReadh);
+  SDFile   *pDFile = (pBlock->last) ? TSDB_READ_LAST_FILE(pReadh) : TSDB_READ_DATA_FILE(pReadh);
   SBlockCol blockCol = {0};
 
   tdResetDataCols(pDataCols);
@@ -700,7 +701,7 @@ static int tsdbLoadBlockDataColsImpl(SReadH *pReadh, SBlock *pBlock, SDataCols *
   int ccol = 0;
   for (int i = 0; i < numOfColIds; i++) {
     int16_t    colId = colIds[i];
-    SDataCol * pDataCol = NULL;
+    SDataCol  *pDataCol = NULL;
     SBlockCol *pBlockCol = NULL;
 
     while (true) {

From 70810ddb39382910b950390897d582cc30eecd90 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Wed, 6 Apr 2022 11:37:34 +0000
Subject: [PATCH 24/25] fix a small bug

---
 source/dnode/vnode/src/meta/metaTDBImpl.c |  8 ++++++++
 source/libs/tdb/src/db/tdbBtree.c         | 20 +++++++++++---------
 2 files changed, 19 insertions(+), 9 deletions(-)

diff --git a/source/dnode/vnode/src/meta/metaTDBImpl.c b/source/dnode/vnode/src/meta/metaTDBImpl.c
index 3ab26c3d4d..812ec84a9a 100644
--- a/source/dnode/vnode/src/meta/metaTDBImpl.c
+++ b/source/dnode/vnode/src/meta/metaTDBImpl.c
@@ -226,6 +226,14 @@ int metaSaveTableToDB(SMeta *pMeta, STbCfg *pTbCfg) {
     uid = metaGenerateUid(pMeta);
   }
 
+  // check name and uid unique
+  if (tdbDbGet(pMetaDb->pTbDB, &uid, sizeof(uid), NULL, NULL) == 0) {
+    return -1;
+  }
+  if (tdbDbGet(pMetaDb->pNameIdx, pTbCfg->name, strlen(pTbCfg->name) + 1, NULL, NULL) == 0) {
+    return -1;
+  }
+
   // save to table.db
   pKey = &uid;
   kLen = sizeof(uid);
diff --git a/source/libs/tdb/src/db/tdbBtree.c b/source/libs/tdb/src/db/tdbBtree.c
index f760598714..9df749bec7 100644
--- a/source/libs/tdb/src/db/tdbBtree.c
+++ b/source/libs/tdb/src/db/tdbBtree.c
@@ -233,7 +233,7 @@ int tdbBtreePGet(SBTree *pBt, const void *pKey, int kLen, void **ppKey, int *pkL
     ASSERT(0);
   }
 
-  if (cret) {
+  if (btc.idx < 0 || cret) {
     tdbBtcClose(&btc);
     return -1;
   }
@@ -253,15 +253,17 @@ int tdbBtreePGet(SBTree *pBt, const void *pKey, int kLen, void **ppKey, int *pkL
     memcpy(*ppKey, cd.pKey, cd.kLen);
   }
 
-  pTVal = TDB_REALLOC(*ppVal, cd.vLen);
-  if (pTVal == NULL) {
-    tdbBtcClose(&btc);
-    ASSERT(0);
-    return -1;
+  if (ppVal) {
+    pTVal = TDB_REALLOC(*ppVal, cd.vLen);
+    if (pTVal == NULL) {
+      tdbBtcClose(&btc);
+      ASSERT(0);
+      return -1;
+    }
+    *ppVal = pTVal;
+    *vLen = cd.vLen;
+    memcpy(*ppVal, cd.pVal, cd.vLen);
   }
-  *ppVal = pTVal;
-  *vLen = cd.vLen;
-  memcpy(*ppVal, cd.pVal, cd.vLen);
 
   tdbBtcClose(&btc);
 

From ead76ab6d2c599cb1fa2c488ad463d91f64ee471 Mon Sep 17 00:00:00 2001
From: Xiaoyu Wang <xiaoyuwang@taosdata.com>
Date: Wed, 6 Apr 2022 19:59:27 +0800
Subject: [PATCH 25/25] normal table rollup translate, and condition rewrite

---
 source/libs/parser/src/parAstParser.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/source/libs/parser/src/parAstParser.c b/source/libs/parser/src/parAstParser.c
index 8d0ddafd26..76af6b7ac7 100644
--- a/source/libs/parser/src/parAstParser.c
+++ b/source/libs/parser/src/parAstParser.c
@@ -65,7 +65,7 @@ int32_t parse(SParseContext* pParseCxt, SQuery** pQuery) {
       }
       default:
         Parse(pParser, t0.type, t0, &cxt);
-        ParseTrace(stdout, "");
+        // ParseTrace(stdout, "");
         if (!cxt.valid) {
           goto abort_parse;
         }