Merge branch 'main' of https://github.com/taosdata/TDengine into main

CMakeLists.txt

@@ -10,6 +10,8 @@ if (NOT DEFINED TD_SOURCE_DIR)
   set( TD_SOURCE_DIR ${PROJECT_SOURCE_DIR} )
 endif()
 
+SET(TD_COMMUNITY_DIR ${PROJECT_SOURCE_DIR})
+
 set(TD_SUPPORT_DIR "${TD_SOURCE_DIR}/cmake")
 set(TD_CONTRIB_DIR "${TD_SOURCE_DIR}/contrib")
 

cmake/cmake.version

@@ -16,7 +16,7 @@ find_program(HAVE_GIT NAMES git)
 IF (DEFINED GITINFO)
   SET(TD_VER_GIT ${GITINFO})
 ELSEIF (HAVE_GIT)
-  execute_process(COMMAND git log -1 --format=%H WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR} OUTPUT_VARIABLE GIT_COMMITID)
+  execute_process(COMMAND git log -1 --format=%H WORKING_DIRECTORY ${TD_COMMUNITY_DIR} OUTPUT_VARIABLE GIT_COMMITID)
   #message(STATUS "git log result:${GIT_COMMITID}")
   IF (GIT_COMMITID)
     string (REGEX REPLACE "[\n\t\r]" "" GIT_COMMITID ${GIT_COMMITID})
@@ -30,6 +30,23 @@ ELSE ()
   SET(TD_VER_GIT "no git commit id")
 ENDIF ()
 
+IF (DEFINED GITINFOI)
+  SET(TD_VER_GIT_INTERNAL ${GITINFOI})
+ELSEIF (HAVE_GIT)
+  execute_process(COMMAND git log -1 --format=%H WORKING_DIRECTORY ${PROJECT_SOURCE_DIR} OUTPUT_VARIABLE GIT_COMMITID)
+  message(STATUS "git log result:${GIT_COMMITID}")
+  IF (GIT_COMMITID)
+    string (REGEX REPLACE "[\n\t\r]" "" GIT_COMMITID ${GIT_COMMITID})
+    SET(TD_VER_GIT_INTERNAL ${GIT_COMMITID})
+  ELSE ()
+    message(STATUS "not a git repository")
+    SET(TD_VER_GIT "no git commit id")
+  ENDIF ()
+ELSE ()
+  message(STATUS "no git cmd")
+  SET(TD_VER_GIT_INTERNAL "no git commit id")
+ENDIF ()
+
 IF (DEFINED VERDATE)
   SET(TD_VER_DATE ${VERDATE})
 ELSE ()

cmake/taosadapter_CMakeLists.txt.in

@@ -2,7 +2,7 @@
 # taosadapter
 ExternalProject_Add(taosadapter
         GIT_REPOSITORY https://github.com/taosdata/taosadapter.git
-        GIT_TAG d8059ff
+        GIT_TAG cb1e89c
         SOURCE_DIR "${TD_SOURCE_DIR}/tools/taosadapter"
         BINARY_DIR ""
         #BUILD_IN_SOURCE TRUE

cmake/taostools_CMakeLists.txt.in

@@ -2,7 +2,7 @@
 # taos-tools
 ExternalProject_Add(taos-tools
         GIT_REPOSITORY https://github.com/taosdata/taos-tools.git
-        GIT_TAG 04296a5
+        GIT_TAG d194dc9
         SOURCE_DIR "${TD_SOURCE_DIR}/tools/taos-tools"
         BINARY_DIR ""
         #BUILD_IN_SOURCE TRUE

docs/en/07-develop/07-tmq.mdx

@@ -293,7 +293,6 @@ You configure the following parameters when creating a consumer:
 |      `auto.offset.reset`       |  enum   | Initial offset for the consumer group                                     | Specify `earliest`, `latest`, or `none`(default) |
 |      `enable.auto.commit`      | boolean | Commit automatically                                             | Specify `true` or `false`.                   |
 |   `auto.commit.interval.ms`    | integer | Interval for automatic commits, in milliseconds                           |
-| `enable.heartbeat.background`  | boolean | Backend heartbeat; if enabled, the consumer does not go offline even if it has not polled for a long time |                                             |
 | `experimental.snapshot.enable` | boolean | Specify whether to consume messages from the WAL or from TSBS                    |                                             |
 |     `msg.with.table.name`      | boolean | Specify whether to deserialize table names from messages                                 |
 
@@ -368,7 +367,6 @@ conf := &tmq.ConfigMap{
  "td.connect.port":              "6030",
  "client.id":                    "test_tmq_c",
  "enable.auto.commit":           "false",
- "enable.heartbeat.background":  "true",
  "experimental.snapshot.enable": "true",
  "msg.with.table.name":          "true",
 }
@@ -418,7 +416,6 @@ Python programs use the following parameters:
 | `auto.commit.interval.ms` | string | Interval for automatic commits, in milliseconds | |
 | `auto.offset.reset` | string | Initial offset for the consumer group | Specify `earliest`, `latest`, or `none`(default) |
 | `experimental.snapshot.enable` | string | Specify whether to consume messages from the WAL or from TSDB | Specify `true` or `false` |
-| `enable.heartbeat.background` | string | Backend heartbeat; if enabled, the consumer does not go offline even if it has not polled for a long time | Specify `true` or `false` |
 
 </TabItem>
 

docs/examples/go/sub/main.go

@@ -35,7 +35,6 @@ func main() {
 		"td.connect.port":              "6030",
 		"client.id":                    "test_tmq_client",
 		"enable.auto.commit":           "false",
-		"enable.heartbeat.background":  "true",
 		"experimental.snapshot.enable": "true",
 		"msg.with.table.name":          "true",
 	})

docs/zh/07-develop/07-tmq.mdx

@@ -291,7 +291,6 @@ CREATE TOPIC topic_name AS DATABASE db_name;
 |      `auto.offset.reset`       |  enum   | 消费组订阅的初始位置                                     | 可选：`earliest`(default), `latest`, `none` |
 |      `enable.auto.commit`      | boolean | 是否启用消费位点自动提交                                 | 合法值：`true`, `false`。                   |
 |   `auto.commit.interval.ms`    | integer | 以毫秒为单位的消费记录自动提交消费位点时间间             | 默认 5000 m                                 |
-| `enable.heartbeat.background`  | boolean | 启用后台心跳，启用后即使长时间不 poll 消息也不会造成离线 | 默认开启                                    |
 | `experimental.snapshot.enable` | boolean | 是否允许从 TSDB 消费数据                                 | 实验功能，默认关闭                          |
 |     `msg.with.table.name`      | boolean | 是否允许从消息中解析表名, 不适用于列订阅（列订阅时可将 tbname 作为列写入 subquery 语句）               | |
 
@@ -366,7 +365,6 @@ conf := &tmq.ConfigMap{
  "td.connect.port":              "6030",
  "client.id":                    "test_tmq_c",
  "enable.auto.commit":           "false",
- "enable.heartbeat.background":  "true",
  "experimental.snapshot.enable": "true",
  "msg.with.table.name":          "true",
 }
@@ -418,7 +416,6 @@ consumer = Consumer({"group.id": "local", "td.connect.ip": "127.0.0.1"})
 | `auto.commit.interval.ms` | string | 以毫秒为单位的自动提交时间间隔 | 默认值：5000 ms |
 | `auto.offset.reset` | string | 消费组订阅的初始位置 | 可选：`earliest`(default), `latest`, `none` |
 | `experimental.snapshot.enable` | string | 是否允许从 TSDB 消费数据 | 合法值：`true`, `false` |
-| `enable.heartbeat.background` | string | 启用后台心跳，启用后即使长时间不 poll 消息也不会造成离线 | 合法值：`true`, `false` |
 
 </TabItem>
 

include/common/tcommon.h

@@ -185,7 +185,7 @@ typedef struct SBlockID {
 typedef struct SDataBlockInfo {
   STimeWindow window;
   int32_t     rowSize;
-  int32_t     rows;  // todo hide this attribute
+  int64_t     rows;  // todo hide this attribute
   uint32_t    capacity;
   SBlockID    id;
   int16_t     hasVarCol;

include/common/tdatablock.h

@@ -178,7 +178,7 @@ int32_t getJsonValueLen(const char* data);
 
 int32_t colDataSetVal(SColumnInfoData* pColumnInfoData, uint32_t rowIndex, const char* pData, bool isNull);
 int32_t colDataAppend(SColumnInfoData* pColumnInfoData, uint32_t rowIndex, const char* pData, bool isNull);
-int32_t colDataSetNItems(SColumnInfoData* pColumnInfoData, uint32_t rowIndex, const char* pData, uint32_t numOfRows);
+int32_t colDataSetNItems(SColumnInfoData* pColumnInfoData, uint32_t rowIndex, const char* pData, uint32_t numOfRows, bool trimValue);
 int32_t colDataMergeCol(SColumnInfoData* pColumnInfoData, int32_t numOfRow1, int32_t* capacity,
                         const SColumnInfoData* pSource, int32_t numOfRow2);
 int32_t colDataAssign(SColumnInfoData* pColumnInfoData, const SColumnInfoData* pSource, int32_t numOfRows,

include/common/tmsgcb.h

@@ -34,7 +34,7 @@ typedef enum {
   WRITE_QUEUE,
   APPLY_QUEUE,
   SYNC_QUEUE,
-  SYNC_CTRL_QUEUE,
+  SYNC_RD_QUEUE,
   STREAM_QUEUE,
   QUEUE_MAX,
 } EQueueType;

include/common/tmsgdef.h

@@ -259,7 +259,7 @@ enum {
 
   TD_NEW_MSG_SEG(TDMT_SYNC_MSG)
   TD_DEF_MSG_TYPE(TDMT_SYNC_TIMEOUT, "sync-timer", NULL, NULL)
-  TD_DEF_MSG_TYPE(TDMT_SYNC_PING, "sync-ping", NULL, NULL)             // no longer used
+  TD_DEF_MSG_TYPE(TDMT_SYNC_TIMEOUT_ELECTION, "sync-elect", NULL, NULL)
   TD_DEF_MSG_TYPE(TDMT_SYNC_PING_REPLY, "sync-ping-reply", NULL, NULL) // no longer used
   TD_DEF_MSG_TYPE(TDMT_SYNC_CLIENT_REQUEST, "sync-client-request", NULL, NULL)
   TD_DEF_MSG_TYPE(TDMT_SYNC_CLIENT_REQUEST_BATCH, "sync-client-request-batch", NULL, NULL)

include/libs/executor/executor.h

@@ -78,7 +78,7 @@ qTaskInfo_t qCreateStreamExecTaskInfo(void* msg, SReadHandle* readers, int32_t v
  * @param SReadHandle
  * @return
  */
-qTaskInfo_t qCreateQueueExecTaskInfo(void* msg, SReadHandle* pReaderHandle, int32_t vgId, int32_t* numOfCols, SSchemaWrapper** pSchema);
+qTaskInfo_t qCreateQueueExecTaskInfo(void* msg, SReadHandle* pReaderHandle, int32_t vgId, int32_t* numOfCols, uint64_t id);
 
 /**
  * set the task Id, usually used by message queue process
@@ -89,6 +89,7 @@ qTaskInfo_t qCreateQueueExecTaskInfo(void* msg, SReadHandle* pReaderHandle, int3
 void qSetTaskId(qTaskInfo_t tinfo, uint64_t taskId, uint64_t queryId);
 
 int32_t qSetStreamOpOpen(qTaskInfo_t tinfo);
+
 /**
  * Set multiple input data blocks for the stream scan.
  * @param tinfo
@@ -149,7 +150,6 @@ int32_t qGetQueryTableSchemaVersion(qTaskInfo_t tinfo, char* dbName, char* table
  * @param handle
  * @return
  */
-
 int32_t qExecTaskOpt(qTaskInfo_t tinfo, SArray* pResList, uint64_t* useconds, bool* hasMore, SLocalFetch* pLocal);
 
 int32_t qExecTask(qTaskInfo_t tinfo, SSDataBlock** pBlock, uint64_t* useconds);
@@ -162,6 +162,7 @@ void qCleanExecTaskBlockBuf(qTaskInfo_t tinfo);
  * @return
  */
 int32_t qAsyncKillTask(qTaskInfo_t tinfo, int32_t rspCode);
+int32_t qKillTask(qTaskInfo_t tinfo, int32_t rspCode);
 
 bool qTaskIsExecuting(qTaskInfo_t qinfo);
 
@@ -171,14 +172,6 @@ bool qTaskIsExecuting(qTaskInfo_t qinfo);
  */
 void qDestroyTask(qTaskInfo_t tinfo);
 
-/**
- * Extract the qualified table id list, and than pass them to the TSDB driver to load the required table data blocks.
- *
- * @param iter  the table iterator to traverse all tables belongs to a super table, or an invert index
- * @return
- */
-int32_t qGetQualifiedTableIdList(void* pTableList, const char* tagCond, int32_t tagCondLen, SArray* pTableIdList);
-
 void qProcessRspMsg(void* parent, struct SRpcMsg* pMsg, struct SEpSet* pEpSet);
 
 int32_t qGetExplainExecInfo(qTaskInfo_t tinfo, SArray* pExecInfoList /*,int32_t* resNum, SExplainExecInfo** pRes*/);

include/libs/function/function.h

@@ -112,7 +112,7 @@ typedef struct SResultDataInfo {
 typedef struct SInputColumnInfoData {
   int32_t           totalRows;        // total rows in current columnar data
   int32_t           startRowIndex;    // handle started row index
-  int32_t           numOfRows;        // the number of rows needs to be handled
+  int64_t           numOfRows;        // the number of rows needs to be handled
   int32_t           numOfInputCols;   // PTS is not included
   bool              colDataSMAIsSet;  // if agg is set or not
   SColumnInfoData  *pPTS;             // primary timestamp column

include/libs/function/functionMgt.h

@@ -241,6 +241,7 @@ int32_t fmGetUdafExecFuncs(int32_t funcId, SFuncExecFuncs* pFpSet);
 int32_t fmSetInvertFunc(int32_t funcId, SFuncExecFuncs* pFpSet);
 int32_t fmSetNormalFunc(int32_t funcId, SFuncExecFuncs* pFpSet);
 bool    fmIsInvertible(int32_t funcId);
+char*   fmGetFuncName(int32_t funcId);
 
 #ifdef __cplusplus
 }

include/libs/stream/streamState.h

@@ -23,20 +23,19 @@ extern "C" {
 #ifndef _STREAM_STATE_H_
 #define _STREAM_STATE_H_
 
-typedef struct SStreamTask SStreamTask;
-
 typedef bool (*state_key_cmpr_fn)(void* pKey1, void* pKey2);
 
 typedef struct STdbState {
-  SStreamTask* pOwner;
+  struct SStreamTask* pOwner;
-  TDB*         db;
+
-  TTB*         pStateDb;
+  TDB* db;
-  TTB*         pFuncStateDb;
+  TTB* pStateDb;
-  TTB*         pFillStateDb;  // todo refactor
+  TTB* pFuncStateDb;
-  TTB*         pSessionStateDb;
+  TTB* pFillStateDb;  // todo refactor
-  TTB*         pParNameDb;
+  TTB* pSessionStateDb;
-  TTB*         pParTagDb;
+  TTB* pParNameDb;
-  TXN*         txn;
+  TTB* pParTagDb;
+  TXN* txn;
 } STdbState;
 
 // incremental state storage
@@ -45,7 +44,7 @@ typedef struct {
   int32_t    number;
 } SStreamState;
 
-SStreamState* streamStateOpen(char* path, SStreamTask* pTask, bool specPath, int32_t szPage, int32_t pages);
+SStreamState* streamStateOpen(char* path, struct SStreamTask* pTask, bool specPath, int32_t szPage, int32_t pages);
 void          streamStateClose(SStreamState* pState);
 int32_t       streamStateBegin(SStreamState* pState);
 int32_t       streamStateCommit(SStreamState* pState);

include/libs/stream/tstream.h

@@ -295,7 +295,7 @@ typedef struct {
   SEpSet  epSet;
 } SStreamChildEpInfo;
 
-typedef struct SStreamTask {
+struct SStreamTask {
   int64_t streamId;
   int32_t taskId;
   int32_t totalLevel;
@@ -362,8 +362,7 @@ typedef struct SStreamTask {
 
   int64_t checkpointingId;
   int32_t checkpointAlignCnt;
-
+};
-} SStreamTask;
 
 int32_t tEncodeStreamEpInfo(SEncoder* pEncoder, const SStreamChildEpInfo* pInfo);
 int32_t tDecodeStreamEpInfo(SDecoder* pDecoder, SStreamChildEpInfo* pInfo);
@@ -587,7 +586,7 @@ int32_t streamProcessDispatchReq(SStreamTask* pTask, SStreamDispatchReq* pReq, S
 int32_t streamProcessDispatchRsp(SStreamTask* pTask, SStreamDispatchRsp* pRsp, int32_t code);
 
 int32_t streamProcessRetrieveReq(SStreamTask* pTask, SStreamRetrieveReq* pReq, SRpcMsg* pMsg);
-int32_t streamProcessRetrieveRsp(SStreamTask* pTask, SStreamRetrieveRsp* pRsp);
+// int32_t streamProcessRetrieveRsp(SStreamTask* pTask, SStreamRetrieveRsp* pRsp);
 
 int32_t streamTryExec(SStreamTask* pTask);
 int32_t streamSchedExec(SStreamTask* pTask);

include/libs/sync/sync.h

@@ -48,7 +48,7 @@ extern "C" {
 #define SYNC_HEARTBEAT_REPLY_SLOW_MS 1500
 #define SYNC_SNAP_RESEND_MS          1000 * 60
 
-#define SYNC_VND_COMMIT_MIN_MS 1000
+#define SYNC_VND_COMMIT_MIN_MS 3000
 
 #define SYNC_MAX_BATCH_SIZE 1
 #define SYNC_INDEX_BEGIN    0

include/util/tarray.h

@@ -69,14 +69,6 @@ void* taosArrayAddBatch(SArray* pArray, const void* pData, int32_t nEles);
  */
 void taosArrayRemoveDuplicate(SArray* pArray, __compar_fn_t comparFn, void (*fp)(void*));
 
-/**
- *
- * @param pArray
- * @param comparFn
- * @param fp
- */
-void taosArrayRemoveDuplicateP(SArray* pArray, __compar_fn_t comparFn, void (*fp)(void*));
-
 /**
  *  add all element from the source array list into the destination
  * @param pArray
@@ -216,6 +208,8 @@ void  taosArrayDestroyP(SArray* pArray, FDelete fp);
 
 void  taosArrayDestroyEx(SArray* pArray, FDelete fp);
 
+void taosArraySwap(SArray* a, SArray* b);
+
 /**
  * sort the array
  * @param pArray
@@ -252,14 +246,6 @@ void taosArraySortPWithExt(SArray* pArray, __ext_compar_fn_t fn, const void* par
 int32_t taosEncodeArray(void** buf, const SArray* pArray, FEncode encode);
 void*   taosDecodeArray(const void* buf, SArray** pArray, FDecode decode, int32_t dataSz);
 
-/**
- * swap array
- * @param a
- * @param b
- * @return
- */
-void taosArraySwap(SArray* a, SArray* b);
-
 #ifdef __cplusplus
 }
 #endif

include/util/tsimplehash.h

@@ -48,6 +48,13 @@ SSHashObj *tSimpleHashInit(size_t capacity, _hash_fn_t fn);
  */
 int32_t tSimpleHashGetSize(const SSHashObj *pHashObj);
 
+/**
+ * set the free function pointer
+ * @param pHashObj
+ * @param freeFp
+ */
+void tSimpleHashSetFreeFp(SSHashObj* pHashObj, _hash_free_fn_t freeFp);
+
 int32_t tSimpleHashPrint(const SSHashObj *pHashObj);
 
 /**

include/util/version.h

@@ -23,6 +23,7 @@ extern "C" {
 extern char version[];
 extern char compatible_version[];
 extern char gitinfo[];
+extern char gitinfoOfInternal[];
 extern char buildinfo[];
 
 #ifdef __cplusplus

packaging/rpm/tdengine.spec

@@ -90,45 +90,33 @@ cp %{_compiledir}/../include/libs/function/taosudf.h       %{buildroot}%{homepat
 cp -r %{_compiledir}/../examples/*                  %{buildroot}%{homepath}/examples
 
 if [ -f %{_compiledir}/build/bin/jemalloc-config ]; then
-    mkdir -p %{buildroot}%{userlocalpath}/bin
+    mkdir -p %{buildroot}%{homepath}/jemalloc/ ||:
-    mkdir -p %{buildroot}%{userlocalpath}/lib
+    mkdir -p %{buildroot}%{homepath}/jemalloc/include/jemalloc/ ||:
-    mkdir -p %{buildroot}%{userlocalpath}/lib/pkgconfig
+    mkdir -p %{buildroot}%{homepath}/jemalloc/lib/ ||:
-    mkdir -p %{buildroot}%{userlocalpath}/include
+    mkdir -p %{buildroot}%{homepath}/jemalloc/lib/pkgconfig ||:
-    mkdir -p %{buildroot}%{userlocalpath}/include/jemalloc
-    mkdir -p %{buildroot}%{userlocalpath}/share
-    mkdir -p %{buildroot}%{userlocalpath}/share/doc
-    mkdir -p %{buildroot}%{userlocalpath}/share/doc/jemalloc
-    mkdir -p %{buildroot}%{userlocalpath}/share/man
-    mkdir -p %{buildroot}%{userlocalpath}/share/man/man3
 
-    cp %{_compiledir}/build/bin/jemalloc-config %{buildroot}%{userlocalpath}/bin/
+    cp %{_compiledir}/build/bin/jemalloc-config %{buildroot}%{homepath}/jemalloc/bin
     if [ -f %{_compiledir}/build/bin/jemalloc.sh ]; then
-        cp %{_compiledir}/build/bin/jemalloc.sh %{buildroot}%{userlocalpath}/bin/
+        cp %{_compiledir}/build/bin/jemalloc.sh %{buildroot}%{homepath}/jemalloc/bin
     fi
     if [ -f %{_compiledir}/build/bin/jeprof ]; then
-        cp %{_compiledir}/build/bin/jeprof %{buildroot}%{userlocalpath}/bin/
+        cp %{_compiledir}/build/bin/jeprof %{buildroot}%{homepath}/jemalloc/bin
     fi
     if [ -f %{_compiledir}/build/include/jemalloc/jemalloc.h ]; then
-        cp %{_compiledir}/build/include/jemalloc/jemalloc.h %{buildroot}%{userlocalpath}/include/jemalloc/
+        cp %{_compiledir}/build/include/jemalloc/jemalloc.h %{buildroot}%{homepath}/jemalloc/include/jemalloc/
     fi
     if [ -f %{_compiledir}/build/lib/libjemalloc.so.2 ]; then
-        cp %{_compiledir}/build/lib/libjemalloc.so.2 %{buildroot}%{userlocalpath}/lib/
+        cp %{_compiledir}/build/lib/libjemalloc.so.2 %{buildroot}%{homepath}/jemalloc/lib
-        ln -sf libjemalloc.so.2 %{buildroot}%{userlocalpath}/lib/libjemalloc.so
+        ln -sf libjemalloc.so.2 %{buildroot}%{homepath}/jemalloc/lib/libjemalloc.so
     fi
     if [ -f %{_compiledir}/build/lib/libjemalloc.a ]; then
-        cp %{_compiledir}/build/lib/libjemalloc.a %{buildroot}%{userlocalpath}/lib/
+        cp %{_compiledir}/build/lib/libjemalloc.a %{buildroot}%{homepath}/jemalloc/lib
     fi
     if [ -f %{_compiledir}/build/lib/libjemalloc_pic.a ]; then
-        cp %{_compiledir}/build/lib/libjemalloc_pic.a %{buildroot}%{userlocalpath}/lib/
+        cp %{_compiledir}/build/lib/libjemalloc_pic.a %{buildroot}%{homepath}/jemalloc/lib
     fi
     if [ -f %{_compiledir}/build/lib/pkgconfig/jemalloc.pc ]; then
-        cp %{_compiledir}/build/lib/pkgconfig/jemalloc.pc %{buildroot}%{userlocalpath}/lib/pkgconfig/
+        cp %{_compiledir}/build/lib/pkgconfig/jemalloc.pc %{buildroot}%{homepath}/jemalloc/lib/pkgconfig
-    fi
-    if [ -f %{_compiledir}/build/share/doc/jemalloc/jemalloc.html ]; then
-        cp %{_compiledir}/build/share/doc/jemalloc/jemalloc.html %{buildroot}%{userlocalpath}/share/doc/jemalloc/
-    fi
-    if [ -f %{_compiledir}/build/share/man/man3/jemalloc.3 ]; then
-        cp %{_compiledir}/build/share/man/man3/jemalloc.3 %{buildroot}%{userlocalpath}/share/man/man3/
     fi
 fi
 

packaging/tools/makepkg.sh

@@ -51,9 +51,9 @@ fi
 
 if [ -d ${top_dir}/tools/taos-tools/packaging/deb ]; then
   cd ${top_dir}/tools/taos-tools/packaging/deb
-  [ -z "$taos_tools_ver" ] && taos_tools_ver="0.1.0"
+  
-
   taostools_ver=$(git for-each-ref --sort=taggerdate --format '%(tag)' refs/tags|grep -v taos | tail -1)
+  [ -z "$taos_tools_ver" ] && taos_tools_ver="0.1.0"
   taostools_install_dir="${release_dir}/${clientName2}Tools-${taostools_ver}"
 
   cd ${curr_dir}

packaging/tools/post.sh

@@ -145,6 +145,59 @@ function install_include() {
     log_print "install include success"
 }
 
+function install_jemalloc() {
+  jemalloc_dir=${script_dir}/../jemalloc
+
+  if [ -d ${jemalloc_dir} ]; then
+    ${csudo}/usr/bin/install -c -d /usr/local/bin
+
+    if [ -f ${jemalloc_dir}/bin/jemalloc-config ]; then
+      ${csudo}/usr/bin/install -c -m 755 ${jemalloc_dir}/bin/jemalloc-config /usr/local/bin
+    fi
+    if [ -f ${jemalloc_dir}/bin/jemalloc.sh ]; then
+      ${csudo}/usr/bin/install -c -m 755 ${jemalloc_dir}/bin/jemalloc.sh /usr/local/bin
+    fi
+    if [ -f ${jemalloc_dir}/bin/jeprof ]; then
+      ${csudo}/usr/bin/install -c -m 755 ${jemalloc_dir}/bin/jeprof /usr/local/bin
+    fi
+    if [ -f ${jemalloc_dir}/include/jemalloc/jemalloc.h ]; then
+      ${csudo}/usr/bin/install -c -d /usr/local/include/jemalloc
+      ${csudo}/usr/bin/install -c -m 644 ${jemalloc_dir}/include/jemalloc/jemalloc.h /usr/local/include/jemalloc
+    fi
+    if [ -f ${jemalloc_dir}/lib/libjemalloc.so.2 ]; then
+      ${csudo}/usr/bin/install -c -d /usr/local/lib
+      ${csudo}/usr/bin/install -c -m 755 ${jemalloc_dir}/lib/libjemalloc.so.2 /usr/local/lib
+      ${csudo}ln -sf libjemalloc.so.2 /usr/local/lib/libjemalloc.so
+      ${csudo}/usr/bin/install -c -d /usr/local/lib
+      if [ -f ${jemalloc_dir}/lib/libjemalloc.a ]; then
+        ${csudo}/usr/bin/install -c -m 755 ${jemalloc_dir}/lib/libjemalloc.a /usr/local/lib
+      fi
+      if [ -f ${jemalloc_dir}/lib/libjemalloc_pic.a ]; then
+        ${csudo}/usr/bin/install -c -m 755 ${jemalloc_dir}/lib/libjemalloc_pic.a /usr/local/lib
+      fi
+      if [ -f ${jemalloc_dir}/lib/libjemalloc_pic.a ]; then
+        ${csudo}/usr/bin/install -c -d /usr/local/lib/pkgconfig
+        ${csudo}/usr/bin/install -c -m 644 ${jemalloc_dir}/lib/pkgconfig/jemalloc.pc /usr/local/lib/pkgconfig
+      fi
+    fi
+    if [ -f ${jemalloc_dir}/share/doc/jemalloc/jemalloc.html ]; then
+      ${csudo}/usr/bin/install -c -d /usr/local/share/doc/jemalloc
+      ${csudo}/usr/bin/install -c -m 644 ${jemalloc_dir}/share/doc/jemalloc/jemalloc.html /usr/local/share/doc/jemalloc
+    fi
+    if [ -f ${jemalloc_dir}/share/man/man3/jemalloc.3 ]; then
+      ${csudo}/usr/bin/install -c -d /usr/local/share/man/man3
+      ${csudo}/usr/bin/install -c -m 644 ${jemalloc_dir}/share/man/man3/jemalloc.3 /usr/local/share/man/man3
+    fi
+
+    if [ -d /etc/ld.so.conf.d ]; then
+      echo "/usr/local/lib" | ${csudo}tee /etc/ld.so.conf.d/jemalloc.conf >/dev/null || echo -e "failed to write /etc/ld.so.conf.d/jemalloc.conf"
+      ${csudo}ldconfig
+    else
+      echo "/etc/ld.so.conf.d not found!"
+    fi
+  fi
+}
+
 function install_lib() {
     log_print "start install lib from ${lib_dir} to ${lib_link_dir}"
     ${csudo}rm -f ${lib_link_dir}/libtaos* || :
@@ -663,6 +716,7 @@ function install_TDengine() {
     # Install include, lib, binary and service
     install_include &&
     install_lib &&
+    install_jemalloc
     install_bin
 
     if [[ "$?" != 0 ]];then

source/client/src/clientMsgHandler.c

@@ -76,7 +76,7 @@ int32_t processConnectRsp(void* param, SDataBuf* pMsg, int32_t code) {
     goto End;
   }
 
-  if ((code = taosCheckVersionCompatibleFromStr(version, connectRsp.sVer, 2)) != 0) {
+  if ((code = taosCheckVersionCompatibleFromStr(version, connectRsp.sVer, 3)) != 0) {
     setErrno(pRequest, code);
     tsem_post(&pRequest->body.rspSem);
     goto End;
@@ -506,7 +506,7 @@ int32_t processShowVariablesRsp(void* param, SDataBuf* pMsg, int32_t code) {
       code = setQueryResultFromRsp(&pRequest->body.resInfo, pRes, false, true);
     }
 
-    if(code != 0){
+    if (code != 0) {
       taosMemoryFree(pRes);
     }
     tFreeSShowVariablesRsp(&rsp);

source/client/src/clientRawBlockWrite.c

@@ -183,7 +183,7 @@ static char* buildAlterSTableJson(void* alterData, int32_t alterDataLen) {
   }
   string = cJSON_PrintUnformatted(json);
 
-end:
+  end:
   cJSON_Delete(json);
   tFreeSMAltertbReq(&req);
   return string;
@@ -205,7 +205,7 @@ static char* processCreateStb(SMqMetaRsp* metaRsp) {
   }
   string = buildCreateTableJson(&req.schemaRow, &req.schemaTag, req.name, req.suid, TSDB_SUPER_TABLE);
   uDebug("processCreateStb %s", string);
-_err:
+  _err:
   tDecoderClear(&coder);
   return string;
 }
@@ -227,7 +227,7 @@ static char* processAlterStb(SMqMetaRsp* metaRsp) {
   string = buildAlterSTableJson(req.alterOriData, req.alterOriDataLen);
   uDebug("processAlterStb %s", string);
 
-_err:
+  _err:
   tDecoderClear(&coder);
   return string;
 }
@@ -309,7 +309,7 @@ static void buildChildElement(cJSON* json, SVCreateTbReq* pCreateReq) {
     cJSON_AddItemToArray(tags, tag);
   }
 
-end:
+  end:
   cJSON_AddItemToObject(json, "tags", tags);
   taosArrayDestroy(pTagVals);
 }
@@ -368,7 +368,7 @@ static char* processCreateTable(SMqMetaRsp* metaRsp) {
     uDebug("processCreateTable :%s", string);
   }
 
-_exit:
+  _exit:
   for (int32_t iReq = 0; iReq < req.nReqs; iReq++) {
     pCreateReq = req.pReqs + iReq;
     taosMemoryFreeClear(pCreateReq->comment);
@@ -408,7 +408,7 @@ static char* processAutoCreateTable(STaosxRsp* rsp) {
   }
   string = buildCreateCTableJson(pCreateReq, rsp->createTableNum);
   uDebug("processAutoCreateTable :%s", string);
-_exit:
+  _exit:
   for (int i = 0; i < rsp->createTableNum; i++) {
     tDecoderClear(&decoder[i]);
     taosMemoryFreeClear(pCreateReq[i].comment);
@@ -535,7 +535,7 @@ static char* processAlterTable(SMqMetaRsp* metaRsp) {
   string = cJSON_PrintUnformatted(json);
   uDebug("processAlterTable :%s", string);
 
-_exit:
+  _exit:
   cJSON_Delete(json);
   tDecoderClear(&decoder);
   return string;
@@ -569,7 +569,7 @@ static char* processDropSTable(SMqMetaRsp* metaRsp) {
 
   string = cJSON_PrintUnformatted(json);
   uDebug("processDropSTable :%s", string);
-_exit:
+  _exit:
   cJSON_Delete(json);
   tDecoderClear(&decoder);
   return string;
@@ -609,7 +609,7 @@ static char* processDeleteTable(SMqMetaRsp* metaRsp) {
 
   string = cJSON_PrintUnformatted(json);
   uDebug("processDeleteTable :%s", string);
-_exit:
+  _exit:
   cJSON_Delete(json);
   tDecoderClear(&coder);
   return string;
@@ -652,7 +652,7 @@ static char* processDropTable(SMqMetaRsp* metaRsp) {
 
   string = cJSON_PrintUnformatted(json);
   uDebug("processDropTable :%s", string);
-_exit:
+  _exit:
   cJSON_Delete(json);
   tDecoderClear(&decoder);
   return string;
@@ -742,7 +742,7 @@ static int32_t taosCreateStb(TAOS* taos, void* meta, int32_t metaLen) {
   code = pRequest->code;
   taosMemoryFree(pCmdMsg.pMsg);
 
-end:
+  end:
   destroyRequest(pRequest);
   tFreeSMCreateStbReq(&pReq);
   tDecoderClear(&coder);
@@ -839,7 +839,7 @@ static int32_t taosDropStb(TAOS* taos, void* meta, int32_t metaLen) {
   code = pRequest->code;
   taosMemoryFree(pCmdMsg.pMsg);
 
-end:
+  end:
   destroyRequest(pRequest);
   tDecoderClear(&coder);
   return code;
@@ -901,9 +901,9 @@ static int32_t taosCreateTable(TAOS* taos, void* meta, int32_t metaLen) {
   taosHashSetFreeFp(pVgroupHashmap, destroyCreateTbReqBatch);
 
   SRequestConnInfo conn = {.pTrans = pTscObj->pAppInfo->pTransporter,
-                           .requestId = pRequest->requestId,
+      .requestId = pRequest->requestId,
-                           .requestObjRefId = pRequest->self,
+      .requestObjRefId = pRequest->self,
-                           .mgmtEps = getEpSet_s(&pTscObj->pAppInfo->mgmtEp)};
+      .mgmtEps = getEpSet_s(&pTscObj->pAppInfo->mgmtEp)};
 
   pRequest->tableList = taosArrayInit(req.nReqs, sizeof(SName));
   // loop to create table
@@ -987,7 +987,7 @@ static int32_t taosCreateTable(TAOS* taos, void* meta, int32_t metaLen) {
 
   code = pRequest->code;
 
-end:
+  end:
   for (int32_t iReq = 0; iReq < req.nReqs; iReq++) {
     pCreateReq = req.pReqs + iReq;
     taosMemoryFreeClear(pCreateReq->comment);
@@ -1058,9 +1058,9 @@ static int32_t taosDropTable(TAOS* taos, void* meta, int32_t metaLen) {
   taosHashSetFreeFp(pVgroupHashmap, destroyDropTbReqBatch);
 
   SRequestConnInfo conn = {.pTrans = pTscObj->pAppInfo->pTransporter,
-                           .requestId = pRequest->requestId,
+      .requestId = pRequest->requestId,
-                           .requestObjRefId = pRequest->self,
+      .requestObjRefId = pRequest->self,
-                           .mgmtEps = getEpSet_s(&pTscObj->pAppInfo->mgmtEp)};
+      .mgmtEps = getEpSet_s(&pTscObj->pAppInfo->mgmtEp)};
   pRequest->tableList = taosArrayInit(req.nReqs, sizeof(SName));
   // loop to create table
   for (int32_t iReq = 0; iReq < req.nReqs; iReq++) {
@@ -1132,7 +1132,7 @@ static int32_t taosDropTable(TAOS* taos, void* meta, int32_t metaLen) {
   }
   code = pRequest->code;
 
-end:
+  end:
   taosHashCleanup(pVgroupHashmap);
   destroyRequest(pRequest);
   tDecoderClear(&coder);
@@ -1201,7 +1201,7 @@ static int32_t taosDeleteData(TAOS* taos, void* meta, int32_t metaLen) {
   }
   taos_free_result(res);
 
-end:
+  end:
   tDecoderClear(&coder);
   return code;
 }
@@ -1249,9 +1249,9 @@ static int32_t taosAlterTable(TAOS* taos, void* meta, int32_t metaLen) {
   }
 
   SRequestConnInfo conn = {.pTrans = pTscObj->pAppInfo->pTransporter,
-                           .requestId = pRequest->requestId,
+      .requestId = pRequest->requestId,
-                           .requestObjRefId = pRequest->self,
+      .requestObjRefId = pRequest->self,
-                           .mgmtEps = getEpSet_s(&pTscObj->pAppInfo->mgmtEp)};
+      .mgmtEps = getEpSet_s(&pTscObj->pAppInfo->mgmtEp)};
 
   SVgroupInfo pInfo = {0};
   SName       pName = {0};
@@ -1311,7 +1311,7 @@ static int32_t taosAlterTable(TAOS* taos, void* meta, int32_t metaLen) {
       code = handleAlterTbExecRes(pRes->res, pCatalog);
     }
   }
-end:
+  end:
   taosArrayDestroy(pArray);
   if (pVgData) taosMemoryFreeClear(pVgData->pData);
   taosMemoryFreeClear(pVgData);
@@ -1399,7 +1399,7 @@ int taos_write_raw_block_with_fields(TAOS* taos, int rows, char* pData, const ch
   launchQueryImpl(pRequest, pQuery, true, NULL);
   code = pRequest->code;
 
-end:
+  end:
   taosMemoryFreeClear(pTableMeta);
   qDestroyQuery(pQuery);
   destroyRequest(pRequest);
@@ -1481,7 +1481,7 @@ int taos_write_raw_block(TAOS* taos, int rows, char* pData, const char* tbname)
   launchQueryImpl(pRequest, pQuery, true, NULL);
   code = pRequest->code;
 
-end:
+  end:
   taosMemoryFreeClear(pTableMeta);
   qDestroyQuery(pQuery);
   destroyRequest(pRequest);
@@ -1601,6 +1601,7 @@ static int32_t tmqWriteRawDataImpl(TAOS* taos, void* data, int32_t dataLen) {
       uError("WriteRaw:rawBlockBindData failed");
       goto end;
     }
+    taosMemoryFreeClear(pTableMeta);
   }
 
   code = smlBuildOutput(pQuery, pVgHash);
@@ -1612,7 +1613,7 @@ static int32_t tmqWriteRawDataImpl(TAOS* taos, void* data, int32_t dataLen) {
   launchQueryImpl(pRequest, pQuery, true, NULL);
   code = pRequest->code;
 
-end:
+  end:
   tDeleteSMqDataRsp(&rspObj.rsp);
   tDecoderClear(&decoder);
   qDestroyQuery(pQuery);
@@ -1707,6 +1708,7 @@ static int32_t tmqWriteRawMetaDataImpl(TAOS* taos, void* data, int32_t dataLen)
       tDecoderInit(&decoderTmp, *dataTmp, *lenTmp);
       if (tDecodeSVCreateTbReq(&decoderTmp, &pCreateReq) < 0) {
         tDecoderClear(&decoderTmp);
+        tDestroySVCreateTbReq(&pCreateReq, TSDB_MSG_FLG_DECODE);
         uError("WriteRaw: tDecodeSVCreateTbReq error");
         code = TSDB_CODE_TMQ_INVALID_MSG;
         goto end;
@@ -1715,15 +1717,19 @@ static int32_t tmqWriteRawMetaDataImpl(TAOS* taos, void* data, int32_t dataLen)
       if (pCreateReq.type != TSDB_CHILD_TABLE) {
         uError("WriteRaw:pCreateReq.type != TSDB_CHILD_TABLE. table name: %s", tbName);
         code = TSDB_CODE_TSC_INVALID_VALUE;
+        tDecoderClear(&decoderTmp);
+        tDestroySVCreateTbReq(&pCreateReq, TSDB_MSG_FLG_DECODE);
         goto end;
       }
       if (strcmp(tbName, pCreateReq.name) == 0) {
         cloneSVreateTbReq(&pCreateReq, &pCreateReqDst);
 //        pCreateReqDst->ctb.suid = processSuid(pCreateReqDst->ctb.suid, pRequest->pDb);
         tDecoderClear(&decoderTmp);
+        tDestroySVCreateTbReq(&pCreateReq, TSDB_MSG_FLG_DECODE);
         break;
       }
       tDecoderClear(&decoderTmp);
+      tDestroySVCreateTbReq(&pCreateReq, TSDB_MSG_FLG_DECODE);
     }
 
     SVgroupInfo vg;
@@ -1774,6 +1780,7 @@ static int32_t tmqWriteRawMetaDataImpl(TAOS* taos, void* data, int32_t dataLen)
       goto end;
     }
     pCreateReqDst = NULL;
+    taosMemoryFreeClear(pTableMeta);
   }
 
   code = smlBuildOutput(pQuery, pVgHash);
@@ -1785,7 +1792,7 @@ static int32_t tmqWriteRawMetaDataImpl(TAOS* taos, void* data, int32_t dataLen)
   launchQueryImpl(pRequest, pQuery, true, NULL);
   code = pRequest->code;
 
-end:
+  end:
   tDeleteSTaosxRsp(&rspObj.rsp);
   tDecoderClear(&decoder);
   qDestroyQuery(pQuery);

source/client/src/clientSml.c

@@ -257,19 +257,14 @@ cleanup:
   kvVal->f = (float)result;
 
 #define SET_BIGINT                                                                                         \
-  if (smlDoubleToInt64OverFlow(result)) {                                                                  \
+  errno = 0;                                                                                             \
-    errno = 0;                                                                                             \
+  int64_t tmp = taosStr2Int64(pVal, &endptr, 10);                                                        \
-    int64_t tmp = taosStr2Int64(pVal, &endptr, 10);                                                        \
+  if (errno == ERANGE) {                                                                                 \
-    if (errno == ERANGE) {                                                                                 \
+    smlBuildInvalidDataMsg(msg, "big int out of range[-9223372036854775808,9223372036854775807]", pVal); \
-      smlBuildInvalidDataMsg(msg, "big int out of range[-9223372036854775808,9223372036854775807]", pVal); \
+    return false;                                                                                        \
-      return false;                                                                                        \
+  }                                                                                                      \
-    }                                                                                                      \
+  kvVal->type = TSDB_DATA_TYPE_BIGINT;                                                                   \
-    kvVal->type = TSDB_DATA_TYPE_BIGINT;                                                                   \
+  kvVal->i = tmp;
-    kvVal->i = tmp;                                                                                        \
-    return true;                                                                                           \
-  }                                                                                                        \
-  kvVal->type = TSDB_DATA_TYPE_BIGINT;                                                                     \
-  kvVal->i = (int64_t)result;
 
 #define SET_INT                                                                    \
   if (!IS_VALID_INT(result)) {                                                     \
@@ -288,19 +283,14 @@ cleanup:
   kvVal->i = result;
 
 #define SET_UBIGINT                                                                               \
-  if (result >= (double)UINT64_MAX || result < 0) {                                               \
+  errno = 0;                                                                                    \
-    errno = 0;                                                                                    \
+  uint64_t tmp = taosStr2UInt64(pVal, &endptr, 10);                                             \
-    uint64_t tmp = taosStr2UInt64(pVal, &endptr, 10);                                             \
+  if (errno == ERANGE || result < 0) {                                                          \
-    if (errno == ERANGE || result < 0) {                                                          \
+    smlBuildInvalidDataMsg(msg, "unsigned big int out of range[0,18446744073709551615]", pVal); \
-      smlBuildInvalidDataMsg(msg, "unsigned big int out of range[0,18446744073709551615]", pVal); \
+    return false;                                                                               \
-      return false;                                                                               \
+  }                                                                                             \
-    }                                                                                             \
+  kvVal->type = TSDB_DATA_TYPE_UBIGINT;                                                         \
-    kvVal->type = TSDB_DATA_TYPE_UBIGINT;                                                         \
+  kvVal->u = tmp;
-    kvVal->u = tmp;                                                                               \
-    return true;                                                                                  \
-  }                                                                                               \
-  kvVal->type = TSDB_DATA_TYPE_UBIGINT;                                                           \
-  kvVal->u = result;
 
 #define SET_UINT                                                                  \
   if (!IS_VALID_UINT(result)) {                                                   \

source/client/src/clientTmq.c

@@ -320,15 +320,16 @@ tmq_conf_res_t tmq_conf_set(tmq_conf_t* conf, const char* key, const char* value
   }
 
   if (strcasecmp(key, "enable.heartbeat.background") == 0) {
-    if (strcasecmp(value, "true") == 0) {
+//    if (strcasecmp(value, "true") == 0) {
-      conf->hbBgEnable = true;
+//      conf->hbBgEnable = true;
-      return TMQ_CONF_OK;
+//      return TMQ_CONF_OK;
-    } else if (strcasecmp(value, "false") == 0) {
+//    } else if (strcasecmp(value, "false") == 0) {
-      conf->hbBgEnable = false;
+//      conf->hbBgEnable = false;
-      return TMQ_CONF_OK;
+//      return TMQ_CONF_OK;
-    } else {
+//    } else {
+      tscError("the default value of enable.heartbeat.background is true, can not be seted");
       return TMQ_CONF_INVALID;
-    }
+//    }
   }
 
   if (strcasecmp(key, "td.connect.ip") == 0) {
@@ -535,10 +536,14 @@ static int32_t doSendCommitMsg(tmq_t* tmq, SMqClientVg* pVg, const char* pTopicN
   atomic_add_fetch_32(&pParamSet->totalRspNum, 1);
 
   SEp* pEp = GET_ACTIVE_EP(&pVg->epSet);
-  tscDebug("consumer:0x%" PRIx64 " topic:%s on vgId:%d send offset:%" PRId64 " prev:%" PRId64
+  char offsetBuf[80] = {0};
-           ", ep:%s:%d, ordinal:%d/%d, req:0x%" PRIx64,
+  tFormatOffset(offsetBuf, tListLen(offsetBuf), &pOffset->val);
-           tmq->consumerId, pOffset->subKey, pVg->vgId, pOffset->val.version, pVg->committedOffset.version, pEp->fqdn,
+
-           pEp->port, index + 1, totalVgroups, pMsgSendInfo->requestId);
+  char commitBuf[80] = {0};
+  tFormatOffset(commitBuf, tListLen(commitBuf), &pVg->committedOffset);
+  tscDebug("consumer:0x%" PRIx64 " topic:%s on vgId:%d send offset:%s prev:%s, ep:%s:%d, ordinal:%d/%d, req:0x%" PRIx64,
+           tmq->consumerId, pOffset->subKey, pVg->vgId, offsetBuf, commitBuf, pEp->fqdn, pEp->port, index + 1,
+           totalVgroups, pMsgSendInfo->requestId);
 
   int64_t transporterId = 0;
   asyncSendMsgToServer(tmq->pTscObj->pAppInfo->pTransporter, &pVg->epSet, &transporterId, pMsgSendInfo);

source/client/test/clientTests.cpp

@@ -998,7 +998,8 @@ TEST(clientCase, sub_db_test) {
 
   // 创建订阅 topics 列表
   tmq_list_t* topicList = tmq_list_new();
-  tmq_list_append(topicList, "topic_db1");
+  tmq_list_append(topicList, "topic_t1");
+  tmq_list_append(topicList, "topic_s2");
 
   // 启动订阅
   tmq_subscribe(tmq, topicList);

source/common/src/tdatablock.c

@@ -147,9 +147,17 @@ int32_t colDataReserve(SColumnInfoData* pColumnInfoData, size_t newSize) {
   return TSDB_CODE_SUCCESS;
 }
 
-static void doCopyNItems(struct SColumnInfoData* pColumnInfoData, int32_t currentRow, const char* pData,
+static int32_t doCopyNItems(struct SColumnInfoData* pColumnInfoData, int32_t currentRow, const char* pData,
-                         int32_t itemLen, int32_t numOfRows) {
+                         int32_t itemLen, int32_t numOfRows, bool trimValue) {
-  ASSERT(pColumnInfoData->info.bytes >= itemLen);
+  if (pColumnInfoData->info.bytes < itemLen) {
+    uWarn("column/tag actual data len %d is bigger than schema len %d, trim it:%d", itemLen, pColumnInfoData->info.bytes, trimValue);
+    if (trimValue) {
+      itemLen = pColumnInfoData->info.bytes;
+    } else {
+      return TSDB_CODE_TDB_INVALID_TABLE_SCHEMA_VER;
+    }
+  }
+  
   size_t start = 1;
 
   // the first item
@@ -178,10 +186,12 @@ static void doCopyNItems(struct SColumnInfoData* pColumnInfoData, int32_t curren
 
     pColumnInfoData->varmeta.length += numOfRows * itemLen;
   }
+
+  return TSDB_CODE_SUCCESS;
 }
 
 int32_t colDataSetNItems(SColumnInfoData* pColumnInfoData, uint32_t currentRow, const char* pData,
-                            uint32_t numOfRows) {
+                            uint32_t numOfRows, bool trimValue) {
   int32_t len = pColumnInfoData->info.bytes;
   if (IS_VAR_DATA_TYPE(pColumnInfoData->info.type)) {
     len = varDataTLen(pData);
@@ -193,8 +203,7 @@ int32_t colDataSetNItems(SColumnInfoData* pColumnInfoData, uint32_t currentRow,
     }
   }
 
-  doCopyNItems(pColumnInfoData, currentRow, pData, len, numOfRows);
+  return doCopyNItems(pColumnInfoData, currentRow, pData, len, numOfRows, trimValue);
-  return TSDB_CODE_SUCCESS;
 }
 
 static void doBitmapMerge(SColumnInfoData* pColumnInfoData, int32_t numOfRow1, const SColumnInfoData* pSource,
@@ -973,7 +982,7 @@ int32_t blockDataSort(SSDataBlock* pDataBlock, SArray* pOrderInfo) {
         taosSort(pColInfoData->pData, pDataBlock->info.rows, pColInfoData->info.bytes, fn);
 
         int64_t p1 = taosGetTimestampUs();
-        uDebug("blockDataSort easy cost:%" PRId64 ", rows:%d\n", p1 - p0, pDataBlock->info.rows);
+        uDebug("blockDataSort easy cost:%" PRId64 ", rows:%" PRId64 "\n", p1 - p0, pDataBlock->info.rows);
 
         return TSDB_CODE_SUCCESS;
       } else {  // var data type
@@ -1180,7 +1189,7 @@ void blockDataCleanup(SSDataBlock* pDataBlock) {
 
 void blockDataEmpty(SSDataBlock* pDataBlock) {
   SDataBlockInfo* pInfo = &pDataBlock->info;
-  if (pInfo->capacity == 0 || pInfo->rows > pDataBlock->info.capacity) {
+  if (pInfo->capacity == 0) {
     return;
   }
 
@@ -1739,14 +1748,14 @@ int32_t tEncodeDataBlock(void** buf, const SSDataBlock* pBlock) {
   int64_t tbUid = pBlock->info.id.uid;
   int16_t numOfCols = taosArrayGetSize(pBlock->pDataBlock);
   int16_t hasVarCol = pBlock->info.hasVarCol;
-  int32_t rows = pBlock->info.rows;
+  int64_t rows = pBlock->info.rows;
   int32_t sz = taosArrayGetSize(pBlock->pDataBlock);
 
   int32_t tlen = 0;
   tlen += taosEncodeFixedI64(buf, tbUid);
   tlen += taosEncodeFixedI16(buf, numOfCols);
   tlen += taosEncodeFixedI16(buf, hasVarCol);
-  tlen += taosEncodeFixedI32(buf, rows);
+  tlen += taosEncodeFixedI64(buf, rows);
   tlen += taosEncodeFixedI32(buf, sz);
   for (int32_t i = 0; i < sz; i++) {
     SColumnInfoData* pColData = (SColumnInfoData*)taosArrayGet(pBlock->pDataBlock, i);
@@ -1777,7 +1786,7 @@ void* tDecodeDataBlock(const void* buf, SSDataBlock* pBlock) {
   buf = taosDecodeFixedU64(buf, &pBlock->info.id.uid);
   buf = taosDecodeFixedI16(buf, &numOfCols);
   buf = taosDecodeFixedI16(buf, &pBlock->info.hasVarCol);
-  buf = taosDecodeFixedI32(buf, &pBlock->info.rows);
+  buf = taosDecodeFixedI64(buf, &pBlock->info.rows);
   buf = taosDecodeFixedI32(buf, &sz);
   pBlock->pDataBlock = taosArrayInit(sz, sizeof(SColumnInfoData));
   for (int32_t i = 0; i < sz; i++) {
@@ -1981,7 +1990,7 @@ char* dumpBlockData(SSDataBlock* pDataBlock, const char* flag, char** pDataBuf)
   int32_t len = 0;
   len += snprintf(dumpBuf + len, size - len,
                   "===stream===%s|block type %d|child id %d|group id:%" PRIu64 "|uid:%" PRId64
-                  "|rows:%d|version:%" PRIu64 "|cal start:%" PRIu64 "|cal end:%" PRIu64 "|tbl:%s\n",
+                  "|rows:%" PRId64 "|version:%" PRIu64 "|cal start:%" PRIu64 "|cal end:%" PRIu64 "|tbl:%s\n",
                   flag, (int32_t)pDataBlock->info.type, pDataBlock->info.childId, pDataBlock->info.id.groupId,
                   pDataBlock->info.id.uid, pDataBlock->info.rows, pDataBlock->info.version,
                   pDataBlock->info.calWin.skey, pDataBlock->info.calWin.ekey, pDataBlock->info.parTbName);

source/common/src/tdataformat.c

@@ -2439,6 +2439,12 @@ _exit:
 int32_t tColDataAddValueByDataBlock(SColData *pColData, int8_t type, int32_t bytes, int32_t nRows, char *lengthOrbitmap,
                                     char *data) {
   int32_t code = 0;
+  if(data == NULL){
+    for (int32_t i = 0; i < nRows; ++i) {
+      code = tColDataAppendValueImpl[pColData->flag][CV_FLAG_NONE](pColData, NULL, 0);
+    }
+    goto _exit;
+  }
 
   if (IS_VAR_DATA_TYPE(type)) {  // var-length data type
     for (int32_t i = 0; i < nRows; ++i) {
@@ -2447,6 +2453,11 @@ int32_t tColDataAddValueByDataBlock(SColData *pColData, int8_t type, int32_t byt
         code = tColDataAppendValueImpl[pColData->flag][CV_FLAG_NULL](pColData, NULL, 0);
         if (code) goto _exit;
       } else {
+        if(ASSERT(varDataTLen(data + offset) <= bytes)){
+          uError("var data length invalid, varDataTLen(data + offset):%d <= bytes:%d", (int)varDataTLen(data + offset), bytes);
+          code = TSDB_CODE_INVALID_PARA;
+          goto _exit;
+        }
         code = tColDataAppendValueImpl[pColData->flag][CV_FLAG_VALUE](pColData, (uint8_t *)varDataVal(data + offset),
                                                                       varDataLen(data + offset));
       }

source/dnode/mgmt/exe/dmMain.c

@@ -18,6 +18,7 @@
 #include "mnode.h"
 #include "tconfig.h"
 #include "tglobal.h"
+#include "version.h"
 
 // clang-format off
 #define DM_APOLLO_URL    "The apollo string to use when configuring the server, such as: -a 'jsonFile:./tests/cfg.json', cfg.json text can be '{\"fqdn\":\"td1\"}'."
@@ -76,28 +77,28 @@ void dmLogCrash(int signum, void *sigInfo, void *context) {
   taosIgnSignal(SIGINT);
   taosIgnSignal(SIGBREAK);
 
-#ifndef WINDOWS  
+#ifndef WINDOWS
   taosIgnSignal(SIGBUS);
 #endif
   taosIgnSignal(SIGABRT);
   taosIgnSignal(SIGFPE);
   taosIgnSignal(SIGSEGV);
 
-  char *pMsg = NULL;
+  char       *pMsg = NULL;
   const char *flags = "UTL FATAL ";
   ELogLevel   level = DEBUG_FATAL;
   int32_t     dflag = 255;
-  int64_t     msgLen= -1;
+  int64_t     msgLen = -1;
-  
+
   if (tsEnableCrashReport) {
     if (taosGenCrashJsonMsg(signum, &pMsg, dmGetClusterId(), global.startTime)) {
       taosPrintLog(flags, level, dflag, "failed to generate crash json msg");
       goto _return;
     } else {
-      msgLen = strlen(pMsg);  
+      msgLen = strlen(pMsg);
     }
   }
-  
+
 _return:
 
   taosLogCrashInfo("taosd", pMsg, msgLen, signum, sigInfo);
@@ -123,7 +124,7 @@ static void dmSetSignalHandle() {
 
 #ifndef WINDOWS
   taosSetSignal(SIGBUS, dmLogCrash);
-#endif  
+#endif
   taosSetSignal(SIGABRT, dmLogCrash);
   taosSetSignal(SIGFPE, dmLogCrash);
   taosSetSignal(SIGSEGV, dmLogCrash);
@@ -134,7 +135,7 @@ static int32_t dmParseArgs(int32_t argc, char const *argv[]) {
 
   int32_t cmdEnvIndex = 0;
   if (argc < 2) return 0;
-  
+
   global.envCmd = taosMemoryMalloc((argc - 1) * sizeof(char *));
   memset(global.envCmd, 0, (argc - 1) * sizeof(char *));
   for (int32_t i = 1; i < argc; ++i) {
@@ -203,6 +204,9 @@ static void dmPrintVersion() {
 #endif
   printf("%s version: %s compatible_version: %s\n", releaseName, version, compatible_version);
   printf("gitinfo: %s\n", gitinfo);
+#ifdef TD_ENTERPRISE
+  printf("gitinfoOfInternal: %s\n", gitinfoOfInternal);
+#endif
   printf("buildInfo: %s\n", buildinfo);
 }
 
@@ -284,7 +288,7 @@ int mainWindows(int argc, char **argv) {
       printf("failed to init memory dbg, error:%s\n", tstrerror(code));
       return code;
     }
-    tsAsyncLog = false;    
+    tsAsyncLog = false;
     printf("memory dbg enabled\n");
   }
 #endif

source/dnode/mgmt/mgmt_mnode/inc/mmInt.h

@@ -34,7 +34,7 @@ typedef struct SMnodeMgmt {
   SSingleWorker  readWorker;
   SSingleWorker  writeWorker;
   SSingleWorker  syncWorker;
-  SSingleWorker  syncCtrlWorker;
+  SSingleWorker  syncRdWorker;
   bool           stopped;
   int32_t        refCount;
   TdThreadRwlock lock;
@@ -54,7 +54,7 @@ int32_t mmStartWorker(SMnodeMgmt *pMgmt);
 void    mmStopWorker(SMnodeMgmt *pMgmt);
 int32_t mmPutMsgToWriteQueue(SMnodeMgmt *pMgmt, SRpcMsg *pMsg);
 int32_t mmPutMsgToSyncQueue(SMnodeMgmt *pMgmt, SRpcMsg *pMsg);
-int32_t mmPutMsgToSyncCtrlQueue(SMnodeMgmt *pMgmt, SRpcMsg *pMsg);
+int32_t mmPutMsgToSyncRdQueue(SMnodeMgmt *pMgmt, SRpcMsg *pMsg);
 int32_t mmPutMsgToReadQueue(SMnodeMgmt *pMgmt, SRpcMsg *pMsg);
 int32_t mmPutMsgToQueryQueue(SMnodeMgmt *pMgmt, SRpcMsg *pMsg);
 int32_t mmPutMsgToFetchQueue(SMnodeMgmt *pMgmt, SRpcMsg *pMsg);

source/dnode/mgmt/mgmt_mnode/src/mmHandle.c

@@ -188,7 +188,8 @@ SArray *mmGetMsgHandles() {
   if (dmSetMgmtHandle(pArray, TDMT_VND_DROP_INDEX_RSP, mmPutMsgToWriteQueue, 0) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_VND_DISABLE_WRITE_RSP, mmPutMsgToWriteQueue, 0) == NULL) goto _OVER;
 
-  if (dmSetMgmtHandle(pArray, TDMT_SYNC_TIMEOUT, mmPutMsgToSyncQueue, 1) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_TIMEOUT_ELECTION, mmPutMsgToSyncQueue, 1) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_HEARTBEAT, mmPutMsgToSyncRdQueue, 1) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_CLIENT_REQUEST, mmPutMsgToSyncQueue, 1) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_CLIENT_REQUEST_BATCH, mmPutMsgToSyncQueue, 1) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_CLIENT_REQUEST_REPLY, mmPutMsgToSyncQueue, 1) == NULL) goto _OVER;
@@ -198,11 +199,12 @@ SArray *mmGetMsgHandles() {
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_APPEND_ENTRIES_BATCH, mmPutMsgToSyncQueue, 1) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_APPEND_ENTRIES_REPLY, mmPutMsgToSyncQueue, 1) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_SNAPSHOT_SEND, mmPutMsgToSyncQueue, 1) == NULL) goto _OVER;
-  if (dmSetMgmtHandle(pArray, TDMT_SYNC_SNAPSHOT_RSP, mmPutMsgToSyncQueue, 1) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_PRE_SNAPSHOT, mmPutMsgToSyncQueue, 1) == NULL) goto _OVER;
-  if (dmSetMgmtHandle(pArray, TDMT_SYNC_PRE_SNAPSHOT_REPLY, mmPutMsgToSyncQueue, 1) == NULL) goto _OVER;
+
-  if (dmSetMgmtHandle(pArray, TDMT_SYNC_HEARTBEAT, mmPutMsgToSyncCtrlQueue, 1) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_TIMEOUT, mmPutMsgToSyncRdQueue, 1) == NULL) goto _OVER;
-  if (dmSetMgmtHandle(pArray, TDMT_SYNC_HEARTBEAT_REPLY, mmPutMsgToSyncCtrlQueue, 1) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_HEARTBEAT_REPLY, mmPutMsgToSyncRdQueue, 1) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_SNAPSHOT_RSP, mmPutMsgToSyncRdQueue, 1) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_PRE_SNAPSHOT_REPLY, mmPutMsgToSyncRdQueue, 1) == NULL) goto _OVER;
 
   code = 0;
 

source/dnode/mgmt/mgmt_mnode/src/mmWorker.c

@@ -111,8 +111,8 @@ int32_t mmPutMsgToSyncQueue(SMnodeMgmt *pMgmt, SRpcMsg *pMsg) {
   return mmPutMsgToWorker(pMgmt, &pMgmt->syncWorker, pMsg);
 }
 
-int32_t mmPutMsgToSyncCtrlQueue(SMnodeMgmt *pMgmt, SRpcMsg *pMsg) {
+int32_t mmPutMsgToSyncRdQueue(SMnodeMgmt *pMgmt, SRpcMsg *pMsg) {
-  return mmPutMsgToWorker(pMgmt, &pMgmt->syncCtrlWorker, pMsg);
+  return mmPutMsgToWorker(pMgmt, &pMgmt->syncRdWorker, pMsg);
 }
 
 int32_t mmPutMsgToReadQueue(SMnodeMgmt *pMgmt, SRpcMsg *pMsg) {
@@ -151,8 +151,8 @@ int32_t mmPutMsgToQueue(SMnodeMgmt *pMgmt, EQueueType qtype, SRpcMsg *pRpc) {
     case SYNC_QUEUE:
       pWorker = &pMgmt->syncWorker;
       break;
-    case SYNC_CTRL_QUEUE:
+    case SYNC_RD_QUEUE:
-      pWorker = &pMgmt->syncCtrlWorker;
+      pWorker = &pMgmt->syncRdWorker;
       break;
     default:
       terrno = TSDB_CODE_INVALID_PARA;
@@ -238,12 +238,12 @@ int32_t mmStartWorker(SMnodeMgmt *pMgmt) {
   SSingleWorkerCfg scCfg = {
       .min = 1,
       .max = 1,
-      .name = "mnode-sync-ctrl",
+      .name = "mnode-sync-rd",
       .fp = (FItem)mmProcessSyncMsg,
       .param = pMgmt,
   };
-  if (tSingleWorkerInit(&pMgmt->syncCtrlWorker, &scCfg) != 0) {
+  if (tSingleWorkerInit(&pMgmt->syncRdWorker, &scCfg) != 0) {
-    dError("failed to start mnode mnode-sync-ctrl worker since %s", terrstr());
+    dError("failed to start mnode mnode-sync-rd worker since %s", terrstr());
     return -1;
   }
 
@@ -259,6 +259,6 @@ void mmStopWorker(SMnodeMgmt *pMgmt) {
   tSingleWorkerCleanup(&pMgmt->readWorker);
   tSingleWorkerCleanup(&pMgmt->writeWorker);
   tSingleWorkerCleanup(&pMgmt->syncWorker);
-  tSingleWorkerCleanup(&pMgmt->syncCtrlWorker);
+  tSingleWorkerCleanup(&pMgmt->syncRdWorker);
   dDebug("mnode workers are closed");
 }

source/dnode/mgmt/mgmt_vnode/inc/vmInt.h

@@ -59,7 +59,7 @@ typedef struct {
   SVnode       *pImpl;
   SMultiWorker  pWriteW;
   SMultiWorker  pSyncW;
-  SMultiWorker  pSyncCtrlW;
+  SMultiWorker  pSyncRdW;
   SMultiWorker  pApplyW;
   STaosQueue   *pQueryQ;
   STaosQueue   *pStreamQ;
@@ -107,7 +107,7 @@ int32_t vmPutRpcMsgToQueue(SVnodeMgmt *pMgmt, EQueueType qtype, SRpcMsg *pRpc);
 
 int32_t vmPutMsgToWriteQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg);
 int32_t vmPutMsgToSyncQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg);
-int32_t vmPutMsgToSyncCtrlQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg);
+int32_t vmPutMsgToSyncRdQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg);
 int32_t vmPutMsgToQueryQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg);
 int32_t vmPutMsgToFetchQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg);
 int32_t vmPutMsgToStreamQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg);

source/dnode/mgmt/mgmt_vnode/src/vmHandle.c

@@ -549,7 +549,8 @@ SArray *vmGetMsgHandles() {
   if (dmSetMgmtHandle(pArray, TDMT_DND_CREATE_VNODE, vmPutMsgToMgmtQueue, 0) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_DND_DROP_VNODE, vmPutMsgToMgmtQueue, 0) == NULL) goto _OVER;
 
-  if (dmSetMgmtHandle(pArray, TDMT_SYNC_TIMEOUT, vmPutMsgToSyncQueue, 0) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_TIMEOUT_ELECTION, vmPutMsgToSyncQueue, 0) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_HEARTBEAT, vmPutMsgToSyncRdQueue, 0) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_CLIENT_REQUEST, vmPutMsgToSyncQueue, 0) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_CLIENT_REQUEST_BATCH, vmPutMsgToSyncQueue, 0) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_CLIENT_REQUEST_REPLY, vmPutMsgToSyncQueue, 0) == NULL) goto _OVER;
@@ -559,12 +560,12 @@ SArray *vmGetMsgHandles() {
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_APPEND_ENTRIES_BATCH, vmPutMsgToSyncQueue, 0) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_APPEND_ENTRIES_REPLY, vmPutMsgToSyncQueue, 0) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_SNAPSHOT_SEND, vmPutMsgToSyncQueue, 0) == NULL) goto _OVER;
-  if (dmSetMgmtHandle(pArray, TDMT_SYNC_SNAPSHOT_RSP, vmPutMsgToSyncQueue, 0) == NULL) goto _OVER;
   if (dmSetMgmtHandle(pArray, TDMT_SYNC_PRE_SNAPSHOT, vmPutMsgToSyncQueue, 0) == NULL) goto _OVER;
-  if (dmSetMgmtHandle(pArray, TDMT_SYNC_PRE_SNAPSHOT_REPLY, vmPutMsgToSyncQueue, 0) == NULL) goto _OVER;
 
-  if (dmSetMgmtHandle(pArray, TDMT_SYNC_HEARTBEAT, vmPutMsgToSyncCtrlQueue, 0) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_TIMEOUT, vmPutMsgToSyncRdQueue, 0) == NULL) goto _OVER;
-  if (dmSetMgmtHandle(pArray, TDMT_SYNC_HEARTBEAT_REPLY, vmPutMsgToSyncCtrlQueue, 0) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_HEARTBEAT_REPLY, vmPutMsgToSyncRdQueue, 0) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_SNAPSHOT_RSP, vmPutMsgToSyncRdQueue, 0) == NULL) goto _OVER;
+  if (dmSetMgmtHandle(pArray, TDMT_SYNC_PRE_SNAPSHOT_REPLY, vmPutMsgToSyncRdQueue, 0) == NULL) goto _OVER;
 
   code = 0;
 

source/dnode/mgmt/mgmt_vnode/src/vmInt.c

@@ -98,9 +98,9 @@ void vmCloseVnode(SVnodeMgmt *pMgmt, SVnodeObj *pVnode, bool commitAndRemoveWal)
         pVnode->pSyncW.queue->threadId);
   tMultiWorkerCleanup(&pVnode->pSyncW);
 
-  dInfo("vgId:%d, wait for vnode sync ctrl queue:%p is empty, thread:%08" PRId64, pVnode->vgId,
+  dInfo("vgId:%d, wait for vnode sync rd queue:%p is empty, thread:%08" PRId64, pVnode->vgId, pVnode->pSyncRdW.queue,
-        pVnode->pSyncCtrlW.queue, pVnode->pSyncCtrlW.queue->threadId);
+        pVnode->pSyncRdW.queue->threadId);
-  tMultiWorkerCleanup(&pVnode->pSyncCtrlW);
+  tMultiWorkerCleanup(&pVnode->pSyncRdW);
 
   dInfo("vgId:%d, wait for vnode apply queue:%p is empty, thread:%08" PRId64, pVnode->vgId, pVnode->pApplyW.queue,
         pVnode->pApplyW.queue->threadId);

source/dnode/mgmt/mgmt_vnode/src/vmWorker.c

@@ -216,9 +216,9 @@ static int32_t vmPutMsgToQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg, EQueueType qtyp
       dGTrace("vgId:%d, msg:%p put into vnode-sync queue", pVnode->vgId, pMsg);
       taosWriteQitem(pVnode->pSyncW.queue, pMsg);
       break;
-    case SYNC_CTRL_QUEUE:
+    case SYNC_RD_QUEUE:
-      dGTrace("vgId:%d, msg:%p put into vnode-sync-ctrl queue", pVnode->vgId, pMsg);
+      dGTrace("vgId:%d, msg:%p put into vnode-sync-rd queue", pVnode->vgId, pMsg);
-      taosWriteQitem(pVnode->pSyncCtrlW.queue, pMsg);
+      taosWriteQitem(pVnode->pSyncRdW.queue, pMsg);
       break;
     case APPLY_QUEUE:
       dGTrace("vgId:%d, msg:%p put into vnode-apply queue", pVnode->vgId, pMsg);
@@ -234,9 +234,7 @@ static int32_t vmPutMsgToQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg, EQueueType qtyp
   return code;
 }
 
-int32_t vmPutMsgToSyncCtrlQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg) {
+int32_t vmPutMsgToSyncRdQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg) { return vmPutMsgToQueue(pMgmt, pMsg, SYNC_RD_QUEUE); }
-  return vmPutMsgToQueue(pMgmt, pMsg, SYNC_CTRL_QUEUE);
-}
 
 int32_t vmPutMsgToSyncQueue(SVnodeMgmt *pMgmt, SRpcMsg *pMsg) { return vmPutMsgToQueue(pMgmt, pMsg, SYNC_QUEUE); }
 
@@ -327,18 +325,18 @@ int32_t vmGetQueueSize(SVnodeMgmt *pMgmt, int32_t vgId, EQueueType qtype) {
 int32_t vmAllocQueue(SVnodeMgmt *pMgmt, SVnodeObj *pVnode) {
   SMultiWorkerCfg wcfg = {.max = 1, .name = "vnode-write", .fp = (FItems)vnodeProposeWriteMsg, .param = pVnode->pImpl};
   SMultiWorkerCfg scfg = {.max = 1, .name = "vnode-sync", .fp = (FItems)vmProcessSyncQueue, .param = pVnode};
-  SMultiWorkerCfg sccfg = {.max = 1, .name = "vnode-sync-ctrl", .fp = (FItems)vmProcessSyncQueue, .param = pVnode};
+  SMultiWorkerCfg sccfg = {.max = 1, .name = "vnode-sync-rd", .fp = (FItems)vmProcessSyncQueue, .param = pVnode};
   SMultiWorkerCfg acfg = {.max = 1, .name = "vnode-apply", .fp = (FItems)vnodeApplyWriteMsg, .param = pVnode->pImpl};
   (void)tMultiWorkerInit(&pVnode->pWriteW, &wcfg);
   (void)tMultiWorkerInit(&pVnode->pSyncW, &scfg);
-  (void)tMultiWorkerInit(&pVnode->pSyncCtrlW, &sccfg);
+  (void)tMultiWorkerInit(&pVnode->pSyncRdW, &sccfg);
   (void)tMultiWorkerInit(&pVnode->pApplyW, &acfg);
 
   pVnode->pQueryQ = tQWorkerAllocQueue(&pMgmt->queryPool, pVnode, (FItem)vmProcessQueryQueue);
   pVnode->pStreamQ = tAutoQWorkerAllocQueue(&pMgmt->streamPool, pVnode, (FItem)vmProcessStreamQueue);
   pVnode->pFetchQ = tWWorkerAllocQueue(&pMgmt->fetchPool, pVnode, (FItems)vmProcessFetchQueue);
 
-  if (pVnode->pWriteW.queue == NULL || pVnode->pSyncW.queue == NULL || pVnode->pSyncCtrlW.queue == NULL ||
+  if (pVnode->pWriteW.queue == NULL || pVnode->pSyncW.queue == NULL || pVnode->pSyncRdW.queue == NULL ||
       pVnode->pApplyW.queue == NULL || pVnode->pQueryQ == NULL || pVnode->pStreamQ == NULL || pVnode->pFetchQ == NULL) {
     terrno = TSDB_CODE_OUT_OF_MEMORY;
     return -1;
@@ -348,8 +346,8 @@ int32_t vmAllocQueue(SVnodeMgmt *pMgmt, SVnodeObj *pVnode) {
         pVnode->pWriteW.queue->threadId);
   dInfo("vgId:%d, sync-queue:%p is alloced, thread:%08" PRId64, pVnode->vgId, pVnode->pSyncW.queue,
         pVnode->pSyncW.queue->threadId);
-  dInfo("vgId:%d, sync-ctrl-queue:%p is alloced, thread:%08" PRId64, pVnode->vgId, pVnode->pSyncCtrlW.queue,
+  dInfo("vgId:%d, sync-rd-queue:%p is alloced, thread:%08" PRId64, pVnode->vgId, pVnode->pSyncRdW.queue,
-        pVnode->pSyncCtrlW.queue->threadId);
+        pVnode->pSyncRdW.queue->threadId);
   dInfo("vgId:%d, apply-queue:%p is alloced, thread:%08" PRId64, pVnode->vgId, pVnode->pApplyW.queue,
         pVnode->pApplyW.queue->threadId);
   dInfo("vgId:%d, query-queue:%p is alloced", pVnode->vgId, pVnode->pQueryQ);

source/dnode/mnode/impl/src/mndConsumer.c

@@ -247,6 +247,13 @@ static SMqRebInfo *mndGetOrCreateRebSub(SHashObj *pHash, const char *key) {
   return pRebInfo;
 }
 
+static void freeRebalanceItem(void* param) {
+  SMqRebInfo* pInfo = param;
+  taosArrayDestroy(pInfo->lostConsumers);
+  taosArrayDestroy(pInfo->newConsumers);
+  taosArrayDestroy(pInfo->removedConsumers);
+}
+
 static int32_t mndProcessMqTimerMsg(SRpcMsg *pMsg) {
   SMnode         *pMnode = pMsg->info.node;
   SSdb           *pSdb = pMnode->pSdb;
@@ -262,8 +269,21 @@ static int32_t mndProcessMqTimerMsg(SRpcMsg *pMsg) {
   }
 
   SMqDoRebalanceMsg *pRebMsg = rpcMallocCont(sizeof(SMqDoRebalanceMsg));
+  if (pRebMsg == NULL) {
+    mError("failed to create the rebalance msg, size:%d, quit mq timer", (int32_t) sizeof(SMqDoRebalanceMsg));
+    mndRebEnd();
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
+
   pRebMsg->rebSubHash = taosHashInit(64, MurmurHash3_32, true, HASH_NO_LOCK);
-  // TODO set cleanfp
+  if (pRebMsg->rebSubHash == NULL) {
+    mError("failed to create rebalance hashmap");
+    rpcFreeCont(pRebMsg);
+    mndRebEnd();
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
+
+  taosHashSetFreeFp(pRebMsg->rebSubHash, freeRebalanceItem);
 
   // iterate all consumers, find all modification
   while (1) {
@@ -356,7 +376,7 @@ static int32_t mndProcessMqTimerMsg(SRpcMsg *pMsg) {
   } else {
     taosHashCleanup(pRebMsg->rebSubHash);
     rpcFreeCont(pRebMsg);
-    mDebug("mq rebalance finished, no modification");
+    mDebug("mq timer finished, no need to re-balance");
     mndRebEnd();
   }
   return 0;
@@ -601,6 +621,13 @@ static void* topicNameDup(void* p){
   return taosStrdup((char*) p);
 }
 
+static void freeItem(void* param) {
+  void* pItem = *(void**)param;
+  if (pItem != NULL) {
+    taosMemoryFree(pItem);
+  }
+}
+
 int32_t mndProcessSubscribeReq(SRpcMsg *pMsg) {
   SMnode *pMnode = pMsg->info.node;
   char   *msgStr = pMsg->pCont;
@@ -616,7 +643,7 @@ int32_t mndProcessSubscribeReq(SRpcMsg *pMsg) {
   int32_t code = -1;
   SArray *pTopicList = subscribe.topicNames;
   taosArraySort(pTopicList, taosArrayCompareString);
-  taosArrayRemoveDuplicateP(pTopicList, taosArrayCompareString, taosMemoryFree);
+  taosArrayRemoveDuplicate(pTopicList, taosArrayCompareString, freeItem);
 
   int32_t newTopicNum = taosArrayGetSize(pTopicList);
 
@@ -845,6 +872,53 @@ static int32_t mndConsumerActionDelete(SSdb *pSdb, SMqConsumerObj *pConsumer) {
   return 0;
 }
 
+static void updateConsumerStatus(SMqConsumerObj* pConsumer) {
+  int32_t status = pConsumer->status;
+
+  if (taosArrayGetSize(pConsumer->rebNewTopics) == 0 && taosArrayGetSize(pConsumer->rebRemovedTopics) == 0) {
+    if (status == MQ_CONSUMER_STATUS__MODIFY || status == MQ_CONSUMER_STATUS__MODIFY_IN_REB) {
+      pConsumer->status = MQ_CONSUMER_STATUS__READY;
+    } else if (status == MQ_CONSUMER_STATUS__LOST_IN_REB || status == MQ_CONSUMER_STATUS__LOST) {
+      pConsumer->status = MQ_CONSUMER_STATUS__LOST_REBD;
+    }
+  } else {
+    if (status == MQ_CONSUMER_STATUS__MODIFY || status == MQ_CONSUMER_STATUS__MODIFY_IN_REB) {
+      pConsumer->status = MQ_CONSUMER_STATUS__MODIFY;
+    } else if (status == MQ_CONSUMER_STATUS__LOST || status == MQ_CONSUMER_STATUS__LOST_IN_REB) {
+      pConsumer->status = MQ_CONSUMER_STATUS__LOST;
+    }
+  }
+}
+
+// remove from new topic
+static void removeFromNewTopicList(SMqConsumerObj* pConsumer, const char* pTopic) {
+  int32_t size = taosArrayGetSize(pConsumer->rebNewTopics);
+  for (int32_t i = 0; i < taosArrayGetSize(pConsumer->rebNewTopics); i++) {
+    char *p = taosArrayGetP(pConsumer->rebNewTopics, i);
+    if (strcmp(pTopic, p) == 0) {
+      taosArrayRemove(pConsumer->rebNewTopics, i);
+      taosMemoryFree(p);
+
+      mDebug("consumer:0x%" PRIx64 " remove new topic:%s in the topic list, remain newTopics:%d", pConsumer->consumerId,
+             pTopic, (int) taosArrayGetSize(pConsumer->rebNewTopics));
+      break;
+    }
+  }
+}
+
+// remove from removed topic
+static void removeFromRemoveTopicList(SMqConsumerObj* pConsumer, const char* pTopic) {
+  int32_t size = taosArrayGetSize(pConsumer->rebRemovedTopics);
+  for (int32_t i = 0; i < size; i++) {
+    char *p = taosArrayGetP(pConsumer->rebRemovedTopics, i);
+    if (strcmp(pTopic, p) == 0) {
+      taosArrayRemove(pConsumer->rebRemovedTopics, i);
+      taosMemoryFree(p);
+      break;
+    }
+  }
+}
+
 static int32_t mndConsumerActionUpdate(SSdb *pSdb, SMqConsumerObj *pOldConsumer, SMqConsumerObj *pNewConsumer) {
   mDebug("consumer:0x%" PRIx64 " perform update action, update type:%d, subscribe-time:%" PRId64 ", uptime:%" PRId64,
          pOldConsumer->consumerId, pNewConsumer->updateType, pOldConsumer->subscribeTime, pOldConsumer->upTime);
@@ -855,6 +929,7 @@ static int32_t mndConsumerActionUpdate(SSdb *pSdb, SMqConsumerObj *pOldConsumer,
     /*A(taosArrayGetSize(pOldConsumer->rebNewTopics) == 0);*/
     /*A(taosArrayGetSize(pOldConsumer->rebRemovedTopics) == 0);*/
 
+    // this new consumer has identical topics with one existed consumers.
     if (taosArrayGetSize(pNewConsumer->rebNewTopics) == 0 && taosArrayGetSize(pNewConsumer->rebRemovedTopics) == 0) {
       pOldConsumer->status = MQ_CONSUMER_STATUS__READY;
     } else {
@@ -871,7 +946,6 @@ static int32_t mndConsumerActionUpdate(SSdb *pSdb, SMqConsumerObj *pOldConsumer,
       pNewConsumer->assignedTopics = tmp;
 
       pOldConsumer->subscribeTime = pNewConsumer->upTime;
-
       pOldConsumer->status = MQ_CONSUMER_STATUS__MODIFY;
     }
   } else if (pNewConsumer->updateType == CONSUMER_UPDATE__LOST) {
@@ -911,71 +985,48 @@ static int32_t mndConsumerActionUpdate(SSdb *pSdb, SMqConsumerObj *pOldConsumer,
     pOldConsumer->rebalanceTime = pNewConsumer->upTime;
 
   } else if (pNewConsumer->updateType == CONSUMER_UPDATE__ADD) {
-    /*A(taosArrayGetSize(pNewConsumer->rebNewTopics) == 1);*/
+    ASSERT(taosArrayGetSize(pNewConsumer->rebNewTopics) == 1 && taosArrayGetSize(pNewConsumer->rebRemovedTopics) == 0);
-    /*A(taosArrayGetSize(pNewConsumer->rebRemovedTopics) == 0);*/
+    char *pNewTopic = taosStrdup(taosArrayGetP(pNewConsumer->rebNewTopics, 0));
 
-    char *addedTopic = taosStrdup(taosArrayGetP(pNewConsumer->rebNewTopics, 0));
     // not exist in current topic
-
+    bool    existing = false;
-    bool existing = false;
-#if 1
     int32_t numOfExistedTopics = taosArrayGetSize(pOldConsumer->currentTopics);
     for (int32_t i = 0; i < numOfExistedTopics; i++) {
       char *topic = taosArrayGetP(pOldConsumer->currentTopics, i);
-      if (strcmp(topic, addedTopic) == 0) {
+      if (strcmp(topic, pNewTopic) == 0) {
         existing = true;
       }
     }
-#endif
 
-    // remove from new topic
+    removeFromNewTopicList(pOldConsumer, pNewTopic);
-    for (int32_t i = 0; i < taosArrayGetSize(pOldConsumer->rebNewTopics); i++) {
-      char *topic = taosArrayGetP(pOldConsumer->rebNewTopics, i);
-      if (strcmp(addedTopic, topic) == 0) {
-        taosArrayRemove(pOldConsumer->rebNewTopics, i);
-        taosMemoryFree(topic);
-        break;
-      }
-    }
 
     // add to current topic
     if (!existing) {
-      taosArrayPush(pOldConsumer->currentTopics, &addedTopic);
+      taosArrayPush(pOldConsumer->currentTopics, &pNewTopic);
       taosArraySort(pOldConsumer->currentTopics, taosArrayCompareString);
     } else {
-      taosMemoryFree(addedTopic);
+      taosMemoryFree(pNewTopic);
     }
 
     // set status
     int32_t status = pOldConsumer->status;
-    if (taosArrayGetSize(pOldConsumer->rebNewTopics) == 0 && taosArrayGetSize(pOldConsumer->rebRemovedTopics) == 0) {
+    updateConsumerStatus(pOldConsumer);
-      if (status == MQ_CONSUMER_STATUS__MODIFY || status == MQ_CONSUMER_STATUS__MODIFY_IN_REB) {
-        pOldConsumer->status = MQ_CONSUMER_STATUS__READY;
-      } else if (status == MQ_CONSUMER_STATUS__LOST_IN_REB || status == MQ_CONSUMER_STATUS__LOST) {
-        pOldConsumer->status = MQ_CONSUMER_STATUS__LOST_REBD;
-      }
-    } else {
-      if (status == MQ_CONSUMER_STATUS__MODIFY || status == MQ_CONSUMER_STATUS__MODIFY_IN_REB) {
-        pOldConsumer->status = MQ_CONSUMER_STATUS__MODIFY_IN_REB;
-      } else if (status == MQ_CONSUMER_STATUS__LOST || status == MQ_CONSUMER_STATUS__LOST_IN_REB) {
-        pOldConsumer->status = MQ_CONSUMER_STATUS__LOST_IN_REB;
-      }
-    }
 
     // the re-balance is triggered when the new consumer is launched.
     pOldConsumer->rebalanceTime = pNewConsumer->upTime;
 
     atomic_add_fetch_32(&pOldConsumer->epoch, 1);
-    mDebug("consumer:0x%" PRIx64 " state (%d)%s -> (%d)%s, new epoch:%d, reb-time:%" PRId64 ", current topics:%d",
+    mDebug("consumer:0x%" PRIx64 " state (%d)%s -> (%d)%s, new epoch:%d, reb-time:%" PRId64
+           ", current topics:%d, newTopics:%d, removeTopics:%d",
            pOldConsumer->consumerId, status, mndConsumerStatusName(status), pOldConsumer->status,
-           mndConsumerStatusName(pOldConsumer->status),
+           mndConsumerStatusName(pOldConsumer->status), pOldConsumer->epoch, pOldConsumer->rebalanceTime,
-           pOldConsumer->epoch, pOldConsumer->rebalanceTime, (int)taosArrayGetSize(pOldConsumer->currentTopics));
+           (int)taosArrayGetSize(pOldConsumer->currentTopics), (int)taosArrayGetSize(pOldConsumer->rebNewTopics),
+           (int)taosArrayGetSize(pOldConsumer->rebRemovedTopics));
+
   } else if (pNewConsumer->updateType == CONSUMER_UPDATE__REMOVE) {
     /*A(taosArrayGetSize(pNewConsumer->rebNewTopics) == 0);*/
     /*A(taosArrayGetSize(pNewConsumer->rebRemovedTopics) == 1);*/
     char *removedTopic = taosArrayGetP(pNewConsumer->rebRemovedTopics, 0);
-
-    // not exist in new topic
 #if 0
     for (int32_t i = 0; i < taosArrayGetSize(pOldConsumer->rebNewTopics); i++) {
       char *topic = taosArrayGetP(pOldConsumer->rebNewTopics, i);
@@ -984,14 +1035,7 @@ static int32_t mndConsumerActionUpdate(SSdb *pSdb, SMqConsumerObj *pOldConsumer,
 #endif
 
     // remove from removed topic
-    for (int32_t i = 0; i < taosArrayGetSize(pOldConsumer->rebRemovedTopics); i++) {
+    removeFromRemoveTopicList(pOldConsumer, removedTopic);
-      char *topic = taosArrayGetP(pOldConsumer->rebRemovedTopics, i);
-      if (strcmp(removedTopic, topic) == 0) {
-        taosArrayRemove(pOldConsumer->rebRemovedTopics, i);
-        taosMemoryFree(topic);
-        break;
-      }
-    }
 
     // remove from current topic
     int32_t i = 0;
@@ -1004,32 +1048,20 @@ static int32_t mndConsumerActionUpdate(SSdb *pSdb, SMqConsumerObj *pOldConsumer,
         break;
       }
     }
-    // must find the topic
-    /*A(i < sz);*/
 
     // set status
     int32_t status = pOldConsumer->status;
-    if (taosArrayGetSize(pOldConsumer->rebNewTopics) == 0 && taosArrayGetSize(pOldConsumer->rebRemovedTopics) == 0) {
+    updateConsumerStatus(pOldConsumer);
-      if (status == MQ_CONSUMER_STATUS__MODIFY || status == MQ_CONSUMER_STATUS__MODIFY_IN_REB) {
-        pOldConsumer->status = MQ_CONSUMER_STATUS__READY;
-      } else if (status == MQ_CONSUMER_STATUS__LOST_IN_REB || status == MQ_CONSUMER_STATUS__LOST) {
-        pOldConsumer->status = MQ_CONSUMER_STATUS__LOST_REBD;
-      }
-    } else {
-      if (status == MQ_CONSUMER_STATUS__MODIFY || status == MQ_CONSUMER_STATUS__MODIFY_IN_REB) {
-        pOldConsumer->status = MQ_CONSUMER_STATUS__MODIFY_IN_REB;
-      } else if (status == MQ_CONSUMER_STATUS__LOST || status == MQ_CONSUMER_STATUS__LOST_IN_REB) {
-        pOldConsumer->status = MQ_CONSUMER_STATUS__LOST_IN_REB;
-      }
-    }
 
     pOldConsumer->rebalanceTime = pNewConsumer->upTime;
     atomic_add_fetch_32(&pOldConsumer->epoch, 1);
 
-    mDebug("consumer:0x%" PRIx64 " state %d(%s) -> %d(%s), new epoch:%d, reb-time:%" PRId64 ", current topics:%d",
+    mDebug("consumer:0x%" PRIx64 " state (%d)%s -> (%d)%s, new epoch:%d, reb-time:%" PRId64
+               ", current topics:%d, newTopics:%d, removeTopics:%d",
            pOldConsumer->consumerId, status, mndConsumerStatusName(status), pOldConsumer->status,
-           mndConsumerStatusName(pOldConsumer->status),
+           mndConsumerStatusName(pOldConsumer->status), pOldConsumer->epoch, pOldConsumer->rebalanceTime,
-           pOldConsumer->epoch, pOldConsumer->rebalanceTime, (int)taosArrayGetSize(pOldConsumer->currentTopics));
+           (int)taosArrayGetSize(pOldConsumer->currentTopics), (int)taosArrayGetSize(pOldConsumer->rebNewTopics),
+           (int)taosArrayGetSize(pOldConsumer->rebRemovedTopics));
   }
 
   taosWUnLockLatch(&pOldConsumer->lock);

source/dnode/mnode/impl/src/mndProfile.c

@@ -226,7 +226,7 @@ static int32_t mndProcessConnectReq(SRpcMsg *pReq) {
     goto _OVER;
   }
 
-  if ((code = taosCheckVersionCompatibleFromStr(connReq.sVer, version, 2)) != 0) {
+  if ((code = taosCheckVersionCompatibleFromStr(connReq.sVer, version, 3)) != 0) {
     terrno = code;
     goto _OVER;
   }

source/dnode/mnode/impl/src/mndSubscribe.c

@@ -197,24 +197,20 @@ static SMqRebInfo *mndGetOrCreateRebSub(SHashObj *pHash, const char *key) {
   return pRebSub;
 }
 
-static int32_t mndDoRebalance(SMnode *pMnode, const SMqRebInputObj *pInput, SMqRebOutputObj *pOutput) {
+static void doRemoveExistedConsumers(SMqRebOutputObj *pOutput, SHashObj *pHash, const SMqRebInputObj *pInput) {
-  int32_t     totalVgNum = pOutput->pSub->vgNum;
+  int32_t     numOfRemoved = taosArrayGetSize(pInput->pRebInfo->removedConsumers);
-  const char *sub = pOutput->pSub->key;
+  const char *pSubKey = pOutput->pSub->key;
-  mInfo("sub:%s mq re-balance %d vgroups", sub, pOutput->pSub->vgNum);
 
-  // 1. build temporary hash(vgId -> SMqRebOutputVg) to store modified vg
-  SHashObj *pHash = taosHashInit(64, taosGetDefaultHashFunction(TSDB_DATA_TYPE_INT), false, HASH_NO_LOCK);
-
-  // 2. check and get actual removed consumers, put their vg into hash
-  int32_t removedNum = taosArrayGetSize(pInput->pRebInfo->removedConsumers);
   int32_t actualRemoved = 0;
-  for (int32_t i = 0; i < removedNum; i++) {
+  for (int32_t i = 0; i < numOfRemoved; i++) {
     uint64_t consumerId = *(uint64_t *)taosArrayGet(pInput->pRebInfo->removedConsumers, i);
 
     SMqConsumerEp *pConsumerEp = taosHashGet(pOutput->pSub->consumerHash, &consumerId, sizeof(int64_t));
 
+    // consumer exists till now
     if (pConsumerEp) {
       actualRemoved++;
+
       int32_t consumerVgNum = taosArrayGetSize(pConsumerEp->vgs);
       for (int32_t j = 0; j < consumerVgNum; j++) {
         SMqVgEp       *pVgEp = taosArrayGetP(pConsumerEp->vgs, j);
@@ -223,52 +219,66 @@ static int32_t mndDoRebalance(SMnode *pMnode, const SMqRebInputObj *pInput, SMqR
             .newConsumerId = -1,
             .pVgEp = pVgEp,
         };
+
         taosHashPut(pHash, &pVgEp->vgId, sizeof(int32_t), &outputVg, sizeof(SMqRebOutputVg));
-        mInfo("sub:%s mq re-balance remove vgId:%d from consumer:0x%" PRIx64, sub, pVgEp->vgId, consumerId);
+        mInfo("sub:%s mq re-balance remove vgId:%d from consumer:%" PRIx64, pSubKey, pVgEp->vgId, consumerId);
       }
+
       taosArrayDestroy(pConsumerEp->vgs);
       taosHashRemove(pOutput->pSub->consumerHash, &consumerId, sizeof(int64_t));
+
       // put into removed
       taosArrayPush(pOutput->removedConsumers, &consumerId);
     }
   }
 
-  if (removedNum != actualRemoved) {
+  if (numOfRemoved != actualRemoved) {
-    mError("sub:%s mq re-balance removedNum:%d not matched with actual:%d", sub, removedNum, actualRemoved);
+    mError("sub:%s mq re-balance removedNum:%d not matched with actual:%d", pSubKey, numOfRemoved, actualRemoved);
+  } else {
+    mInfo("sub:%s removed %d consumers", pSubKey, numOfRemoved);
   }
+}
 
-  // if previously no consumer, there are vgs not assigned
+static void doAddNewConsumers(SMqRebOutputObj *pOutput, const SMqRebInputObj *pInput) {
-  {
+  int32_t numOfNewConsumers = taosArrayGetSize(pInput->pRebInfo->newConsumers);
-    int32_t consumerVgNum = taosArrayGetSize(pOutput->pSub->unassignedVgs);
+  const char *pSubKey = pOutput->pSub->key;
-    for (int32_t i = 0; i < consumerVgNum; i++) {
+
-      SMqVgEp       *pVgEp = *(SMqVgEp **)taosArrayPop(pOutput->pSub->unassignedVgs);
+  for (int32_t i = 0; i < numOfNewConsumers; i++) {
-      SMqRebOutputVg rebOutput = {
+    int64_t consumerId = *(int64_t *)taosArrayGet(pInput->pRebInfo->newConsumers, i);
-          .oldConsumerId = -1,
+
-          .newConsumerId = -1,
+    SMqConsumerEp newConsumerEp;
-          .pVgEp = pVgEp,
+    newConsumerEp.consumerId = consumerId;
-      };
+    newConsumerEp.vgs = taosArrayInit(0, sizeof(void *));
-      taosHashPut(pHash, &pVgEp->vgId, sizeof(int32_t), &rebOutput, sizeof(SMqRebOutputVg));
+
-      mInfo("sub:%s mq re-balance remove vgId:%d from unassigned", sub, pVgEp->vgId);
+    taosHashPut(pOutput->pSub->consumerHash, &consumerId, sizeof(int64_t), &newConsumerEp, sizeof(SMqConsumerEp));
-    }
+    taosArrayPush(pOutput->newConsumers, &consumerId);
+    mInfo("sub:%s mq rebalance add new consumer:%" PRIx64, pSubKey, consumerId);
   }
+}
 
-  // 3. calc vg number of each consumer
+static void addUnassignedVgroups(SMqRebOutputObj *pOutput, SHashObj* pHash) {
-  int32_t afterRebConsumerNum = pInput->oldConsumerNum + taosArrayGetSize(pInput->pRebInfo->newConsumers) -
+  const char *pSubKey = pOutput->pSub->key;
-                                taosArrayGetSize(pInput->pRebInfo->removedConsumers);
+  int32_t numOfVgroups = taosArrayGetSize(pOutput->pSub->unassignedVgs);
-  int32_t minVgCnt = 0;
+
-  int32_t imbConsumerNum = 0;
+  for (int32_t i = 0; i < numOfVgroups; i++) {
-  // calc num
+    SMqVgEp       *pVgEp = *(SMqVgEp **)taosArrayPop(pOutput->pSub->unassignedVgs);
-  if (afterRebConsumerNum) {
+    SMqRebOutputVg rebOutput = {
-    minVgCnt = totalVgNum / afterRebConsumerNum;
+        .oldConsumerId = -1,
-    imbConsumerNum = totalVgNum % afterRebConsumerNum;
+        .newConsumerId = -1,
+        .pVgEp = pVgEp,
+    };
+
+    taosHashPut(pHash, &pVgEp->vgId, sizeof(int32_t), &rebOutput, sizeof(SMqRebOutputVg));
+    mInfo("sub:%s mq re-balance remove vgId:%d from unassigned", pSubKey, pVgEp->vgId);
   }
+}
 
-  mInfo("sub:%s mq re-balance %d consumers: at least %d vgs each, %d consumers has more vgs", sub,
+static void transferVgroupsForConsumers(SMqRebOutputObj *pOutput, SHashObj* pHash, int32_t minVgCnt, int32_t imbConsumerNum) {
-        afterRebConsumerNum, minVgCnt, imbConsumerNum);
+  const char *pSubKey = pOutput->pSub->key;
 
-  // 4. first scan: remove consumer more than wanted, put to remove hash
   int32_t imbCnt = 0;
   void   *pIter = NULL;
+
   while (1) {
     pIter = taosHashIterate(pOutput->pSub->consumerHash, pIter);
     if (pIter == NULL) {
@@ -276,8 +286,8 @@ static int32_t mndDoRebalance(SMnode *pMnode, const SMqRebInputObj *pInput, SMqR
     }
 
     SMqConsumerEp *pConsumerEp = (SMqConsumerEp *)pIter;
+    int32_t        consumerVgNum = taosArrayGetSize(pConsumerEp->vgs);
 
-    int32_t consumerVgNum = taosArrayGetSize(pConsumerEp->vgs);
     // all old consumers still existing are touched
     // TODO optimize: touch only consumer whose vgs changed
     taosArrayPush(pOutput->touchedConsumers, &pConsumerEp->consumerId);
@@ -296,13 +306,13 @@ static int32_t mndDoRebalance(SMnode *pMnode, const SMqRebInputObj *pInput, SMqR
                 .pVgEp = pVgEp,
             };
             taosHashPut(pHash, &pVgEp->vgId, sizeof(int32_t), &outputVg, sizeof(SMqRebOutputVg));
-            mInfo("sub:%s mq rebalance remove vgId:%d from consumer:0x%" PRIx64 ",(first scan)", sub, pVgEp->vgId,
+            mInfo("sub:%s mq rebalance remove vgId:%d from consumer:0x%" PRIx64 ",(first scan)", pSubKey, pVgEp->vgId,
                   pConsumerEp->consumerId);
           }
           imbCnt++;
         }
       } else {
-        // pop until equal minVg
+        // all the remain consumers should only have the number of vgroups, which is equalled to the value of minVg
         while (taosArrayGetSize(pConsumerEp->vgs) > minVgCnt) {
           SMqVgEp       *pVgEp = *(SMqVgEp **)taosArrayPop(pConsumerEp->vgs);
           SMqRebOutputVg outputVg = {
@@ -311,36 +321,67 @@ static int32_t mndDoRebalance(SMnode *pMnode, const SMqRebInputObj *pInput, SMqR
               .pVgEp = pVgEp,
           };
           taosHashPut(pHash, &pVgEp->vgId, sizeof(int32_t), &outputVg, sizeof(SMqRebOutputVg));
-          mInfo("sub:%s mq rebalance remove vgId:%d from consumer:0x%" PRIx64 ",(first scan)", sub, pVgEp->vgId,
+          mInfo("sub:%s mq rebalance remove vgId:%d from consumer:0x%" PRIx64 ",(first scan)", pSubKey, pVgEp->vgId,
                 pConsumerEp->consumerId);
         }
       }
     }
   }
+}
 
-  // 5. add new consumer into sub
+static int32_t mndDoRebalance(SMnode *pMnode, const SMqRebInputObj *pInput, SMqRebOutputObj *pOutput) {
-  {
+  int32_t     totalVgNum = pOutput->pSub->vgNum;
-    int32_t consumerNum = taosArrayGetSize(pInput->pRebInfo->newConsumers);
+  const char *pSubKey = pOutput->pSub->key;
-    for (int32_t i = 0; i < consumerNum; i++) {
-      int64_t consumerId = *(int64_t *)taosArrayGet(pInput->pRebInfo->newConsumers, i);
 
-      SMqConsumerEp newConsumerEp;
+  int32_t numOfRemoved = taosArrayGetSize(pInput->pRebInfo->removedConsumers);
-      newConsumerEp.consumerId = consumerId;
+  int32_t numOfAdded = taosArrayGetSize(pInput->pRebInfo->newConsumers);
-      newConsumerEp.vgs = taosArrayInit(0, sizeof(void *));
+  mInfo("sub:%s mq re-balance %d vgroups, existed consumers:%d, added:%d, removed:%d", pSubKey, totalVgNum,
-      taosHashPut(pOutput->pSub->consumerHash, &consumerId, sizeof(int64_t), &newConsumerEp, sizeof(SMqConsumerEp));
+        pInput->oldConsumerNum, numOfAdded, numOfRemoved);
-      taosArrayPush(pOutput->newConsumers, &consumerId);
+
-      mInfo("sub:%s mq rebalance add new consumer:0x%" PRIx64, sub, consumerId);
+  // 1. build temporary hash(vgId -> SMqRebOutputVg) to store modified vg
-    }
+  SHashObj *pHash = taosHashInit(64, taosGetDefaultHashFunction(TSDB_DATA_TYPE_INT), false, HASH_NO_LOCK);
+
+  // 2. check and get actual removed consumers, put their vg into hash
+  doRemoveExistedConsumers(pOutput, pHash, pInput);
+
+  // 3. if previously no consumer, there are vgs not assigned
+  addUnassignedVgroups(pOutput, pHash);
+
+  // 4. calc vg number of each consumer
+  int32_t numOfFinal = pInput->oldConsumerNum + numOfAdded - numOfRemoved;
+
+  int32_t minVgCnt = 0;
+  int32_t imbConsumerNum = 0;
+
+  // calc num
+  if (numOfFinal) {
+    minVgCnt = totalVgNum / numOfFinal;
+    imbConsumerNum = totalVgNum % numOfFinal;
+    mInfo("sub:%s mq re-balance %d consumers: at least %d vgs each, %d consumers has 1 more vgroups than avg value",
+          pSubKey, numOfFinal, minVgCnt, imbConsumerNum);
+  } else {
+    mInfo("sub:%s no consumer subscribe this topic", pSubKey);
   }
 
-  // 6. second scan: find consumer do not have enough vg, extract from temporary hash and assign to new consumer.
+  // 5. first scan: remove vgroups from te consumers, who have more vgroups than the threashold value that is
+  // minVgCnt, and then put them into the recycled hash list
+  transferVgroupsForConsumers(pOutput, pHash, minVgCnt, imbConsumerNum);
+
+  // 6. add new consumer into sub
+  doAddNewConsumers(pOutput, pInput);
+
+  // 7. second scan: find consumer do not have enough vgroups, extract from temporary hash and assign to them
   // All related vg should be put into rebVgs
   SMqRebOutputVg *pRebVg = NULL;
   void           *pRemovedIter = NULL;
-  pIter = NULL;
+  void           *pIter = NULL;
+
   while (1) {
     pIter = taosHashIterate(pOutput->pSub->consumerHash, pIter);
-    if (pIter == NULL) break;
+    if (pIter == NULL) {
+      break;
+    }
+
     SMqConsumerEp *pConsumerEp = (SMqConsumerEp *)pIter;
 
     // push until equal minVg
@@ -348,8 +389,8 @@ static int32_t mndDoRebalance(SMnode *pMnode, const SMqRebInputObj *pInput, SMqR
       // iter hash and find one vg
       pRemovedIter = taosHashIterate(pHash, pRemovedIter);
       if (pRemovedIter == NULL) {
-        mError("sub:%s removed iter is null", sub);
+        mError("sub:%s removed iter is null", pSubKey);
-        continue;
+        break;
       }
 
       pRebVg = (SMqRebOutputVg *)pRemovedIter;
@@ -409,15 +450,15 @@ static int32_t mndDoRebalance(SMnode *pMnode, const SMqRebInputObj *pInput, SMqR
 
       taosArrayPush(pOutput->pSub->unassignedVgs, &pRebOutput->pVgEp);
       taosArrayPush(pOutput->rebVgs, pRebOutput);
-      mInfo("sub:%s mq re-balance unassign vgId:%d (second scan)", sub, pRebOutput->pVgEp->vgId);
+      mInfo("sub:%s mq re-balance unassign vgId:%d (second scan)", pSubKey, pRebOutput->pVgEp->vgId);
     }
   }
 
   // 8. generate logs
-  mInfo("sub:%s mq re-balance calculation completed, re-balanced vg", sub);
+  mInfo("sub:%s mq re-balance calculation completed, re-balanced vg", pSubKey);
   for (int32_t i = 0; i < taosArrayGetSize(pOutput->rebVgs); i++) {
     SMqRebOutputVg *pOutputRebVg = taosArrayGet(pOutput->rebVgs, i);
-    mInfo("sub:%s mq re-balance vgId:%d, moved from consumer:0x%" PRIx64 ", to consumer:0x%" PRIx64, sub,
+    mInfo("sub:%s mq re-balance vgId:%d, moved from consumer:0x%" PRIx64 ", to consumer:0x%" PRIx64, pSubKey,
           pOutputRebVg->pVgEp->vgId, pOutputRebVg->oldConsumerId, pOutputRebVg->newConsumerId);
   }
   {
@@ -427,10 +468,10 @@ static int32_t mndDoRebalance(SMnode *pMnode, const SMqRebInputObj *pInput, SMqR
       if (pIter == NULL) break;
       SMqConsumerEp *pConsumerEp = (SMqConsumerEp *)pIter;
       int32_t        sz = taosArrayGetSize(pConsumerEp->vgs);
-      mInfo("sub:%s mq re-balance final cfg: consumer:0x%" PRIx64 " has %d vg", sub, pConsumerEp->consumerId, sz);
+      mInfo("sub:%s mq re-balance final cfg: consumer:0x%" PRIx64 " has %d vg", pSubKey, pConsumerEp->consumerId, sz);
       for (int32_t i = 0; i < sz; i++) {
         SMqVgEp *pVgEp = taosArrayGetP(pConsumerEp->vgs, i);
-        mInfo("sub:%s mq re-balance final cfg: vg %d to consumer:0x%" PRIx64, sub, pVgEp->vgId,
+        mInfo("sub:%s mq re-balance final cfg: vg %d to consumer:0x%" PRIx64, pSubKey, pVgEp->vgId,
               pConsumerEp->consumerId);
       }
     }
@@ -555,17 +596,23 @@ static int32_t mndProcessRebalanceReq(SRpcMsg *pMsg) {
   SMnode            *pMnode = pMsg->info.node;
   SMqDoRebalanceMsg *pReq = pMsg->pCont;
   void              *pIter = NULL;
+  bool               rebalanceOnce = false;  // to ensure only once.
 
-  mInfo("mq re-balance start");
+  mInfo("mq re-balance start, total required re-balanced trans:%d", taosHashGetSize(pReq->rebSubHash));
 
+  // here we only handle one topic rebalance requirement to ensure the atomic execution of this transaction.
   while (1) {
+    if (rebalanceOnce) {
+      break;
+    }
+
     pIter = taosHashIterate(pReq->rebSubHash, pIter);
     if (pIter == NULL) {
       break;
     }
 
+    // todo handle the malloc failure
     SMqRebInputObj rebInput = {0};
-
     SMqRebOutputObj rebOutput = {0};
     rebOutput.newConsumers = taosArrayInit(0, sizeof(int64_t));
     rebOutput.removedConsumers = taosArrayInit(0, sizeof(int64_t));
@@ -582,9 +629,10 @@ static int32_t mndProcessRebalanceReq(SRpcMsg *pMsg) {
       char topic[TSDB_TOPIC_FNAME_LEN];
       char cgroup[TSDB_CGROUP_LEN];
       mndSplitSubscribeKey(pRebInfo->key, topic, cgroup, true);
+
       SMqTopicObj *pTopic = mndAcquireTopic(pMnode, topic);
       if (pTopic == NULL) {
-        mError("mq re-balance %s ignored since topic %s not exist", pRebInfo->key, topic);
+        mError("mq re-balance %s ignored since topic %s doesn't exist", pRebInfo->key, topic);
         continue;
       }
 
@@ -604,11 +652,13 @@ static int32_t mndProcessRebalanceReq(SRpcMsg *pMsg) {
       mndReleaseTopic(pMnode, pTopic);
 
       rebInput.oldConsumerNum = 0;
+      mInfo("topic:%s has no consumers sub yet", topic);
     } else {
       taosRLockLatch(&pSub->lock);
       rebInput.oldConsumerNum = taosHashGetSize(pSub->consumerHash);
       rebOutput.pSub = tCloneSubscribeObj(pSub);
       taosRUnLockLatch(&pSub->lock);
+      mInfo("topic:%s has %d consumers sub till now", pRebInfo->key, rebInput.oldConsumerNum);
       mndReleaseSubscribe(pMnode, pSub);
     }
 
@@ -623,16 +673,14 @@ static int32_t mndProcessRebalanceReq(SRpcMsg *pMsg) {
       mError("mq re-balance persist output error, possibly vnode splitted or dropped");
     }
 
-    taosArrayDestroy(pRebInfo->lostConsumers);
-    taosArrayDestroy(pRebInfo->newConsumers);
-    taosArrayDestroy(pRebInfo->removedConsumers);
-
     taosArrayDestroy(rebOutput.newConsumers);
     taosArrayDestroy(rebOutput.touchedConsumers);
     taosArrayDestroy(rebOutput.removedConsumers);
     taosArrayDestroy(rebOutput.rebVgs);
     tDeleteSubscribeObj(rebOutput.pSub);
     taosMemoryFree(rebOutput.pSub);
+
+    rebalanceOnce = true;
   }
 
   // reset flag

source/dnode/mnode/impl/src/mndSync.c

@@ -33,7 +33,7 @@ static int32_t mndSyncEqCtrlMsg(const SMsgCb *msgcb, SRpcMsg *pMsg) {
     return -1;
   }
 
-  int32_t code = tmsgPutToQueue(msgcb, SYNC_CTRL_QUEUE, pMsg);
+  int32_t code = tmsgPutToQueue(msgcb, SYNC_RD_QUEUE, pMsg);
   if (code != 0) {
     rpcFreeCont(pMsg->pCont);
     pMsg->pCont = NULL;

source/dnode/vnode/CMakeLists.txt

@@ -57,7 +57,7 @@ target_sources(
 
     # tq
     "src/tq/tq.c"
-    "src/tq/tqExec.c"
+    "src/tq/tqScan.c"
     "src/tq/tqMeta.c"
     "src/tq/tqRead.c"
     "src/tq/tqOffset.c"

source/dnode/vnode/inc/vnode.h

@@ -178,11 +178,11 @@ typedef struct STsdbReader STsdbReader;
 
 int32_t tsdbSetTableList(STsdbReader *pReader, const void *pTableList, int32_t num);
 int32_t tsdbReaderOpen(SVnode *pVnode, SQueryTableDataCond *pCond, void *pTableList, int32_t numOfTables,
-                       SSDataBlock *pResBlock, STsdbReader **ppReader, const char *idstr);
+                       SSDataBlock *pResBlock, STsdbReader **ppReader, const char *idstr, bool countOnly);
 
 void         tsdbReaderSetId(STsdbReader* pReader, const char* idstr);
 void         tsdbReaderClose(STsdbReader *pReader);
-bool         tsdbNextDataBlock(STsdbReader *pReader);
+int32_t      tsdbNextDataBlock(STsdbReader *pReader, bool *hasNext);
 int32_t      tsdbRetrieveDatablockSMA(STsdbReader *pReader, SSDataBlock *pDataBlock, bool *allHave);
 void         tsdbReleaseDataBlock(STsdbReader *pReader);
 SSDataBlock *tsdbRetrieveDataBlock(STsdbReader *pTsdbReadHandle, SArray *pColumnIdList);

source/dnode/vnode/src/inc/tq.h

@@ -139,8 +139,7 @@ int32_t tqScanData(STQ* pTq, const STqHandle* pHandle, SMqDataRsp* pRsp, STqOffs
 int32_t tqFetchLog(STQ* pTq, STqHandle* pHandle, int64_t* fetchOffset, SWalCkHead** pHeadWithCkSum, uint64_t reqId);
 
 // tqExec
-int32_t tqTaosxScanLog(STQ* pTq, STqHandle* pHandle, SPackedData submit, STaosxRsp* pRsp);
+int32_t tqTaosxScanLog(STQ* pTq, STqHandle* pHandle, SPackedData submit, STaosxRsp* pRsp, int32_t* totalRows);
-// int32_t tqTaosxScanLog(STQ* pTq, STqHandle* pHandle, SSubmitReq* pReq, STaosxRsp* pRsp);
 int32_t tqAddBlockDataToRsp(const SSDataBlock* pBlock, SMqDataRsp* pRsp, int32_t numOfCols, int8_t precision);
 int32_t tqSendDataRsp(STQ* pTq, const SRpcMsg* pMsg, const SMqPollReq* pReq, const SMqDataRsp* pRsp, int32_t type);
 int32_t tqPushDataRsp(STQ* pTq, STqPushEntry* pPushEntry);

source/dnode/vnode/src/inc/tsdb.h

@@ -224,6 +224,8 @@ int32_t tsdbTbDataIterCreate(STbData *pTbData, TSDBKEY *pFrom, int8_t backward,
 void   *tsdbTbDataIterDestroy(STbDataIter *pIter);
 void    tsdbTbDataIterOpen(STbData *pTbData, TSDBKEY *pFrom, int8_t backward, STbDataIter *pIter);
 bool    tsdbTbDataIterNext(STbDataIter *pIter);
+void    tsdbMemTableCountRows(SMemTable *pMemTable, SHashObj*        pTableMap, int64_t *rowsNum);
+
 // STbData
 int32_t tsdbGetNRowsInTbData(STbData *pTbData);
 // tsdbFile.c ==============================================================================================

source/dnode/vnode/src/inc/vnodeInt.h

@@ -193,7 +193,7 @@ STQ*    tqOpen(const char* path, SVnode* pVnode);
 void    tqClose(STQ*);
 int     tqPushMsg(STQ*, void* msg, int32_t msgLen, tmsg_t msgType, int64_t ver);
 int     tqRegisterPushEntry(STQ* pTq, void* pHandle, const SMqPollReq* pRequest, SRpcMsg* pRpcMsg, SMqDataRsp* pDataRsp, int32_t type);
-int     tqRemovePushEntry(STQ* pTq, const char* pKey, int32_t keyLen, uint64_t consumerId, bool rspConsumer);
+int     tqUnregisterPushEntry(STQ* pTq, const char* pKey, int32_t keyLen, uint64_t consumerId, bool rspConsumer);
 
 int     tqCommit(STQ*);
 int32_t tqUpdateTbUidList(STQ* pTq, const SArray* tbUidList, bool isAdd);

source/dnode/vnode/src/sma/smaRollup.c

@@ -669,7 +669,7 @@ static int32_t tdRSmaExecAndSubmitResult(SSma *pSma, qTaskInfo_t taskInfo, SRSma
 #endif
     for (int32_t i = 0; i < taosArrayGetSize(pResList); ++i) {
       SSDataBlock *output = taosArrayGetP(pResList, i);
-      smaDebug("result block, uid:%" PRIu64 ", groupid:%" PRIu64 ", rows:%d", output->info.id.uid,
+      smaDebug("result block, uid:%" PRIu64 ", groupid:%" PRIu64 ", rows:%" PRId64, output->info.id.uid,
                output->info.id.groupId, output->info.rows);
 
       STsdb       *sinkTsdb = (pItem->level == TSDB_RETENTION_L1 ? pSma->pRSmaTsdb[0] : pSma->pRSmaTsdb[1]);

source/dnode/vnode/src/tq/tq.c

@@ -223,19 +223,6 @@ static int32_t doSendDataRsp(const SRpcHandleInfo* pRpcHandleInfo, const SMqData
 
 int32_t tqPushDataRsp(STQ* pTq, STqPushEntry* pPushEntry) {
   SMqDataRsp* pRsp = pPushEntry->pDataRsp;
-
-#if 0
-  A(taosArrayGetSize(pRsp->blockData) == pRsp->blockNum);
-  A(taosArrayGetSize(pRsp->blockDataLen) == pRsp->blockNum);
-
-  A(!pRsp->withSchema);
-  A(taosArrayGetSize(pRsp->blockSchema) == 0);
-
-  if (pRsp->reqOffset.type == TMQ_OFFSET__LOG) {
-    A(pRsp->rspOffset.version > pRsp->reqOffset.version);
-  }
-#endif
-
   SMqRspHead* pHeader = &pPushEntry->pDataRsp->head;
   doSendDataRsp(&pPushEntry->info, pRsp, pHeader->epoch, pHeader->consumerId, pHeader->mqMsgType);
 
@@ -412,8 +399,8 @@ static int32_t extractResetOffsetVal(STqOffsetVal* pOffsetVal, STQ* pTq, STqHand
 
     char formatBuf[80];
     tFormatOffset(formatBuf, 80, pOffsetVal);
-    tqDebug("tmq poll: consumer:0x%" PRIx64 ", subkey %s, vgId:%d, existed offset found, offset reset to %s and continue.",
+    tqDebug("tmq poll: consumer:0x%" PRIx64 ", subkey %s, vgId:%d, existed offset found, offset reset to %s and continue. reqId:0x%"PRIx64,
-            consumerId, pHandle->subKey, vgId, formatBuf);
+            consumerId, pHandle->subKey, vgId, formatBuf, pRequest->reqId);
     return 0;
   } else {
     // no poll occurs in this vnode for this topic, let's seek to the right offset value.
@@ -486,6 +473,10 @@ static int32_t extractDataAndRspForNormalSubscribe(STQ* pTq, STqHandle* pHandle,
 
   qSetTaskId(pHandle->execHandle.task, consumerId, pRequest->reqId);
   code = tqScanData(pTq, pHandle, &dataRsp, pOffset);
+  if (code != TSDB_CODE_SUCCESS) {
+    taosWUnLockLatch(&pTq->lock);
+    return code;
+  }
 
   // till now, all data has been transferred to consumer, new data needs to push client once arrived.
   if (dataRsp.blockNum == 0 && dataRsp.reqOffset.type == TMQ_OFFSET__LOG &&
@@ -538,123 +529,133 @@ static int32_t doPollDataForMq(STQ* pTq, STqHandle* pHandle, const SMqPollReq* p
   // this is a normal subscribe requirement
   if (pHandle->execHandle.subType == TOPIC_SUB_TYPE__COLUMN) {
     return extractDataAndRspForNormalSubscribe(pTq, pHandle, pRequest, pMsg, &offset);
-  } else { // for taosX
+  }
-    // todo handle the case where re-balance occurs.
-    SMqMetaRsp metaRsp = {0};
-    STaosxRsp  taosxRsp = {0};
-    tqInitTaosxRsp(&taosxRsp, pRequest);
 
-    if (offset.type != TMQ_OFFSET__LOG) {
+  // todo handle the case where re-balance occurs.
-      if (tqScanTaosx(pTq, pHandle, &taosxRsp, &metaRsp, &offset) < 0) {
+  // for taosx
-        return -1;
+  SMqMetaRsp metaRsp = {0};
-      }
+  STaosxRsp taosxRsp = {0};
+  tqInitTaosxRsp(&taosxRsp, pRequest);
 
-      if (metaRsp.metaRspLen > 0) {
+  if (offset.type != TMQ_OFFSET__LOG) {
-        code = tqSendMetaPollRsp(pTq, pMsg, pRequest, &metaRsp);
+    if (tqScanTaosx(pTq, pHandle, &taosxRsp, &metaRsp, &offset) < 0) {
-        tqDebug("tmq poll: consumer:0x%" PRIx64 " subkey:%s vgId:%d, send meta offset type:%d,uid:%" PRId64
+      return -1;
-                ",ts:%" PRId64,
-                consumerId, pHandle->subKey, vgId, metaRsp.rspOffset.type, metaRsp.rspOffset.uid, metaRsp.rspOffset.ts);
-        taosMemoryFree(metaRsp.metaRsp);
-        tDeleteSTaosxRsp(&taosxRsp);
-        return code;
-      }
-
-      if (taosxRsp.blockNum > 0) {
-        code = tqSendDataRsp(pTq, pMsg, pRequest, (SMqDataRsp*)&taosxRsp, TMQ_MSG_TYPE__TAOSX_RSP);
-        tDeleteSTaosxRsp(&taosxRsp);
-        return code;
-      } else {
-        offset = taosxRsp.rspOffset;
-      }
-
-      tqDebug("taosx poll: consumer:0x%" PRIx64 " subkey:%s vgId:%d, send data blockNum:%d, offset type:%d,uid:%" PRId64
-              ",version:%" PRId64,
-              consumerId, pHandle->subKey, vgId, taosxRsp.blockNum, taosxRsp.rspOffset.type, taosxRsp.rspOffset.uid,
-              taosxRsp.rspOffset.version);
     }
 
-    if (offset.type == TMQ_OFFSET__LOG) {
+    if (metaRsp.metaRspLen > 0) {
-      int64_t fetchVer = offset.version + 1;
+      code = tqSendMetaPollRsp(pTq, pMsg, pRequest, &metaRsp);
-      pCkHead = taosMemoryMalloc(sizeof(SWalCkHead) + 2048);
+      tqDebug("tmq poll: consumer:0x%" PRIx64 " subkey:%s vgId:%d, send meta offset type:%d,uid:%" PRId64
-      if (pCkHead == NULL) {
+                  ",ts:%" PRId64,
-        tDeleteSTaosxRsp(&taosxRsp);
+              consumerId, pHandle->subKey, vgId, metaRsp.rspOffset.type, metaRsp.rspOffset.uid,
-        terrno = TSDB_CODE_OUT_OF_MEMORY;
+              metaRsp.rspOffset.ts);
-        return -1;
+      taosMemoryFree(metaRsp.metaRsp);
+      tDeleteSTaosxRsp(&taosxRsp);
+      return code;
+    }
+
+    if (taosxRsp.blockNum > 0) {
+      code = tqSendDataRsp(pTq, pMsg, pRequest, (SMqDataRsp*)&taosxRsp, TMQ_MSG_TYPE__TAOSX_RSP);
+      tDeleteSTaosxRsp(&taosxRsp);
+      return code;
+    } else {
+      offset = taosxRsp.rspOffset;
+    }
+
+    tqDebug("taosx poll: consumer:0x%" PRIx64 " subkey:%s vgId:%d, send data blockNum:%d, offset type:%d,uid:%" PRId64
+                ",version:%" PRId64,
+            consumerId, pHandle->subKey, vgId, taosxRsp.blockNum, taosxRsp.rspOffset.type, taosxRsp.rspOffset.uid,
+            taosxRsp.rspOffset.version);
+  }
+
+  if (offset.type == TMQ_OFFSET__LOG) {
+    int64_t fetchVer = offset.version + 1;
+    pCkHead = taosMemoryMalloc(sizeof(SWalCkHead) + 2048);
+    if (pCkHead == NULL) {
+      tDeleteSTaosxRsp(&taosxRsp);
+      terrno = TSDB_CODE_OUT_OF_MEMORY;
+      return -1;
+    }
+    walSetReaderCapacity(pHandle->pWalReader, 2048);
+    int totalRows = 0;
+    while (1) {
+      // todo refactor: this is not correct.
+      int32_t savedEpoch = atomic_load_32(&pHandle->epoch);
+      if (savedEpoch > pRequest->epoch) {
+        tqWarn("tmq poll: consumer:0x%" PRIx64 " (epoch %d), subkey:%s vgId:%d offset %" PRId64
+                   ", found new consumer epoch %d, discard req epoch %d",
+               consumerId, pRequest->epoch, pHandle->subKey, vgId, fetchVer, savedEpoch, pRequest->epoch);
+        break;
       }
 
-      walSetReaderCapacity(pHandle->pWalReader, 2048);
+      if (tqFetchLog(pTq, pHandle, &fetchVer, &pCkHead, pRequest->reqId) < 0) {
+        tqOffsetResetToLog(&taosxRsp.rspOffset, fetchVer);
+        code = tqSendDataRsp(pTq, pMsg, pRequest, (SMqDataRsp*)&taosxRsp, TMQ_MSG_TYPE__TAOSX_RSP);
+        tDeleteSTaosxRsp(&taosxRsp);
+        taosMemoryFreeClear(pCkHead);
+        return code;
+      }
 
-      while (1) {
+      SWalCont* pHead = &pCkHead->head;
-        // todo refactor: this is not correct.
+      tqDebug("tmq poll: consumer:0x%" PRIx64 " (epoch %d) iter log, vgId:%d offset %" PRId64 " msgType %d", consumerId,
-        int32_t savedEpoch = atomic_load_32(&pHandle->epoch);
+              pRequest->epoch, vgId, fetchVer, pHead->msgType);
-        if (savedEpoch > pRequest->epoch) {
-          tqWarn("tmq poll: consumer:0x%" PRIx64 " (epoch %d), subkey:%s vgId:%d offset %" PRId64
-                 ", found new consumer epoch %d, discard req epoch %d",
-                 consumerId, pRequest->epoch, pHandle->subKey, vgId, fetchVer, savedEpoch, pRequest->epoch);
-          break;
-        }
 
-        if (tqFetchLog(pTq, pHandle, &fetchVer, &pCkHead, pRequest->reqId) < 0) {
+      // process meta
-          tqOffsetResetToLog(&taosxRsp.rspOffset, fetchVer);
+      if (pHead->msgType != TDMT_VND_SUBMIT) {
+        if(totalRows > 0) {
+          tqOffsetResetToLog(&taosxRsp.rspOffset, fetchVer - 1);
           code = tqSendDataRsp(pTq, pMsg, pRequest, (SMqDataRsp*)&taosxRsp, TMQ_MSG_TYPE__TAOSX_RSP);
           tDeleteSTaosxRsp(&taosxRsp);
           taosMemoryFreeClear(pCkHead);
           return code;
         }
 
-        SWalCont* pHead = &pCkHead->head;
+        tqDebug("fetch meta msg, ver:%" PRId64 ", type:%s", pHead->version, TMSG_INFO(pHead->msgType));
-        tqDebug("tmq poll: consumer:0x%" PRIx64 " (epoch %d) iter log, vgId:%d offset %" PRId64 " msgType %d",
+        tqOffsetResetToLog(&metaRsp.rspOffset, fetchVer);
-                consumerId, pRequest->epoch, vgId, fetchVer, pHead->msgType);
+        metaRsp.resMsgType = pHead->msgType;
-
+        metaRsp.metaRspLen = pHead->bodyLen;
-        if (pHead->msgType == TDMT_VND_SUBMIT) {
+        metaRsp.metaRsp = pHead->body;
-          SPackedData submit = {
+        if (tqSendMetaPollRsp(pTq, pMsg, pRequest, &metaRsp) < 0) {
-              .msgStr = POINTER_SHIFT(pHead->body, sizeof(SSubmitReq2Msg)),
+          code = -1;
-              .msgLen = pHead->bodyLen - sizeof(SSubmitReq2Msg),
-              .ver = pHead->version,
-          };
-
-          if (tqTaosxScanLog(pTq, pHandle, submit, &taosxRsp) < 0) {
-            tqError("tmq poll: tqTaosxScanLog error %" PRId64 ", in vgId:%d, subkey %s", consumerId, vgId,
-                    pRequest->subKey);
-            return -1;
-          }
-
-          if (taosxRsp.blockNum > 0) {
-            tqOffsetResetToLog(&taosxRsp.rspOffset, fetchVer);
-            code = tqSendDataRsp(pTq, pMsg, pRequest, (SMqDataRsp*)&taosxRsp, TMQ_MSG_TYPE__TAOSX_RSP);
-            tDeleteSTaosxRsp(&taosxRsp);
-            taosMemoryFreeClear(pCkHead);
-            return code;
-          } else {
-            fetchVer++;
-          }
-
-        } else {
-          /*A(pHandle->fetchMeta);*/
-          /*A(IS_META_MSG(pHead->msgType));*/
-          tqDebug("fetch meta msg, ver:%" PRId64 ", type:%s", pHead->version, TMSG_INFO(pHead->msgType));
-          tqOffsetResetToLog(&metaRsp.rspOffset, fetchVer);
-          metaRsp.resMsgType = pHead->msgType;
-          metaRsp.metaRspLen = pHead->bodyLen;
-          metaRsp.metaRsp = pHead->body;
-          if (tqSendMetaPollRsp(pTq, pMsg, pRequest, &metaRsp) < 0) {
-            code = -1;
-            taosMemoryFreeClear(pCkHead);
-            tDeleteSTaosxRsp(&taosxRsp);
-            return code;
-          }
-          code = 0;
           taosMemoryFreeClear(pCkHead);
           tDeleteSTaosxRsp(&taosxRsp);
           return code;
         }
+        code = 0;
+        taosMemoryFreeClear(pCkHead);
+        tDeleteSTaosxRsp(&taosxRsp);
+        return code;
+      }
+
+      // process data
+      SPackedData submit = {
+          .msgStr = POINTER_SHIFT(pHead->body, sizeof(SSubmitReq2Msg)),
+          .msgLen = pHead->bodyLen - sizeof(SSubmitReq2Msg),
+          .ver = pHead->version,
+      };
+
+      if (tqTaosxScanLog(pTq, pHandle, submit, &taosxRsp, &totalRows) < 0) {
+        tqError("tmq poll: tqTaosxScanLog error %" PRId64 ", in vgId:%d, subkey %s", consumerId, vgId,
+                pRequest->subKey);
+        taosMemoryFreeClear(pCkHead);
+        tDeleteSTaosxRsp(&taosxRsp);
+        return -1;
+      }
+
+      if (totalRows >= 4096 || taosxRsp.createTableNum > 0) {
+        tqOffsetResetToLog(&taosxRsp.rspOffset, fetchVer);
+        code = tqSendDataRsp(pTq, pMsg, pRequest, (SMqDataRsp*)&taosxRsp, TMQ_MSG_TYPE__TAOSX_RSP);
+        tDeleteSTaosxRsp(&taosxRsp);
+        taosMemoryFreeClear(pCkHead);
+        return code;
+      } else {
+        fetchVer++;
       }
     }
-
-    tDeleteSTaosxRsp(&taosxRsp);
-    taosMemoryFreeClear(pCkHead);
-    return 0;
   }
+
+  tDeleteSTaosxRsp(&taosxRsp);
+  taosMemoryFreeClear(pCkHead);
+  return 0;
 }
 
 int32_t tqProcessPollReq(STQ* pTq, SRpcMsg* pMsg) {
@@ -798,7 +799,6 @@ int32_t tqProcessSubscribeReq(STQ* pTq, int64_t sversion, char* msg, int32_t msg
 
     STqHandle tqHandle = {0};
     pHandle = &tqHandle;
-    /*taosInitRWLatch(&pExec->lock);*/
 
     uint64_t oldConsumerId = pHandle->consumerId;
     memcpy(pHandle->subKey, req.subKey, TSDB_SUBSCRIBE_KEY_LEN);
@@ -833,7 +833,7 @@ int32_t tqProcessSubscribeReq(STQ* pTq, int64_t sversion, char* msg, int32_t msg
       req.qmsg = NULL;
 
       pHandle->execHandle.task =
-          qCreateQueueExecTaskInfo(pHandle->execHandle.execCol.qmsg, &handle, vgId, &pHandle->execHandle.numOfCols, NULL);
+          qCreateQueueExecTaskInfo(pHandle->execHandle.execCol.qmsg, &handle, vgId, &pHandle->execHandle.numOfCols, req.newConsumerId);
       void* scanner = NULL;
       qExtractStreamScanner(pHandle->execHandle.task, &scanner);
       pHandle->execHandle.pExecReader = qExtractReaderFromStreamScanner(scanner);
@@ -846,7 +846,7 @@ int32_t tqProcessSubscribeReq(STQ* pTq, int64_t sversion, char* msg, int32_t msg
       buildSnapContext(handle.meta, handle.version, 0, pHandle->execHandle.subType, pHandle->fetchMeta,
                        (SSnapContext**)(&handle.sContext));
 
-      pHandle->execHandle.task = qCreateQueueExecTaskInfo(NULL, &handle, vgId, NULL, NULL);
+      pHandle->execHandle.task = qCreateQueueExecTaskInfo(NULL, &handle, vgId, NULL, req.newConsumerId);
     } else if (pHandle->execHandle.subType == TOPIC_SUB_TYPE__TABLE) {
       pHandle->pWalReader = walOpenReader(pVnode->pWal, NULL);
       pHandle->execHandle.execTb.suid = req.suid;
@@ -864,7 +864,7 @@ int32_t tqProcessSubscribeReq(STQ* pTq, int64_t sversion, char* msg, int32_t msg
 
       buildSnapContext(handle.meta, handle.version, req.suid, pHandle->execHandle.subType, pHandle->fetchMeta,
                        (SSnapContext**)(&handle.sContext));
-      pHandle->execHandle.task = qCreateQueueExecTaskInfo(NULL, &handle, vgId, NULL, NULL);
+      pHandle->execHandle.task = qCreateQueueExecTaskInfo(NULL, &handle, vgId, NULL, req.newConsumerId);
     }
 
     taosHashPut(pTq->pHandle, req.subKey, strlen(req.subKey), pHandle, sizeof(STqHandle));
@@ -885,17 +885,23 @@ int32_t tqProcessSubscribeReq(STQ* pTq, int64_t sversion, char* msg, int32_t msg
       tqInfo("vgId:%d switch consumer from Id:0x%" PRIx64 " to Id:0x%" PRIx64, req.vgId, pHandle->consumerId,
              req.newConsumerId);
 
+      // kill executing task
+      qTaskInfo_t pTaskInfo = pHandle->execHandle.task;
+      if (pTaskInfo != NULL) {
+        qKillTask(pTaskInfo, TSDB_CODE_SUCCESS);
+      }
+
       taosWLockLatch(&pTq->lock);
       atomic_store_32(&pHandle->epoch, -1);
 
       // remove if it has been register in the push manager, and return one empty block to consumer
-      tqRemovePushEntry(pTq, req.subKey, (int32_t)strlen(req.subKey), pHandle->consumerId, true);
+      tqUnregisterPushEntry(pTq, req.subKey, (int32_t)strlen(req.subKey), pHandle->consumerId, true);
 
       atomic_store_64(&pHandle->consumerId, req.newConsumerId);
       atomic_add_fetch_32(&pHandle->epoch, 1);
 
       if (pHandle->execHandle.subType == TOPIC_SUB_TYPE__COLUMN) {
-        qStreamCloseTsdbReader(pHandle->execHandle.task);
+        qStreamCloseTsdbReader(pTaskInfo);
       }
 
       taosWUnLockLatch(&pTq->lock);

source/dnode/vnode/src/tq/tqMeta.c

@@ -307,7 +307,7 @@ int32_t tqMetaRestoreHandle(STQ* pTq) {
 
     if (handle.execHandle.subType == TOPIC_SUB_TYPE__COLUMN) {
       handle.execHandle.task =
-          qCreateQueueExecTaskInfo(handle.execHandle.execCol.qmsg, &reader, vgId, &handle.execHandle.numOfCols, NULL);
+          qCreateQueueExecTaskInfo(handle.execHandle.execCol.qmsg, &reader, vgId, &handle.execHandle.numOfCols, 0);
       if (handle.execHandle.task == NULL) {
         tqError("cannot create exec task for %s", handle.subKey);
         code = -1;
@@ -332,7 +332,7 @@ int32_t tqMetaRestoreHandle(STQ* pTq) {
 
       buildSnapContext(reader.meta, reader.version, 0, handle.execHandle.subType, handle.fetchMeta,
                        (SSnapContext**)(&reader.sContext));
-      handle.execHandle.task = qCreateQueueExecTaskInfo(NULL, &reader, vgId, NULL, NULL);
+      handle.execHandle.task = qCreateQueueExecTaskInfo(NULL, &reader, vgId, NULL, 0);
     } else if (handle.execHandle.subType == TOPIC_SUB_TYPE__TABLE) {
       handle.pWalReader = walOpenReader(pTq->pVnode->pWal, NULL);
 
@@ -341,7 +341,7 @@ int32_t tqMetaRestoreHandle(STQ* pTq) {
       tqDebug("vgId:%d, tq try to get all ctb, suid:%" PRId64, pTq->pVnode->config.vgId, handle.execHandle.execTb.suid);
       for (int32_t i = 0; i < taosArrayGetSize(tbUidList); i++) {
         int64_t tbUid = *(int64_t*)taosArrayGet(tbUidList, i);
-        tqDebug("vgId:%d, idx %d, uid:%" PRId64, TD_VID(pTq->pVnode), i, tbUid);
+        tqDebug("vgId:%d, idx %d, uid:%" PRId64, vgId, i, tbUid);
       }
       handle.execHandle.pExecReader = tqOpenReader(pTq->pVnode);
       tqReaderSetTbUidList(handle.execHandle.pExecReader, tbUidList);
@@ -349,9 +349,9 @@ int32_t tqMetaRestoreHandle(STQ* pTq) {
 
       buildSnapContext(reader.meta, reader.version, handle.execHandle.execTb.suid, handle.execHandle.subType,
                        handle.fetchMeta, (SSnapContext**)(&reader.sContext));
-      handle.execHandle.task = qCreateQueueExecTaskInfo(NULL, &reader, vgId, NULL, NULL);
+      handle.execHandle.task = qCreateQueueExecTaskInfo(NULL, &reader, vgId, NULL, 0);
     }
-    tqDebug("tq restore %s consumer %" PRId64 " vgId:%d", handle.subKey, handle.consumerId, TD_VID(pTq->pVnode));
+    tqDebug("tq restore %s consumer %" PRId64 " vgId:%d", handle.subKey, handle.consumerId, vgId);
     taosHashPut(pTq->pHandle, pKey, kLen, &handle, sizeof(STqHandle));
   }
 

source/dnode/vnode/src/tq/tqPush.c

@@ -206,7 +206,84 @@ int32_t tqPushMsgNew(STQ* pTq, void* msg, int32_t msgLen, tmsg_t msgType, int64_
 }
 #endif
 
-int tqPushMsg(STQ* pTq, void* msg, int32_t msgLen, tmsg_t msgType, int64_t ver) {
+typedef struct {
+  void* pKey;
+  int64_t keyLen;
+} SItem;
+
+static void recordPushedEntry(SArray* cachedKey, void* pIter);
+
+static void freeItem(void* param) {
+  SItem* p = (SItem*) param;
+  taosMemoryFree(p->pKey);
+}
+
+static void doRemovePushedEntry(SArray* pCachedKeys, STQ* pTq) {
+  int32_t vgId = TD_VID(pTq->pVnode);
+  int32_t  numOfKeys = (int32_t) taosArrayGetSize(pCachedKeys);
+
+  for (int32_t i = 0; i < numOfKeys; i++) {
+    SItem* pItem = taosArrayGet(pCachedKeys, i);
+    if (taosHashRemove(pTq->pPushMgr, pItem->pKey, pItem->keyLen) != 0) {
+      tqError("vgId:%d, tq push hash remove key error, key: %s", vgId, (char*) pItem->pKey);
+    }
+  }
+
+  if (numOfKeys > 0) {
+    tqDebug("vgId:%d, pushed %d items and remain:%d", vgId, numOfKeys, (int32_t)taosHashGetSize(pTq->pPushMgr));
+  }
+}
+
+static void doPushDataForEntry(void* pIter, STqExecHandle* pExec, STQ* pTq, int64_t ver, int32_t vgId, char* pData,
+                               int32_t dataLen, SArray* pCachedKey) {
+  STqPushEntry* pPushEntry = *(STqPushEntry**)pIter;
+
+  SMqDataRsp* pRsp = pPushEntry->pDataRsp;
+  if (pRsp->reqOffset.version >= ver) {
+    tqDebug("vgId:%d, push entry req version %" PRId64 ", while push version %" PRId64 ", skip", vgId,
+            pRsp->reqOffset.version, ver);
+    return;
+  }
+
+  qTaskInfo_t pTaskInfo = pExec->task;
+
+  // prepare scan mem data
+  SPackedData submit = {.msgStr = pData, .msgLen = dataLen, .ver = ver};
+
+  if (qStreamSetScanMemData(pTaskInfo, submit) != 0) {
+    return;
+  }
+
+  // here start to scan submit block to extract the subscribed data
+  int32_t totalRows = 0;
+
+  while (1) {
+    SSDataBlock* pDataBlock = NULL;
+    uint64_t     ts = 0;
+    if (qExecTask(pTaskInfo, &pDataBlock, &ts) < 0) {
+      tqDebug("vgId:%d, tq exec error since %s", vgId, terrstr());
+    }
+
+    if (pDataBlock == NULL) {
+      break;
+    }
+
+    tqAddBlockDataToRsp(pDataBlock, pRsp, pExec->numOfCols, pTq->pVnode->config.tsdbCfg.precision);
+    pRsp->blockNum++;
+    totalRows += pDataBlock->info.rows;
+  }
+
+  tqDebug("vgId:%d, tq handle push, subkey:%s, block num:%d, rows:%d", vgId, pPushEntry->subKey, pRsp->blockNum,
+      totalRows);
+
+  if (pRsp->blockNum > 0) {
+    tqOffsetResetToLog(&pRsp->rspOffset, ver);
+    tqPushDataRsp(pTq, pPushEntry);
+    recordPushedEntry(pCachedKey, pIter);
+  }
+}
+
+int32_t tqPushMsg(STQ* pTq, void* msg, int32_t msgLen, tmsg_t msgType, int64_t ver) {
   void*   pReq = POINTER_SHIFT(msg, sizeof(SSubmitReq2Msg));
   int32_t len = msgLen - sizeof(SSubmitReq2Msg);
   int32_t vgId = TD_VID(pTq->pVnode);
@@ -220,24 +297,19 @@ int tqPushMsg(STQ* pTq, void* msg, int32_t msgLen, tmsg_t msgType, int64_t ver)
       tqDebug("vgId:%d tq push msg version:%" PRId64 " type:%s, head:%p, body:%p len:%d, numOfPushed consumers:%d",
           vgId, ver, TMSG_INFO(msgType), msg, pReq, len, numOfRegisteredPush);
 
-      SArray* cachedKeys = taosArrayInit(0, sizeof(void*));
+      void* data = taosMemoryMalloc(len);
-      SArray* cachedKeyLens = taosArrayInit(0, sizeof(size_t));
-
-      void*   data = taosMemoryMalloc(len);
       if (data == NULL) {
         terrno = TSDB_CODE_OUT_OF_MEMORY;
-        tqError("failed to copy data for stream since out of memory");
+        tqError("failed to copy data for stream since out of memory, vgId:%d", vgId);
-        taosArrayDestroyP(cachedKeys, (FDelete)taosMemoryFree);
-        taosArrayDestroy(cachedKeyLens);
-
-        // unlock
         taosWUnLockLatch(&pTq->lock);
         return -1;
       }
 
       memcpy(data, pReq, len);
 
-      void* pIter = NULL;
+      SArray* cachedKey = taosArrayInit(0, sizeof(SItem));
+      void*   pIter = NULL;
+
       while (1) {
         pIter = taosHashIterate(pTq->pPushMgr, pIter);
         if (pIter == NULL) {
@@ -248,83 +320,29 @@ int tqPushMsg(STQ* pTq, void* msg, int32_t msgLen, tmsg_t msgType, int64_t ver)
 
         STqHandle* pHandle = taosHashGet(pTq->pHandle, pPushEntry->subKey, strlen(pPushEntry->subKey));
         if (pHandle == NULL) {
-          tqDebug("vgId:%d, cannot find handle %s", pTq->pVnode->config.vgId, pPushEntry->subKey);
+          tqDebug("vgId:%d, failed to find handle %s in pushing data to consumer, ignore", pTq->pVnode->config.vgId, pPushEntry->subKey);
-          continue;
-        }
-
-        SMqDataRsp* pRsp = pPushEntry->pDataRsp;
-        if (pRsp->reqOffset.version >= ver) {
-          tqDebug("vgId:%d, push entry req version %" PRId64 ", while push version %" PRId64 ", skip", vgId,
-                  pRsp->reqOffset.version, ver);
           continue;
         }
 
         STqExecHandle* pExec = &pHandle->execHandle;
-        qTaskInfo_t    task = pExec->task;
+        doPushDataForEntry(pIter, pExec, pTq, ver, vgId, data, len, cachedKey);
-
-        // prepare scan mem data
-        SPackedData submit = {
-            .msgStr = data,
-            .msgLen = len,
-            .ver = ver,
-        };
-        if(qStreamSetScanMemData(task, submit) != 0){
-          continue;
-        }
-
-        // here start to scan submit block to extract the subscribed data
-        while (1) {
-          SSDataBlock* pDataBlock = NULL;
-          uint64_t     ts = 0;
-          if (qExecTask(task, &pDataBlock, &ts) < 0) {
-            tqDebug("vgId:%d, tq exec error since %s", vgId, terrstr());
-          }
-
-          if (pDataBlock == NULL) {
-            break;
-          }
-
-          tqAddBlockDataToRsp(pDataBlock, pRsp, pExec->numOfCols, pTq->pVnode->config.tsdbCfg.precision);
-          pRsp->blockNum++;
-        }
-
-        tqDebug("vgId:%d, tq handle push, subkey:%s, block num:%d", vgId, pPushEntry->subKey, pRsp->blockNum);
-        if (pRsp->blockNum > 0) {
-          // set offset
-          tqOffsetResetToLog(&pRsp->rspOffset, ver);
-
-          // remove from hash
-          size_t kLen;
-          void*  key = taosHashGetKey(pIter, &kLen);
-          void*  keyCopy = taosMemoryCalloc(1, kLen + 1);
-          memcpy(keyCopy, key, kLen);
-
-          taosArrayPush(cachedKeys, &keyCopy);
-          taosArrayPush(cachedKeyLens, &kLen);
-
-          tqPushDataRsp(pTq, pPushEntry);
-        }
       }
 
-      // delete entry
+      doRemovePushedEntry(cachedKey, pTq);
-      for (int32_t i = 0; i < taosArrayGetSize(cachedKeys); i++) {
+      taosArrayDestroyEx(cachedKey, freeItem);
-        void*  key = taosArrayGetP(cachedKeys, i);
-        size_t kLen = *(size_t*)taosArrayGet(cachedKeyLens, i);
-        if (taosHashRemove(pTq->pPushMgr, key, kLen) != 0) {
-          tqError("vgId:%d, tq push hash remove key error, key: %s", pTq->pVnode->config.vgId, (char*)key);
-        }
-      }
-
-      taosArrayDestroyP(cachedKeys, (FDelete)taosMemoryFree);
-      taosArrayDestroy(cachedKeyLens);
       taosMemoryFree(data);
     }
+
     // unlock
     taosWUnLockLatch(&pTq->lock);
   }
 
+  // push data for stream processing
   if (!tsDisableStream && vnodeIsRoleLeader(pTq->pVnode)) {
-    if (taosHashGetSize(pTq->pStreamMeta->pTasks) == 0) return 0;
+    if (taosHashGetSize(pTq->pStreamMeta->pTasks) == 0) {
+      return 0;
+    }
+
     if (msgType == TDMT_VND_SUBMIT) {
       void* data = taosMemoryMalloc(len);
       if (data == NULL) {
@@ -332,12 +350,9 @@ int tqPushMsg(STQ* pTq, void* msg, int32_t msgLen, tmsg_t msgType, int64_t ver)
         tqError("failed to copy data for stream since out of memory");
         return -1;
       }
+
       memcpy(data, pReq, len);
-      SPackedData submit = {
+      SPackedData submit = {.msgStr = data, .msgLen = len, .ver = ver};
-          .msgStr = data,
-          .msgLen = len,
-          .ver = ver,
-      };
 
       tqDebug("tq copy write msg %p %d %" PRId64 " from %p", data, len, ver, pReq);
       tqProcessSubmitReq(pTq, submit);
@@ -351,6 +366,13 @@ int tqPushMsg(STQ* pTq, void* msg, int32_t msgLen, tmsg_t msgType, int64_t ver)
   return 0;
 }
 
+void recordPushedEntry(SArray* cachedKey, void* pIter) {
+  size_t kLen = 0;
+  void*  key = taosHashGetKey(pIter, &kLen);
+  SItem item = {.pKey = strndup(key, kLen), .keyLen = kLen};
+  taosArrayPush(cachedKey, &item);
+}
+
 int32_t tqRegisterPushEntry(STQ* pTq, void* pHandle, const SMqPollReq* pRequest, SRpcMsg* pRpcMsg,
                             SMqDataRsp* pDataRsp, int32_t type) {
   uint64_t   consumerId = pRequest->consumerId;
@@ -388,8 +410,8 @@ int32_t tqRegisterPushEntry(STQ* pTq, void* pHandle, const SMqPollReq* pRequest,
   return 0;
 }
 
-int32_t tqRemovePushEntry(STQ* pTq, const char* pKey, int32_t keyLen, uint64_t consumerId, bool rspConsumer) {
+int32_t tqUnregisterPushEntry(STQ* pTq, const char* pKey, int32_t keyLen, uint64_t consumerId, bool rspConsumer) {
-  int32_t       vgId = TD_VID(pTq->pVnode);
+  int32_t        vgId = TD_VID(pTq->pVnode);
   STqPushEntry** pEntry = taosHashGet(pTq->pPushMgr, pKey, keyLen);
 
   if (pEntry != NULL) {

source/dnode/vnode/src/tq/tqRead.c

@@ -340,7 +340,7 @@ int32_t tqNextBlock(STqReader* pReader, SFetchRet* ret) {
         continue;
       }
       ret->fetchType = FETCH_TYPE__DATA;
-      tqDebug("return data rows %d", ret->data.info.rows);
+      tqDebug("return data rows %" PRId64, ret->data.info.rows);
       return 0;
     }
 

source/dnode/vnode/src/tq/tqScan.c

@@ -18,7 +18,9 @@
 int32_t tqAddBlockDataToRsp(const SSDataBlock* pBlock, SMqDataRsp* pRsp, int32_t numOfCols, int8_t precision) {
   int32_t dataStrLen = sizeof(SRetrieveTableRsp) + blockGetEncodeSize(pBlock);
   void*   buf = taosMemoryCalloc(1, dataStrLen);
-  if (buf == NULL) return -1;
+  if (buf == NULL) {
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
 
   SRetrieveTableRsp* pRetrieve = (SRetrieveTableRsp*)buf;
   pRetrieve->useconds = 0;
@@ -31,7 +33,8 @@ int32_t tqAddBlockDataToRsp(const SSDataBlock* pBlock, SMqDataRsp* pRsp, int32_t
   actualLen += sizeof(SRetrieveTableRsp);
   taosArrayPush(pRsp->blockDataLen, &actualLen);
   taosArrayPush(pRsp->blockData, &buf);
-  return 0;
+
+  return TSDB_CODE_SUCCESS;
 }
 
 static int32_t tqAddBlockSchemaToRsp(const STqExecHandle* pExec, STaosxRsp* pRsp) {
@@ -62,69 +65,85 @@ static int32_t tqAddTbNameToRsp(const STQ* pTq, int64_t uid, STaosxRsp* pRsp, in
 }
 
 int32_t tqScanData(STQ* pTq, const STqHandle* pHandle, SMqDataRsp* pRsp, STqOffsetVal* pOffset) {
-  const STqExecHandle* pExec = &pHandle->execHandle;
+  const int32_t MAX_ROWS_TO_RETURN = 4096;
+  int32_t       vgId = TD_VID(pTq->pVnode);
+  int32_t       code = 0;
+  int32_t       totalRows = 0;
 
-  qTaskInfo_t task = pExec->task;
+  const STqExecHandle* pExec = &pHandle->execHandle;
+  qTaskInfo_t          task = pExec->task;
 
   if (qStreamPrepareScan(task, pOffset, pHandle->execHandle.subType) < 0) {
-    tqDebug("prepare scan failed, return");
+    tqDebug("prepare scan failed, vgId:%d, consumer:0x%"PRIx64, vgId, pHandle->consumerId);
     if (pOffset->type == TMQ_OFFSET__LOG) {
       pRsp->rspOffset = *pOffset;
-      return 0;
+      return code;
     } else {
       tqOffsetResetToLog(pOffset, pHandle->snapshotVer);
       if (qStreamPrepareScan(task, pOffset, pHandle->execHandle.subType) < 0) {
-        tqDebug("prepare scan failed, return");
+        tqDebug("prepare scan failed, vgId:%d, consumer:0x%"PRIx64, vgId, pHandle->consumerId);
         pRsp->rspOffset = *pOffset;
-        return 0;
+        return code;
       }
     }
   }
 
-  int32_t rowCnt = 0;
   while (1) {
     SSDataBlock* pDataBlock = NULL;
     uint64_t     ts = 0;
 
-    tqDebug("vgId:%d, tmq task start to execute", pTq->pVnode->config.vgId);
+    tqDebug("vgId:%d, tmq task start to execute, consumer:0x%"PRIx64, vgId, pHandle->consumerId);
-    if (qExecTask(task, &pDataBlock, &ts) < 0) {
+
-      tqError("vgId:%d, task exec error since %s", pTq->pVnode->config.vgId, terrstr());
+    code = qExecTask(task, &pDataBlock, &ts);
-      return -1;
+    if (code != TSDB_CODE_SUCCESS) {
+      tqError("vgId:%d, task exec error since %s, consumer:0x%" PRIx64, vgId, terrstr(),
+              pHandle->consumerId);
+      return code;
     }
 
-    tqDebug("consumer:0x%"PRIx64" vgId:%d, tmq task executed, get %p", pHandle->consumerId, pTq->pVnode->config.vgId, pDataBlock);
+    // current scan should be stopped ASAP, since the re-balance occurs.
-
-    // current scan should be stopped asap, since the rebalance occurs.
     if (pDataBlock == NULL) {
       break;
     }
 
-    tqAddBlockDataToRsp(pDataBlock, pRsp, pExec->numOfCols, pTq->pVnode->config.tsdbCfg.precision);
+    code = tqAddBlockDataToRsp(pDataBlock, pRsp, pExec->numOfCols, pTq->pVnode->config.tsdbCfg.precision);
+    if (code != TSDB_CODE_SUCCESS) {
+      tqError("vgId:%d, failed to add block to rsp msg", vgId);
+      return code;
+    }
+
     pRsp->blockNum++;
 
+    tqDebug("vgId:%d, consumer:0x%" PRIx64 " tmq task executed, rows:%" PRId64 ", total blocks:%d", vgId, pHandle->consumerId,
+            pDataBlock->info.rows, pRsp->blockNum);
+
     if (pOffset->type == TMQ_OFFSET__SNAPSHOT_DATA) {
-      rowCnt += pDataBlock->info.rows;
+      totalRows += pDataBlock->info.rows;
-      if (rowCnt >= 4096) {
+      if (totalRows >= MAX_ROWS_TO_RETURN) {
         break;
       }
     }
   }
 
-  if (qStreamExtractOffset(task, &pRsp->rspOffset) < 0) {
+  qStreamExtractOffset(task, &pRsp->rspOffset);
-    return -1;
-  }
 
   if (pRsp->rspOffset.type == 0) {
-    tqError("expected rsp offset: type %d %" PRId64 " %" PRId64 " %" PRId64, pRsp->rspOffset.type, pRsp->rspOffset.ts,
+    code = TSDB_CODE_INVALID_PARA;
-            pRsp->rspOffset.uid, pRsp->rspOffset.version);
+    tqError("vgId:%d, expected rsp offset: type %d %" PRId64 " %" PRId64 " %" PRId64, vgId, pRsp->rspOffset.type,
-    return -1;
+            pRsp->rspOffset.ts, pRsp->rspOffset.uid, pRsp->rspOffset.version);
+    return code;
   }
 
   if (pRsp->withTbName || pRsp->withSchema) {
-    tqError("get column should not with meta:%d,%d", pRsp->withTbName, pRsp->withSchema);
+    code = TSDB_CODE_INVALID_PARA;
-    return -1;
+    tqError("vgId:%d, get column should not with meta:%d,%d", vgId, pRsp->withTbName, pRsp->withSchema);
+    return code;
   }
-  return 0;
+
+  tqDebug("vgId:%d, consumer:0x%" PRIx64 " tmq task executed, total blocks:%d, rows:%d", vgId, pHandle->consumerId,
+          pRsp->blockNum, totalRows);
+
+  return code;
 }
 
 int32_t tqScanTaosx(STQ* pTq, const STqHandle* pHandle, STaosxRsp* pRsp, SMqMetaRsp* pMetaRsp, STqOffsetVal* pOffset) {
@@ -230,23 +249,15 @@ int32_t tqScanTaosx(STQ* pTq, const STqHandle* pHandle, STaosxRsp* pRsp, SMqMeta
   return 0;
 }
 
-int32_t tqTaosxScanLog(STQ* pTq, STqHandle* pHandle, SPackedData submit, STaosxRsp* pRsp) {
+int32_t tqTaosxScanLog(STQ* pTq, STqHandle* pHandle, SPackedData submit, STaosxRsp* pRsp, int32_t* totalRows) {
   STqExecHandle* pExec = &pHandle->execHandle;
-  /*A(pExec->subType != TOPIC_SUB_TYPE__COLUMN);*/
-
   SArray* pBlocks = taosArrayInit(0, sizeof(SSDataBlock));
   SArray* pSchemas = taosArrayInit(0, sizeof(void*));
 
   if (pExec->subType == TOPIC_SUB_TYPE__TABLE) {
     STqReader* pReader = pExec->pExecReader;
-    /*tqReaderSetDataMsg(pReader, pReq, 0);*/
     tqReaderSetSubmitReq2(pReader, submit.msgStr, submit.msgLen, submit.ver);
     while (tqNextDataBlock2(pReader)) {
-      /*SSDataBlock block = {0};*/
-      /*if (tqRetrieveDataBlock(&block, pReader) < 0) {*/
-      /*if (terrno == TSDB_CODE_TQ_TABLE_SCHEMA_NOT_FOUND) continue;*/
-      /*}*/
-
       taosArrayClear(pBlocks);
       taosArrayClear(pSchemas);
       SSubmitTbData* pSubmitTbDataRet = NULL;
@@ -254,7 +265,6 @@ int32_t tqTaosxScanLog(STQ* pTq, STqHandle* pHandle, SPackedData submit, STaosxR
         if (terrno == TSDB_CODE_TQ_TABLE_SCHEMA_NOT_FOUND) continue;
       }
       if (pRsp->withTbName) {
-        /*int64_t uid = pExec->pExecReader->msgIter.uid;*/
         int64_t uid = pExec->pExecReader->lastBlkUid;
         if (tqAddTbNameToRsp(pTq, uid, pRsp, taosArrayGetSize(pBlocks)) < 0) {
           taosArrayDestroyEx(pBlocks, (FDelete)blockDataFreeRes);
@@ -296,6 +306,7 @@ int32_t tqTaosxScanLog(STQ* pTq, STqHandle* pHandle, SPackedData submit, STaosxR
         SSDataBlock* pBlock = taosArrayGet(pBlocks, i);
         tqAddBlockDataToRsp(pBlock, (SMqDataRsp*)pRsp, taosArrayGetSize(pBlock->pDataBlock),
                             pTq->pVnode->config.tsdbCfg.precision);
+        totalRows += pBlock->info.rows;
         blockDataFreeRes(pBlock);
         SSchemaWrapper* pSW = taosArrayGetP(pSchemas, i);
         taosArrayPush(pRsp->blockSchema, &pSW);
@@ -304,13 +315,8 @@ int32_t tqTaosxScanLog(STQ* pTq, STqHandle* pHandle, SPackedData submit, STaosxR
     }
   } else if (pExec->subType == TOPIC_SUB_TYPE__DB) {
     STqReader* pReader = pExec->pExecReader;
-    /*tqReaderSetDataMsg(pReader, pReq, 0);*/
     tqReaderSetSubmitReq2(pReader, submit.msgStr, submit.msgLen, submit.ver);
     while (tqNextDataBlockFilterOut2(pReader, pExec->execDb.pFilterOutTbUid)) {
-      /*SSDataBlock block = {0};*/
-      /*if (tqRetrieveDataBlock(&block, pReader) < 0) {*/
-      /*if (terrno == TSDB_CODE_TQ_TABLE_SCHEMA_NOT_FOUND) continue;*/
-      /*}*/
       taosArrayClear(pBlocks);
       taosArrayClear(pSchemas);
       SSubmitTbData* pSubmitTbDataRet = NULL;
@@ -355,15 +361,11 @@ int32_t tqTaosxScanLog(STQ* pTq, STqHandle* pHandle, SPackedData submit, STaosxR
 
         tEncoderClear(&encoder);
       }
-      /*tqAddBlockDataToRsp(&block, (SMqDataRsp*)pRsp, taosArrayGetSize(block.pDataBlock),*/
-      /*pTq->pVnode->config.tsdbCfg.precision);*/
-      /*blockDataFreeRes(&block);*/
-      /*tqAddBlockSchemaToRsp(pExec, (SMqDataRsp*)pRsp);*/
-      /*pRsp->blockNum++;*/
       for (int32_t i = 0; i < taosArrayGetSize(pBlocks); i++) {
         SSDataBlock* pBlock = taosArrayGet(pBlocks, i);
         tqAddBlockDataToRsp(pBlock, (SMqDataRsp*)pRsp, taosArrayGetSize(pBlock->pDataBlock),
                             pTq->pVnode->config.tsdbCfg.precision);
+        *totalRows += pBlock->info.rows;
         blockDataFreeRes(pBlock);
         SSchemaWrapper* pSW = taosArrayGetP(pSchemas, i);
         taosArrayPush(pRsp->blockSchema, &pSW);
@@ -373,9 +375,5 @@ int32_t tqTaosxScanLog(STQ* pTq, STqHandle* pHandle, SPackedData submit, STaosxR
   }
   taosArrayDestroy(pBlocks);
   taosArrayDestroy(pSchemas);
-//  if (pRsp->blockNum == 0) {
-//    return -1;
-//  }
-
   return 0;
 }

source/dnode/vnode/src/tsdb/tsdbMemTable.c

@@ -282,6 +282,40 @@ bool tsdbTbDataIterNext(STbDataIter *pIter) {
   return true;
 }
 
+int64_t tsdbCountTbDataRows(STbData *pTbData) {
+  SMemSkipListNode *pNode = pTbData->sl.pHead;
+  int64_t rowsNum = 0;
+  
+  while (NULL != pNode) {
+    pNode = SL_GET_NODE_FORWARD(pNode, 0);
+    if (pNode == pTbData->sl.pTail) {
+      return rowsNum;
+    }
+
+    rowsNum++;
+  }
+
+  return rowsNum;
+}
+
+void tsdbMemTableCountRows(SMemTable *pMemTable, SHashObj*        pTableMap, int64_t *rowsNum) {
+  taosRLockLatch(&pMemTable->latch);
+  for (int32_t i = 0; i < pMemTable->nBucket; ++i) {
+    STbData *pTbData = pMemTable->aBucket[i];
+    while (pTbData) {
+      void* p = taosHashGet(pTableMap, &pTbData->uid, sizeof(pTbData->uid));
+      if (p == NULL) {
+        pTbData = pTbData->next;
+        continue;
+      }
+      
+      *rowsNum += tsdbCountTbDataRows(pTbData);
+      pTbData = pTbData->next;
+    }
+  }
+  taosRUnLockLatch(&pMemTable->latch);
+}
+
 static int32_t tsdbMemTableRehash(SMemTable *pMemTable) {
   int32_t code = 0;
 
@@ -787,4 +821,4 @@ SArray *tsdbMemTableGetTbDataArray(SMemTable *pMemTable) {
 
 _exit:
   return aTbDataP;
-}
+}

source/dnode/vnode/src/tsdb/tsdbRead.c

@@ -15,6 +15,7 @@
 
 #include "osDef.h"
 #include "tsdb.h"
+#include "tsimplehash.h"
 
 #define ASCENDING_TRAVERSE(o) (o == TSDB_ORDER_ASC)
 
@@ -24,6 +25,11 @@ typedef enum {
   EXTERNAL_ROWS_NEXT = 0x3,
 } EContentData;
 
+typedef enum {
+  READ_MODE_COUNT_ONLY = 0x1,
+  READ_MODE_ALL,
+} EReadMode;
+
 typedef struct {
   STbDataIter* iter;
   int32_t      index;
@@ -167,6 +173,8 @@ struct STsdbReader {
   uint64_t           suid;
   int16_t            order;
   bool               freeBlock;
+  EReadMode          readMode;
+  uint64_t           rowsNum;
   STimeWindow        window;  // the primary query time window that applies to all queries
   SSDataBlock*       pResBlock;
   int32_t            capacity;
@@ -176,14 +184,15 @@ struct STsdbReader {
   SBlockLoadSuppInfo suppInfo;
   STsdbReadSnap*     pReadSnap;
   SIOCostSummary     cost;
-  STSchema*          pSchema;      // the newest version schema
+  STSchema*          pSchema;  // the newest version schema
-  STSchema*          pMemSchema;   // the previous schema for in-memory data, to avoid load schema too many times
+  //  STSchema*          pMemSchema;   // the previous schema for in-memory data, to avoid load schema too many times
-  SDataFReader*      pFileReader;  // the file reader
+  SSHashObj*    pSchemaMap;   // keep the retrieved schema info, to avoid the overhead by repeatly load schema
-  SDelFReader*       pDelFReader;  // the del file reader
+  SDataFReader* pFileReader;  // the file reader
-  SArray*            pDelIdx;      // del file block index;
+  SDelFReader*  pDelFReader;  // the del file reader
-  SBlockInfoBuf      blockInfoBuf;
+  SArray*       pDelIdx;      // del file block index;
-  int32_t            step;
+  SBlockInfoBuf blockInfoBuf;
-  STsdbReader*       innerReader[2];
+  int32_t       step;
+  STsdbReader*  innerReader[2];
 };
 
 static SFileDataBlockInfo* getCurrentBlockInfo(SDataBlockIter* pBlockIter);
@@ -949,14 +958,17 @@ static void setBlockAllDumped(SFileBlockDumpInfo* pDumpInfo, int64_t maxKey, int
   pDumpInfo->lastKey = maxKey + step;
 }
 
-static void doCopyColVal(SColumnInfoData* pColInfoData, int32_t rowIndex, int32_t colIndex, SColVal* pColVal,
+static int32_t doCopyColVal(SColumnInfoData* pColInfoData, int32_t rowIndex, int32_t colIndex, SColVal* pColVal,
                          SBlockLoadSuppInfo* pSup) {
   if (IS_VAR_DATA_TYPE(pColVal->type)) {
     if (!COL_VAL_IS_VALUE(pColVal)) {
       colDataSetNULL(pColInfoData, rowIndex);
     } else {
       varDataSetLen(pSup->buildBuf[colIndex], pColVal->value.nData);
-      ASSERT(pColVal->value.nData <= pColInfoData->info.bytes);
+      if (pColVal->value.nData > pColInfoData->info.bytes) {
+        tsdbWarn("column cid:%d actual data len %d is bigger than schema len %d", pColVal->cid, pColVal->value.nData, pColInfoData->info.bytes);
+        return TSDB_CODE_TDB_INVALID_TABLE_SCHEMA_VER;
+      }
       if (pColVal->value.nData > 0) {  // pData may be null, if nData is 0
         memcpy(varDataVal(pSup->buildBuf[colIndex]), pColVal->value.pData, pColVal->value.nData);
       }
@@ -966,6 +978,8 @@ static void doCopyColVal(SColumnInfoData* pColInfoData, int32_t rowIndex, int32_
   } else {
     colDataSetVal(pColInfoData, rowIndex, (const char*)&pColVal->value, !COL_VAL_IS_VALUE(pColVal));
   }
+
+  return TSDB_CODE_SUCCESS;
 }
 
 static SFileDataBlockInfo* getCurrentBlockInfo(SDataBlockIter* pBlockIter) {
@@ -1167,6 +1181,7 @@ static int32_t copyBlockDataToSDataBlock(STsdbReader* pReader) {
   SDataBlk*           pBlock = getCurrentBlock(pBlockIter);
   SSDataBlock*        pResBlock = pReader->pResBlock;
   int32_t             numOfOutputCols = pSupInfo->numOfCols;
+  int32_t             code = TSDB_CODE_SUCCESS;
 
   SColVal cv = {0};
   int64_t st = taosGetTimestampUs();
@@ -1244,7 +1259,10 @@ static int32_t copyBlockDataToSDataBlock(STsdbReader* pReader) {
         } else {  // varchar/nchar type
           for (int32_t j = pDumpInfo->rowIndex; rowIndex < dumpedRows; j += step) {
             tColDataGetValue(pData, j, &cv);
-            doCopyColVal(pColData, rowIndex++, i, &cv, pSupInfo);
+            code = doCopyColVal(pColData, rowIndex++, i, &cv, pSupInfo);
+            if (code) {
+              return code;
+            }
           }
         }
       }
@@ -1766,7 +1784,7 @@ static int32_t buildDataBlockFromBuf(STsdbReader* pReader, STableBlockScanInfo*
   setComposedBlockFlag(pReader, true);
 
   double elapsedTime = (taosGetTimestampUs() - st) / 1000.0;
-  tsdbDebug("%p build data block from cache completed, elapsed time:%.2f ms, numOfRows:%d, brange:%" PRId64
+  tsdbDebug("%p build data block from cache completed, elapsed time:%.2f ms, numOfRows:%" PRId64 ", brange:%" PRId64
             " - %" PRId64 ", uid:%" PRIu64 ",  %s",
             pReader, elapsedTime, pBlock->info.rows, pBlock->info.window.skey, pBlock->info.window.ekey,
             pBlockScanInfo->uid, pReader->idStr);
@@ -1776,23 +1794,29 @@ static int32_t buildDataBlockFromBuf(STsdbReader* pReader, STableBlockScanInfo*
 }
 
 static bool tryCopyDistinctRowFromFileBlock(STsdbReader* pReader, SBlockData* pBlockData, int64_t key,
-                                            SFileBlockDumpInfo* pDumpInfo) {
+                                            SFileBlockDumpInfo* pDumpInfo, bool *copied) {
   // opt version
   // 1. it is not a border point
   // 2. the direct next point is not an duplicated timestamp
+  int32_t code = TSDB_CODE_SUCCESS;
+
+  *copied = false;
   bool asc = (pReader->order == TSDB_ORDER_ASC);
   if ((pDumpInfo->rowIndex < pDumpInfo->totalRows - 1 && asc) || (pDumpInfo->rowIndex > 0 && (!asc))) {
     int32_t step = pReader->order == TSDB_ORDER_ASC ? 1 : -1;
 
     int64_t nextKey = pBlockData->aTSKEY[pDumpInfo->rowIndex + step];
     if (nextKey != key) {  // merge is not needed
-      doAppendRowFromFileBlock(pReader->pResBlock, pReader, pBlockData, pDumpInfo->rowIndex);
+      code = doAppendRowFromFileBlock(pReader->pResBlock, pReader, pBlockData, pDumpInfo->rowIndex);
+      if (code) {
+        return code;
+      }
       pDumpInfo->rowIndex += step;
-      return true;
+      *copied = true;
     }
   }
 
-  return false;
+  return code;
 }
 
 static bool nextRowFromLastBlocks(SLastBlockReader* pLastBlockReader, STableBlockScanInfo* pScanInfo,
@@ -1819,20 +1843,34 @@ static bool nextRowFromLastBlocks(SLastBlockReader* pLastBlockReader, STableBloc
 }
 
 static bool tryCopyDistinctRowFromSttBlock(TSDBROW* fRow, SLastBlockReader* pLastBlockReader,
-                                           STableBlockScanInfo* pScanInfo, int64_t ts, STsdbReader* pReader) {
+                                           STableBlockScanInfo* pScanInfo, int64_t ts, STsdbReader* pReader, bool *copied) {
+  int32_t code = TSDB_CODE_SUCCESS;
+
+  *copied = false;
+
   bool hasVal = nextRowFromLastBlocks(pLastBlockReader, pScanInfo, &pReader->verRange);
   if (hasVal) {
     int64_t next1 = getCurrentKeyInLastBlock(pLastBlockReader);
     if (next1 != ts) {
-      doAppendRowFromFileBlock(pReader->pResBlock, pReader, fRow->pBlockData, fRow->iRow);
+      code = doAppendRowFromFileBlock(pReader->pResBlock, pReader, fRow->pBlockData, fRow->iRow);
-      return true;
+      if (code) {
+        return code;
+      }
+      
+      *copied = true;
+      return code;
     }
   } else {
-    doAppendRowFromFileBlock(pReader->pResBlock, pReader, fRow->pBlockData, fRow->iRow);
+    code = doAppendRowFromFileBlock(pReader->pResBlock, pReader, fRow->pBlockData, fRow->iRow);
-    return true;
+    if (code) {
+      return code;
+    }
+      
+    *copied = true;
+    return code;
   }
 
-  return false;
+  return code;
 }
 
 static FORCE_INLINE STSchema* getLatestTableSchema(STsdbReader* pReader, uint64_t uid) {
@@ -1858,28 +1896,23 @@ static FORCE_INLINE STSchema* doGetSchemaForTSRow(int32_t sversion, STsdbReader*
     return pReader->pSchema;
   }
 
-  if (pReader->pMemSchema == NULL) {
+  void** p = tSimpleHashGet(pReader->pSchemaMap, &sversion, sizeof(sversion));
-    int32_t code =
+  if (p != NULL) {
-        metaGetTbTSchemaEx(pReader->pTsdb->pVnode->pMeta, pReader->suid, uid, sversion, &pReader->pMemSchema);
+    return *(STSchema**)p;
-    if (code != TSDB_CODE_SUCCESS) {
-      terrno = code;
-      return NULL;
-    } else {
-      return pReader->pMemSchema;
-    }
   }
 
-  if (pReader->pMemSchema->version == sversion) {
+  STSchema* ptr = NULL;
-    return pReader->pMemSchema;
+  int32_t   code = metaGetTbTSchemaEx(pReader->pTsdb->pVnode->pMeta, pReader->suid, uid, sversion, &ptr);
-  }
+  if (code != TSDB_CODE_SUCCESS) {
-
-  taosMemoryFreeClear(pReader->pMemSchema);
-  int32_t code = metaGetTbTSchemaEx(pReader->pTsdb->pVnode->pMeta, pReader->suid, uid, sversion, &pReader->pMemSchema);
-  if (code != TSDB_CODE_SUCCESS || pReader->pMemSchema == NULL) {
     terrno = code;
     return NULL;
   } else {
-    return pReader->pMemSchema;
+    code = tSimpleHashPut(pReader->pSchemaMap, &sversion, sizeof(sversion), &ptr, POINTER_BYTES);
+    if (code != TSDB_CODE_SUCCESS) {
+      terrno = code;
+      return NULL;
+    }
+    return ptr;
   }
 }
 
@@ -2022,11 +2055,12 @@ static int32_t doMergeBufAndFileRows(STsdbReader* pReader, STableBlockScanInfo*
     return code;
   }
 
-  doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
+  code = doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
 
   taosMemoryFree(pTSRow);
   tsdbRowMergerClear(&merge);
-  return TSDB_CODE_SUCCESS;
+
+  return code;
 }
 
 static int32_t doMergeFileBlockAndLastBlock(SLastBlockReader* pLastBlockReader, STsdbReader* pReader,
@@ -2034,7 +2068,8 @@ static int32_t doMergeFileBlockAndLastBlock(SLastBlockReader* pLastBlockReader,
                                             bool mergeBlockData) {
   SFileBlockDumpInfo* pDumpInfo = &pReader->status.fBlockDumpInfo;
   int64_t             tsLastBlock = getCurrentKeyInLastBlock(pLastBlockReader);
-
+  bool       copied = false;
+  int32_t    code = TSDB_CODE_SUCCESS;
   SRow*      pTSRow = NULL;
   SRowMerger merge = {0};
   TSDBROW    fRow = tMergeTreeGetRow(&pLastBlockReader->mergeTree);
@@ -2042,7 +2077,12 @@ static int32_t doMergeFileBlockAndLastBlock(SLastBlockReader* pLastBlockReader,
 
   // only last block exists
   if ((!mergeBlockData) || (tsLastBlock != pBlockData->aTSKEY[pDumpInfo->rowIndex])) {
-    if (tryCopyDistinctRowFromSttBlock(&fRow, pLastBlockReader, pBlockScanInfo, tsLastBlock, pReader)) {
+    code = tryCopyDistinctRowFromSttBlock(&fRow, pLastBlockReader, pBlockScanInfo, tsLastBlock, pReader, &copied);
+    if (code) {
+      return code;
+    }
+    
+    if (copied) {
       pBlockScanInfo->lastKey = tsLastBlock;
       return TSDB_CODE_SUCCESS;
     } else {
@@ -2060,10 +2100,15 @@ static int32_t doMergeFileBlockAndLastBlock(SLastBlockReader* pLastBlockReader,
         return code;
       }
 
-      doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
+      code = doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
 
       taosMemoryFree(pTSRow);
       tsdbRowMergerClear(&merge);
+
+      if (code != TSDB_CODE_SUCCESS) {
+        return code;
+      }
+
     }
   } else {  // not merge block data
     int32_t code = tsdbRowMergerInit(&merge, &fRow, pReader->pSchema);
@@ -2083,10 +2128,14 @@ static int32_t doMergeFileBlockAndLastBlock(SLastBlockReader* pLastBlockReader,
       return code;
     }
 
-    doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
+    code = doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
 
     taosMemoryFree(pTSRow);
     tsdbRowMergerClear(&merge);
+
+    if (code != TSDB_CODE_SUCCESS) {
+      return code;
+    }
   }
 
   return TSDB_CODE_SUCCESS;
@@ -2131,7 +2180,7 @@ static int32_t mergeFileBlockAndLastBlock(STsdbReader* pReader, SLastBlockReader
           return code;
         }
 
-        doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
+        code = doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
 
         taosMemoryFree(pTSRow);
         tsdbRowMergerClear(&merge);
@@ -2353,7 +2402,7 @@ static int32_t doMergeMultiLevelRows(STsdbReader* pReader, STableBlockScanInfo*
     return code;
   }
 
-  doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
+  code = doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
 
   taosMemoryFree(pTSRow);
   tsdbRowMergerClear(&merge);
@@ -2508,7 +2557,13 @@ bool hasDataInFileBlock(const SBlockData* pBlockData, const SFileBlockDumpInfo*
 int32_t mergeRowsInFileBlocks(SBlockData* pBlockData, STableBlockScanInfo* pBlockScanInfo, int64_t key,
                               STsdbReader* pReader) {
   SFileBlockDumpInfo* pDumpInfo = &pReader->status.fBlockDumpInfo;
-  if (tryCopyDistinctRowFromFileBlock(pReader, pBlockData, key, pDumpInfo)) {
+  bool copied = false;
+  int32_t code = tryCopyDistinctRowFromFileBlock(pReader, pBlockData, key, pDumpInfo, &copied);
+  if (code) {
+    return code;
+  }
+  
+  if (copied) {
     pBlockScanInfo->lastKey = key;
     return TSDB_CODE_SUCCESS;
   } else {
@@ -2528,11 +2583,11 @@ int32_t mergeRowsInFileBlocks(SBlockData* pBlockData, STableBlockScanInfo* pBloc
       return code;
     }
 
-    doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
+    code = doAppendRowFromTSRow(pReader->pResBlock, pReader, pTSRow, pBlockScanInfo);
 
     taosMemoryFree(pTSRow);
     tsdbRowMergerClear(&merge);
-    return TSDB_CODE_SUCCESS;
+    return code;
   }
 }
 
@@ -2649,7 +2704,10 @@ static int32_t buildComposedDataBlock(STsdbReader* pReader) {
     if (isCleanFileDataBlock(pReader, pBlockInfo, pBlock, pBlockScanInfo, keyInBuf, pLastBlockReader) &&
         pBlock->nRow <= pReader->capacity) {
       if (asc || ((!asc) && (!hasDataInLastBlock(pLastBlockReader)))) {
-        copyBlockDataToSDataBlock(pReader);
+        code = copyBlockDataToSDataBlock(pReader);
+        if (code) {
+          goto _end;
+        }
 
         // record the last key value
         pBlockScanInfo->lastKey = asc ? pBlock->maxKey.ts : pBlock->minKey.ts;
@@ -2696,8 +2754,11 @@ static int32_t buildComposedDataBlock(STsdbReader* pReader) {
       break;
     }
 
-    buildComposedDataBlockImpl(pReader, pBlockScanInfo, pBlockData, pLastBlockReader);
+    code = buildComposedDataBlockImpl(pReader, pBlockScanInfo, pBlockData, pLastBlockReader);
-
+    if (code) {
+      goto _end;
+    }
+    
     // currently loaded file data block is consumed
     if ((pBlockData->nRow > 0) && (pDumpInfo->rowIndex >= pBlockData->nRow || pDumpInfo->rowIndex < 0)) {
       SDataBlk* pBlock = getCurrentBlock(&pReader->status.blockIter);
@@ -2716,7 +2777,7 @@ _end:
 
   if (pResBlock->info.rows > 0) {
     tsdbDebug("%p uid:%" PRIu64 ", composed data block created, brange:%" PRIu64 "-%" PRIu64
-              " rows:%d, elapsed time:%.2f ms %s",
+              " rows:%" PRId64 ", elapsed time:%.2f ms %s",
               pReader, pResBlock->info.id.uid, pResBlock->info.window.skey, pResBlock->info.window.ekey,
               pResBlock->info.rows, el, pReader->idStr);
   }
@@ -2922,6 +2983,7 @@ static int32_t doLoadLastBlockSequentially(STsdbReader* pReader) {
   SReaderStatus*    pStatus = &pReader->status;
   SLastBlockReader* pLastBlockReader = pStatus->fileIter.pLastBlockReader;
   STableUidList*    pUidList = &pStatus->uidList;
+  int32_t           code = TSDB_CODE_SUCCESS;
 
   if (taosHashGetSize(pStatus->pTableMap) == 0) {
     return TSDB_CODE_SUCCESS;
@@ -2952,7 +3014,11 @@ static int32_t doLoadLastBlockSequentially(STsdbReader* pReader) {
         break;
       }
 
-      buildComposedDataBlockImpl(pReader, pScanInfo, &pReader->status.fileBlockData, pLastBlockReader);
+      code = buildComposedDataBlockImpl(pReader, pScanInfo, &pReader->status.fileBlockData, pLastBlockReader);
+      if (code) {
+        return code;
+      }
+      
       if (pResBlock->info.rows >= pReader->capacity) {
         break;
       }
@@ -2963,7 +3029,7 @@ static int32_t doLoadLastBlockSequentially(STsdbReader* pReader) {
 
     if (pResBlock->info.rows > 0) {
       tsdbDebug("%p uid:%" PRIu64 ", composed data block created, brange:%" PRIu64 "-%" PRIu64
-                " rows:%d, elapsed time:%.2f ms %s",
+                " rows:%" PRId64 ", elapsed time:%.2f ms %s",
                 pReader, pResBlock->info.id.uid, pResBlock->info.window.skey, pResBlock->info.window.ekey,
                 pResBlock->info.rows, el, pReader->idStr);
       return TSDB_CODE_SUCCESS;
@@ -3032,7 +3098,11 @@ static int32_t doBuildDataBlock(STsdbReader* pReader) {
           break;
         }
 
-        buildComposedDataBlockImpl(pReader, pScanInfo, &pReader->status.fileBlockData, pLastBlockReader);
+        code = buildComposedDataBlockImpl(pReader, pScanInfo, &pReader->status.fileBlockData, pLastBlockReader);
+        if (code) {
+          return code;
+        }
+        
         if (pResBlock->info.rows >= pReader->capacity) {
           break;
         }
@@ -3043,7 +3113,7 @@ static int32_t doBuildDataBlock(STsdbReader* pReader) {
 
       if (pResBlock->info.rows > 0) {
         tsdbDebug("%p uid:%" PRIu64 ", composed data block created, brange:%" PRIu64 "-%" PRIu64
-                  " rows:%d, elapsed time:%.2f ms %s",
+                  " rows:%" PRId64 ", elapsed time:%.2f ms %s",
                   pReader, pResBlock->info.id.uid, pResBlock->info.window.skey, pResBlock->info.window.ekey,
                   pResBlock->info.rows, el, pReader->idStr);
       }
@@ -3069,6 +3139,151 @@ static int32_t doBuildDataBlock(STsdbReader* pReader) {
   return code;
 }
 
+
+static int32_t doSumFileBlockRows(STsdbReader* pReader, SDataFReader* pFileReader) {
+  int64_t    st = taosGetTimestampUs();
+  LRUHandle* handle = NULL;
+  int32_t    code = tsdbCacheGetBlockIdx(pFileReader->pTsdb->biCache, pFileReader, &handle);
+  if (code != TSDB_CODE_SUCCESS || handle == NULL) {
+    goto _end;
+  }
+
+  int32_t numOfTables = taosHashGetSize(pReader->status.pTableMap);
+
+  SArray* aBlockIdx = (SArray*)taosLRUCacheValue(pFileReader->pTsdb->biCache, handle);
+  size_t  num = taosArrayGetSize(aBlockIdx);
+  if (num == 0) {
+    tsdbBICacheRelease(pFileReader->pTsdb->biCache, handle);
+    return TSDB_CODE_SUCCESS;
+  }
+
+  SBlockIdx*     pBlockIdx = NULL;
+  int32_t i = 0;
+  for (int32_t i = 0; i < num; ++i) {
+    pBlockIdx = (SBlockIdx*)taosArrayGet(aBlockIdx, i);
+    if (pBlockIdx->suid != pReader->suid) {
+      continue;
+    }
+
+    STableBlockScanInfo** p = taosHashGet(pReader->status.pTableMap, &pBlockIdx->uid, sizeof(pBlockIdx->uid));
+    if (p == NULL) {
+      continue;
+    }
+
+    STableBlockScanInfo *pScanInfo = *p;
+    tMapDataReset(&pScanInfo->mapData);
+    tsdbReadDataBlk(pReader->pFileReader, pBlockIdx, &pScanInfo->mapData);
+
+    SDataBlk block = {0};
+    for (int32_t j = 0; j < pScanInfo->mapData.nItem; ++j) {
+      tGetDataBlk(pScanInfo->mapData.pData + pScanInfo->mapData.aOffset[j], &block);
+      pReader->rowsNum += block.nRow;
+    }
+  }
+
+_end:
+  tsdbBICacheRelease(pFileReader->pTsdb->biCache, handle);
+  return code;
+}
+
+
+static int32_t doSumSttBlockRows(STsdbReader* pReader) {
+  int32_t   code = TSDB_CODE_SUCCESS;
+  SLastBlockReader*    pLastBlockReader = pReader->status.fileIter.pLastBlockReader;
+  SSttBlockLoadInfo*   pBlockLoadInfo = NULL;
+
+  for (int32_t i = 0; i < pReader->pFileReader->pSet->nSttF; ++i) {  // open all last file
+    pBlockLoadInfo = &pLastBlockReader->pInfo[i];
+    
+    code = tsdbReadSttBlk(pReader->pFileReader, i, pBlockLoadInfo->aSttBlk);
+    if (code) {
+      return code;
+    }
+
+    size_t size = taosArrayGetSize(pBlockLoadInfo->aSttBlk);
+    if (size >= 1) {
+      SSttBlk *pStart = taosArrayGet(pBlockLoadInfo->aSttBlk, 0);
+      SSttBlk *pEnd = taosArrayGet(pBlockLoadInfo->aSttBlk, size - 1);
+  
+      // all identical
+      if (pStart->suid == pEnd->suid) {
+        if (pStart->suid != pReader->suid) {
+          // no qualified stt block existed
+          taosArrayClear(pBlockLoadInfo->aSttBlk);
+          continue;
+        }
+        for (int32_t i = 0; i < size; ++i) {
+          SSttBlk *p = taosArrayGet(pBlockLoadInfo->aSttBlk, i);
+          pReader->rowsNum += p->nRow;
+        }
+      } else {
+        for (int32_t i = 0; i < size; ++i) {
+          SSttBlk *p = taosArrayGet(pBlockLoadInfo->aSttBlk, i);
+          uint64_t s = p->suid;
+          if (s < pReader->suid) {
+            continue;
+          }
+  
+          if (s == pReader->suid) {
+            pReader->rowsNum += p->nRow;
+          } else if (s > pReader->suid) {
+            break;
+          }
+        }
+      }
+    }
+  }
+
+  return code;
+}
+
+static int32_t readRowsCountFromFiles(STsdbReader* pReader) {
+  int32_t   code = TSDB_CODE_SUCCESS;
+
+  while (1) {
+    bool    hasNext = false;
+    int32_t code = filesetIteratorNext(&pReader->status.fileIter, pReader, &hasNext);
+    if (code) {
+      return code;
+    }
+
+    if (!hasNext) {  // no data files on disk
+      break;
+    }
+
+    code = doSumFileBlockRows(pReader, pReader->pFileReader);
+    if (code != TSDB_CODE_SUCCESS) {
+      return code;
+    }
+
+    code = doSumSttBlockRows(pReader);
+    if (code != TSDB_CODE_SUCCESS) {
+      return code;
+    }    
+  }
+
+  pReader->status.loadFromFile = false;
+
+  return code;
+}
+
+static int32_t readRowsCountFromMem(STsdbReader* pReader) {
+  int32_t   code = TSDB_CODE_SUCCESS;
+  int64_t   memNum = 0, imemNum = 0;
+  if (pReader->pReadSnap->pMem != NULL) {
+    tsdbMemTableCountRows(pReader->pReadSnap->pMem, pReader->status.pTableMap, &memNum);
+  }
+
+  if (pReader->pReadSnap->pIMem != NULL) {
+    tsdbMemTableCountRows(pReader->pReadSnap->pIMem, pReader->status.pTableMap, &imemNum);
+  }
+
+  pReader->rowsNum += memNum + imemNum;
+
+  return code;
+}
+
+
 static int32_t buildBlockFromBufferSequentially(STsdbReader* pReader) {
   SReaderStatus* pStatus = &pReader->status;
   STableUidList* pUidList = &pStatus->uidList;
@@ -3677,8 +3892,9 @@ int32_t doMergeMemIMemRows(TSDBROW* pRow, TSDBROW* piRow, STableBlockScanInfo* p
 
   if (ASCENDING_TRAVERSE(pReader->order)) {  // ascending order imem --> mem
     STSchema* pSchema = doGetSchemaForTSRow(TSDBROW_SVERSION(pRow), pReader, pBlockScanInfo->uid);
+    STSchema* piSchema = doGetSchemaForTSRow(TSDBROW_SVERSION(piRow), pReader, pBlockScanInfo->uid);
 
-    int32_t code = tsdbRowMergerInit(&merge, piRow, pSchema);
+    int32_t code = tsdbRowMergerInit2(&merge, pSchema, piRow, piSchema);
     if (code != TSDB_CODE_SUCCESS) {
       return code;
     }
@@ -3689,7 +3905,8 @@ int32_t doMergeMemIMemRows(TSDBROW* pRow, TSDBROW* piRow, STableBlockScanInfo* p
       return code;
     }
 
-    tsdbRowMerge(&merge, pRow);
+    pSchema = doGetSchemaForTSRow(TSDBROW_SVERSION(pRow), pReader, pBlockScanInfo->uid);
+    tsdbRowMergerAdd(&merge, pRow, pSchema);
     code =
         doMergeRowsInBuf(&pBlockScanInfo->iter, pBlockScanInfo->uid, k.ts, pBlockScanInfo->delSkyline, &merge, pReader);
     if (code != TSDB_CODE_SUCCESS) {
@@ -3784,6 +4001,7 @@ int32_t tsdbGetNextRowInMem(STableBlockScanInfo* pBlockScanInfo, STsdbReader* pR
 int32_t doAppendRowFromTSRow(SSDataBlock* pBlock, STsdbReader* pReader, SRow* pTSRow, STableBlockScanInfo* pScanInfo) {
   int32_t outputRowIndex = pBlock->info.rows;
   int64_t uid = pScanInfo->uid;
+  int32_t code =  TSDB_CODE_SUCCESS;
 
   int32_t numOfCols = (int32_t)taosArrayGetSize(pBlock->pDataBlock);
 
@@ -3806,7 +4024,10 @@ int32_t doAppendRowFromTSRow(SSDataBlock* pBlock, STsdbReader* pReader, SRow* pT
       SColumnInfoData* pColInfoData = taosArrayGet(pBlock->pDataBlock, pSupInfo->slotId[i]);
 
       tRowGet(pTSRow, pSchema, j, &colVal);
-      doCopyColVal(pColInfoData, outputRowIndex, i, &colVal, pSupInfo);
+      code = doCopyColVal(pColInfoData, outputRowIndex, i, &colVal, pSupInfo);
+      if (code) {
+        return code;
+      }
       i += 1;
       j += 1;
     } else if (colId < pSchema->columns[j].colId) {
@@ -3836,6 +4057,7 @@ int32_t doAppendRowFromFileBlock(SSDataBlock* pResBlock, STsdbReader* pReader, S
                                  int32_t rowIndex) {
   int32_t i = 0, j = 0;
   int32_t outputRowIndex = pResBlock->info.rows;
+  int32_t code = TSDB_CODE_SUCCESS;
 
   SBlockLoadSuppInfo* pSupInfo = &pReader->suppInfo;
   if (pReader->suppInfo.colId[i] == PRIMARYKEY_TIMESTAMP_COL_ID) {
@@ -3858,7 +4080,10 @@ int32_t doAppendRowFromFileBlock(SSDataBlock* pResBlock, STsdbReader* pReader, S
     SColumnInfoData* pCol = TARRAY_GET_ELEM(pResBlock->pDataBlock, pSupInfo->slotId[i]);
     if (pData->cid == pSupInfo->colId[i]) {
       tColDataGetValue(pData, rowIndex, &cv);
-      doCopyColVal(pCol, outputRowIndex, i, &cv, pSupInfo);
+      code = doCopyColVal(pCol, outputRowIndex, i, &cv, pSupInfo);
+      if (code) {
+        return code;
+      }
       j += 1;
     } else if (pData->cid > pCol->info.colId) {
       // the specified column does not exist in file block, fill with null data
@@ -3882,6 +4107,7 @@ int32_t doAppendRowFromFileBlock(SSDataBlock* pResBlock, STsdbReader* pReader, S
 int32_t buildDataBlockFromBufImpl(STableBlockScanInfo* pBlockScanInfo, int64_t endKey, int32_t capacity,
                                   STsdbReader* pReader) {
   SSDataBlock* pBlock = pReader->pResBlock;
+  int32_t code = TSDB_CODE_SUCCESS;
 
   do {
     //    SRow* pTSRow = NULL;
@@ -3893,13 +4119,20 @@ int32_t buildDataBlockFromBufImpl(STableBlockScanInfo* pBlockScanInfo, int64_t e
     }
 
     if (row.type == TSDBROW_ROW_FMT) {
-      doAppendRowFromTSRow(pBlock, pReader, row.pTSRow, pBlockScanInfo);
+      code = doAppendRowFromTSRow(pBlock, pReader, row.pTSRow, pBlockScanInfo);
 
       if (freeTSRow) {
         taosMemoryFree(row.pTSRow);
       }
+
+      if (code) {
+        return code;
+      }
     } else {
-      doAppendRowFromFileBlock(pBlock, pReader, row.pBlockData, row.iRow);
+      code = doAppendRowFromFileBlock(pBlock, pReader, row.pBlockData, row.iRow);
+      if (code) {
+        break;
+      }
     }
 
     // no data in buffer, return immediately
@@ -3912,7 +4145,7 @@ int32_t buildDataBlockFromBufImpl(STableBlockScanInfo* pBlockScanInfo, int64_t e
     }
   } while (1);
 
-  return TSDB_CODE_SUCCESS;
+  return code;
 }
 
 // TODO refactor: with createDataBlockScanInfo
@@ -3929,8 +4162,13 @@ int32_t tsdbSetTableList(STsdbReader* pReader, const void* pTableList, int32_t n
     if (code) {
       return code;
     }
-    pReader->status.uidList.tableUidList =
+
-        (uint64_t*)taosMemoryRealloc(pReader->status.uidList.tableUidList, sizeof(uint64_t) * num);
+    char* p1 = taosMemoryRealloc(pReader->status.uidList.tableUidList, sizeof(uint64_t) * num);
+    if (p1 == NULL) {
+      return TSDB_CODE_OUT_OF_MEMORY;
+    }
+
+    pReader->status.uidList.tableUidList = (uint64_t*)p1;
   }
 
   taosHashClear(pReader->status.pTableMap);
@@ -3986,6 +4224,8 @@ static int32_t doOpenReaderImpl(STsdbReader* pReader) {
   int32_t code = TSDB_CODE_SUCCESS;
   if (pStatus->fileIter.numOfFiles == 0) {
     pStatus->loadFromFile = false;
+  } else if (READ_MODE_COUNT_ONLY == pReader->readMode) {
+    // DO NOTHING
   } else {
     code = initForFirstBlockInFile(pReader, pBlockIter);
   }
@@ -3997,9 +4237,14 @@ static int32_t doOpenReaderImpl(STsdbReader* pReader) {
   return code;
 }
 
+static void freeSchemaFunc(void* param) {
+  void* p = *(void**)param;
+  taosMemoryFree(p);
+}
+
 // ====================================== EXPOSED APIs ======================================
 int32_t tsdbReaderOpen(SVnode* pVnode, SQueryTableDataCond* pCond, void* pTableList, int32_t numOfTables,
-                       SSDataBlock* pResBlock, STsdbReader** ppReader, const char* idstr) {
+                       SSDataBlock* pResBlock, STsdbReader** ppReader, const char* idstr, bool countOnly) {
   STimeWindow window = pCond->twindows;
   if (pCond->type == TIMEWINDOW_RANGE_EXTERNAL) {
     pCond->twindows.skey += 1;
@@ -4073,6 +4318,14 @@ int32_t tsdbReaderOpen(SVnode* pVnode, SQueryTableDataCond* pCond, void* pTableL
     }
   }
 
+  pReader->pSchemaMap = tSimpleHashInit(8, taosFastHash);
+  if (pReader->pSchemaMap == NULL) {
+    tsdbError("failed init schema hash for reader %s", pReader->idStr);
+    code = TSDB_CODE_OUT_OF_MEMORY;
+    goto _err;
+  }
+
+  tSimpleHashSetFreeFp(pReader->pSchemaMap, freeSchemaFunc);
   if (pReader->pSchema != NULL) {
     code = updateBlockSMAInfo(pReader->pSchema, &pReader->suppInfo);
     if (code != TSDB_CODE_SUCCESS) {
@@ -4091,6 +4344,10 @@ int32_t tsdbReaderOpen(SVnode* pVnode, SQueryTableDataCond* pCond, void* pTableL
 
   pReader->suspended = true;
 
+  if (countOnly) {
+    pReader->readMode = READ_MODE_COUNT_ONLY;
+  }
+  
   tsdbDebug("%p total numOfTable:%d in this query %s", pReader, numOfTables, pReader->idStr);
   return code;
 
@@ -4115,7 +4372,7 @@ void tsdbReaderClose(STsdbReader* pReader) {
       p->status.uidList.tableUidList = NULL;
       p->pReadSnap = NULL;
       p->pSchema = NULL;
-      p->pMemSchema = NULL;
+      p->pSchemaMap = NULL;
 
       p = pReader->innerReader[1];
 
@@ -4123,7 +4380,7 @@ void tsdbReaderClose(STsdbReader* pReader) {
       p->status.uidList.tableUidList = NULL;
       p->pReadSnap = NULL;
       p->pSchema = NULL;
-      p->pMemSchema = NULL;
+      p->pSchemaMap = NULL;
 
       tsdbReaderClose(pReader->innerReader[0]);
       tsdbReaderClose(pReader->innerReader[1]);
@@ -4203,10 +4460,7 @@ void tsdbReaderClose(STsdbReader* pReader) {
   taosMemoryFree(pReader->idStr);
   taosMemoryFree(pReader->pSchema);
 
-  if (pReader->pMemSchema != pReader->pSchema) {
+  tSimpleHashCleanup(pReader->pSchemaMap);
-    taosMemoryFree(pReader->pMemSchema);
-  }
-
   taosMemoryFreeClear(pReader);
 }
 
@@ -4366,14 +4620,14 @@ int32_t tsdbReaderResume(STsdbReader* pReader) {
       pPrevReader->status.pTableMap = pReader->status.pTableMap;
       pPrevReader->status.uidList = pReader->status.uidList;
       pPrevReader->pSchema = pReader->pSchema;
-      pPrevReader->pMemSchema = pReader->pMemSchema;
+      pPrevReader->pSchemaMap = pReader->pSchemaMap;
       pPrevReader->pReadSnap = pReader->pReadSnap;
 
       pNextReader->capacity = 1;
       pNextReader->status.pTableMap = pReader->status.pTableMap;
       pNextReader->status.uidList = pReader->status.uidList;
       pNextReader->pSchema = pReader->pSchema;
-      pNextReader->pMemSchema = pReader->pMemSchema;
+      pNextReader->pSchemaMap = pReader->pSchemaMap;
       pNextReader->pReadSnap = pReader->pReadSnap;
 
       code = doOpenReaderImpl(pPrevReader);
@@ -4394,43 +4648,82 @@ _err:
   return code;
 }
 
-static bool doTsdbNextDataBlock(STsdbReader* pReader) {
+static bool tsdbReadRowsCountOnly(STsdbReader* pReader) {
+  int32_t code =  TSDB_CODE_SUCCESS;
+  SSDataBlock* pBlock = pReader->pResBlock;
+
+  if (pReader->status.loadFromFile == false) {
+    return false;
+  }
+
+  code = readRowsCountFromFiles(pReader);
+  if (code != TSDB_CODE_SUCCESS) {
+    return false;
+  }
+
+  code = readRowsCountFromMem(pReader);
+  if (code != TSDB_CODE_SUCCESS) {
+    return false;
+  }
+
+  pBlock->info.rows = pReader->rowsNum;
+  pBlock->info.id.uid = 0;
+  pBlock->info.dataLoad = 0;
+  
+  pReader->rowsNum = 0;
+  
+  return pBlock->info.rows > 0;
+}
+
+static int32_t doTsdbNextDataBlock(STsdbReader* pReader, bool *hasNext) {
+  int32_t code = TSDB_CODE_SUCCESS;
+  
   // cleanup the data that belongs to the previous data block
   SSDataBlock* pBlock = pReader->pResBlock;
   blockDataCleanup(pBlock);
 
+  *hasNext = false;
+
   SReaderStatus* pStatus = &pReader->status;
   if (taosHashGetSize(pStatus->pTableMap) == 0) {
-    return false;
+    return code;
+  }
+
+  if (READ_MODE_COUNT_ONLY == pReader->readMode) {
+    return tsdbReadRowsCountOnly(pReader);
   }
 
   if (pStatus->loadFromFile) {
-    int32_t code = buildBlockFromFiles(pReader);
+    code = buildBlockFromFiles(pReader);
     if (code != TSDB_CODE_SUCCESS) {
-      return false;
+      return code;
     }
 
-    if (pBlock->info.rows > 0) {
+    if (pBlock->info.rows <= 0) {
-      return true;
-    } else {
       resetTableListIndex(&pReader->status);
-      buildBlockFromBufferSequentially(pReader);
+      code = buildBlockFromBufferSequentially(pReader);
-      return pBlock->info.rows > 0;
     }
   } else {  // no data in files, let's try the buffer
-    buildBlockFromBufferSequentially(pReader);
+    code = buildBlockFromBufferSequentially(pReader);
-    return pBlock->info.rows > 0;
   }
+
+  *hasNext = pBlock->info.rows > 0;
+
+  return code;
 }
 
-bool tsdbNextDataBlock(STsdbReader* pReader) {
+int32_t tsdbNextDataBlock(STsdbReader* pReader, bool *hasNext) {
+  int32_t code = TSDB_CODE_SUCCESS;
+
+  *hasNext = false;
+  
   if (isEmptyQueryTimeWindow(&pReader->window) || pReader->step == EXTERNAL_ROWS_NEXT) {
-    return false;
+    return code;
   }
 
   SReaderStatus* pStatus = &pReader->status;
 
-  int32_t code = tsdbAcquireReader(pReader);
+  code = tsdbAcquireReader(pReader);
   qTrace("tsdb/read: %p, take read mutex, code: %d", pReader, code);
 
   if (pReader->suspended) {
@@ -4438,16 +4731,21 @@ bool tsdbNextDataBlock(STsdbReader* pReader) {
   }
 
   if (pReader->innerReader[0] != NULL && pReader->step == 0) {
-    bool ret = doTsdbNextDataBlock(pReader->innerReader[0]);
+    code = doTsdbNextDataBlock(pReader->innerReader[0], hasNext);
+    if (code) {
+      tsdbReleaseReader(pReader);
+      return code;
+    }
+    
     pReader->step = EXTERNAL_ROWS_PREV;
-    if (ret) {
+    if (*hasNext) {
       pStatus = &pReader->innerReader[0]->status;
       if (pStatus->composedDataBlock) {
         qTrace("tsdb/read: %p, unlock read mutex", pReader);
         tsdbReleaseReader(pReader);
       }
 
-      return ret;
+      return code;
     }
   }
 
@@ -4464,14 +4762,19 @@ bool tsdbNextDataBlock(STsdbReader* pReader) {
     pReader->step = EXTERNAL_ROWS_MAIN;
   }
 
-  bool ret = doTsdbNextDataBlock(pReader);
+  code = doTsdbNextDataBlock(pReader, hasNext);
-  if (ret) {
+  if (code != TSDB_CODE_SUCCESS) {
+    tsdbReleaseReader(pReader);
+    return code;
+  }
+  
+  if (*hasNext) {
     if (pStatus->composedDataBlock) {
       qTrace("tsdb/read: %p, unlock read mutex", pReader);
       tsdbReleaseReader(pReader);
     }
 
-    return ret;
+    return code;
   }
 
   if (pReader->step == EXTERNAL_ROWS_MAIN && pReader->innerReader[1] != NULL) {
@@ -4483,23 +4786,28 @@ bool tsdbNextDataBlock(STsdbReader* pReader) {
       return code;
     }
 
-    ret = doTsdbNextDataBlock(pReader->innerReader[1]);
+    code = doTsdbNextDataBlock(pReader->innerReader[1], hasNext);
+    if (code != TSDB_CODE_SUCCESS) {
+      tsdbReleaseReader(pReader);
+      return code;
+    }
+    
     pReader->step = EXTERNAL_ROWS_NEXT;
-    if (ret) {
+    if (*hasNext) {
       pStatus = &pReader->innerReader[1]->status;
       if (pStatus->composedDataBlock) {
         qTrace("tsdb/read: %p, unlock read mutex", pReader);
         tsdbReleaseReader(pReader);
       }
 
-      return ret;
+      return code;
     }
   }
 
   qTrace("tsdb/read: %p, unlock read mutex", pReader);
   tsdbReleaseReader(pReader);
 
-  return false;
+  return code;
 }
 
 static void doFillNullColSMA(SBlockLoadSuppInfo* pSup, int32_t numOfRows, int32_t numOfCols, SColumnDataAgg* pTsAgg) {
@@ -4644,20 +4952,27 @@ STableBlockScanInfo* getTableBlockScanInfo(SHashObj* pTableMap, uint64_t uid, co
 
 static SSDataBlock* doRetrieveDataBlock(STsdbReader* pReader) {
   SReaderStatus*       pStatus = &pReader->status;
+  int32_t              code = TSDB_CODE_SUCCESS;
   SFileDataBlockInfo*  pBlockInfo = getCurrentBlockInfo(&pStatus->blockIter);
   STableBlockScanInfo* pBlockScanInfo = getTableBlockScanInfo(pStatus->pTableMap, pBlockInfo->uid, pReader->idStr);
   if (pBlockScanInfo == NULL) {
     return NULL;
   }
 
-  int32_t code = doLoadFileBlockData(pReader, &pStatus->blockIter, &pStatus->fileBlockData, pBlockScanInfo->uid);
+  code = doLoadFileBlockData(pReader, &pStatus->blockIter, &pStatus->fileBlockData, pBlockScanInfo->uid);
+  if (code != TSDB_CODE_SUCCESS) {
+    tBlockDataDestroy(&pStatus->fileBlockData);
+    terrno = code;
+    return NULL;
+  }
+
+  code = copyBlockDataToSDataBlock(pReader);
   if (code != TSDB_CODE_SUCCESS) {
     tBlockDataDestroy(&pStatus->fileBlockData);
     terrno = code;
     return NULL;
   }
 
-  copyBlockDataToSDataBlock(pReader);
   return pReader->pResBlock;
 }
 
@@ -4698,7 +5013,7 @@ int32_t tsdbReaderReset(STsdbReader* pReader, SQueryTableDataCond* pCond) {
     return TSDB_CODE_SUCCESS;
   }
 
-  SReaderStatus* pStatus = &pReader->status;
+  SReaderStatus*  pStatus = &pReader->status;
   SDataBlockIter* pBlockIter = &pStatus->blockIter;
 
   pReader->order = pCond->order;
@@ -4719,9 +5034,9 @@ int32_t tsdbReaderReset(STsdbReader* pReader, SQueryTableDataCond* pCond) {
   resetDataBlockIterator(pBlockIter, pReader->order);
   resetTableListIndex(&pReader->status);
 
-  bool asc = ASCENDING_TRAVERSE(pReader->order);
+  bool    asc = ASCENDING_TRAVERSE(pReader->order);
-  int32_t step = asc? 1:-1;
+  int32_t step = asc ? 1 : -1;
-  int64_t ts = asc? pReader->window.skey - 1 : pReader->window.ekey + 1;
+  int64_t ts = asc ? pReader->window.skey - 1 : pReader->window.ekey + 1;
   resetAllDataBlockScanInfo(pStatus->pTableMap, ts, step);
 
   int32_t code = 0;

source/dnode/vnode/src/vnd/vnodeCommit.c

@@ -156,13 +156,10 @@ int vnodeShouldCommit(SVnode *pVnode) {
   bool           needCommit = false;
 
   taosThreadMutexLock(&pVnode->mutex);
-  if (!pVnode->inUse || !diskAvail) {
+  if (pVnode->inUse && diskAvail) {
-    goto _out;
+    needCommit =
+        ((pVnode->inUse->size > pVnode->inUse->node.size) && (pSched->commitMs + SYNC_VND_COMMIT_MIN_MS < nowMs));
   }
-  needCommit =
-      (((pVnode->inUse->size > pVnode->inUse->node.size) && (pSched->commitMs + SYNC_VND_COMMIT_MIN_MS < nowMs)) ||
-       (pVnode->inUse->size > 0 && pSched->commitMs + pSched->maxWaitMs < nowMs));
-_out:
   taosThreadMutexUnlock(&pVnode->mutex);
   return needCommit;
 }

source/dnode/vnode/src/vnd/vnodeSync.c

@@ -378,7 +378,7 @@ static int32_t vnodeSyncEqCtrlMsg(const SMsgCb *msgcb, SRpcMsg *pMsg) {
     return -1;
   }
 
-  int32_t code = tmsgPutToQueue(msgcb, SYNC_CTRL_QUEUE, pMsg);
+  int32_t code = tmsgPutToQueue(msgcb, SYNC_RD_QUEUE, pMsg);
   if (code != 0) {
     rpcFreeCont(pMsg->pCont);
     pMsg->pCont = NULL;

source/libs/executor/inc/executil.h

@@ -107,6 +107,7 @@ int32_t         tableListGetGroupList(const STableListInfo* pTableList, int32_t
 uint64_t        tableListGetSize(const STableListInfo* pTableList);
 uint64_t        tableListGetSuid(const STableListInfo* pTableList);
 STableKeyInfo*  tableListGetInfo(const STableListInfo* pTableList, int32_t index);
+int32_t         tableListFind(const STableListInfo* pTableList, uint64_t uid, int32_t startIndex);
 
 size_t getResultRowSize(struct SqlFunctionCtx* pCtx, int32_t numOfOutput);
 void   initResultRowInfo(SResultRowInfo* pResultRowInfo);

source/libs/executor/inc/executorimpl.h

@@ -143,10 +143,7 @@ typedef struct {
   SQueryTableDataCond tableCond;
   int64_t             fillHistoryVer1;
   int64_t             fillHistoryVer2;
-
+  SStreamState*       pState;
-  // int8_t        triggerSaved;
-  // int64_t       deleteMarkSaved;
-  SStreamState* pState;
 } SStreamTaskInfo;
 
 typedef struct {
@@ -168,15 +165,14 @@ typedef struct STaskStopInfo {
 } STaskStopInfo;
 
 struct SExecTaskInfo {
-  STaskIdInfo   id;
+  STaskIdInfo           id;
-  uint32_t      status;
+  uint32_t              status;
-  STimeWindow   window;
+  STimeWindow           window;
-  STaskCostInfo cost;
+  STaskCostInfo         cost;
-  int64_t       owner;  // if it is in execution
+  int64_t               owner;      // if it is in execution
-  int32_t       code;
+  int32_t               code;
-  int32_t       qbufQuota;  // total available buffer (in KB) during execution query
+  int32_t               qbufQuota;  // total available buffer (in KB) during execution query
-
+  int64_t               version;    // used for stream to record wal version, why not move to sschemainfo
-  int64_t               version;  // used for stream to record wal version, why not move to sschemainfo
   SStreamTaskInfo       streamInfo;
   SSchemaInfo           schemaInfo;
   STableListInfo*       pTableInfoList;  // this is a table list
@@ -188,6 +184,7 @@ struct SExecTaskInfo {
   SLocalFetch           localFetch;
   SArray*               pResultBlockList;  // result block list
   STaskStopInfo         stopInfo;
+  SRWLatch              lock;              // secure the access of STableListInfo
 };
 
 enum {
@@ -339,6 +336,7 @@ typedef struct STableScanInfo {
   int8_t          scanMode;
   int8_t          assignBlockUid;
   bool            hasGroupByTag;
+  bool            countOnly;
 } STableScanInfo;
 
 typedef struct STableMergeScanInfo {
@@ -485,12 +483,6 @@ typedef struct SStreamScanInfo {
 } SStreamScanInfo;
 
 typedef struct {
-  //  int8_t    subType;
-  //  bool      withMeta;
-  //  int64_t   suid;
-  //  int64_t   snapVersion;
-  //  void     *metaInfo;
-  //  void     *dataInfo;
   SVnode*       vnode;
   SSDataBlock   pRes;  // result SSDataBlock
   STsdbReader*  dataReader;
@@ -559,6 +551,7 @@ typedef struct SStreamIntervalOperatorInfo {
   STimeWindowAggSupp twAggSup;
   bool               invertible;
   bool               ignoreExpiredData;
+  bool               ignoreExpiredDataSaved;
   SArray*            pDelWins;  // SWinRes
   int32_t            delIndex;
   SSDataBlock*       pDelRes;
@@ -620,6 +613,7 @@ typedef struct SStreamSessionAggOperatorInfo {
   SPhysiNode*         pPhyNode;   // create new child
   bool                isFinal;
   bool                ignoreExpiredData;
+  bool                ignoreExpiredDataSaved;
   SArray*             pUpdated;
   SSHashObj*          pStUpdated;
 } SStreamSessionAggOperatorInfo;
@@ -637,6 +631,7 @@ typedef struct SStreamStateAggOperatorInfo {
   void*               pDelIterator;
   SArray*             pChildren;  // cache for children's result;
   bool                ignoreExpiredData;
+  bool                ignoreExpiredDataSaved;
   SArray*             pUpdated;
   SSHashObj*          pSeUpdated;
 } SStreamStateAggOperatorInfo;
@@ -690,6 +685,8 @@ typedef struct SStreamFillOperatorInfo {
 #define OPTR_IS_OPENED(_optr)  (((_optr)->status & OP_OPENED) == OP_OPENED)
 #define OPTR_SET_OPENED(_optr) ((_optr)->status |= OP_OPENED)
 
+SExecTaskInfo* doCreateExecTaskInfo(uint64_t queryId, uint64_t taskId, int32_t vgId, EOPTR_EXEC_MODEL model, char* dbFName);
+
 SOperatorFpSet createOperatorFpSet(__optr_open_fn_t openFn, __optr_fn_t nextFn, __optr_fn_t cleanup,
                                    __optr_close_fn_t closeFn, __optr_reqBuf_fn_t reqBufFn, __optr_explain_fn_t explain);
 int32_t        optrDummyOpenFn(SOperatorInfo* pOperator);

source/libs/executor/inc/tfill.h

@@ -92,8 +92,8 @@ typedef struct SResultRowData {
 
 typedef struct SStreamFillLinearInfo {
   TSKEY   nextEnd;
-  SArray* pDeltaVal;      // double. value for Fill(linear).
+  SArray* pEndPoints;
-  SArray* pNextDeltaVal;  // double. value for Fill(linear).
+  SArray* pNextEndPoints;
   int64_t winIndex;
   bool    hasNext;
 } SStreamFillLinearInfo;

source/libs/executor/src/exchangeoperator.c

@@ -212,6 +212,11 @@ static SSDataBlock* loadRemoteData(SOperatorInfo* pOperator) {
       return NULL;
     }
 
+    doFilter(pBlock, pOperator->exprSupp.pFilterInfo, NULL);
+    if (blockDataGetNumOfRows(pBlock) == 0) {
+      continue;
+    }
+    
     SLimitInfo* pLimitInfo = &pExchangeInfo->limitInfo;
     if (hasLimitOffsetInfo(pLimitInfo)) {
       int32_t status = handleLimitOffset(pOperator, pLimitInfo, pBlock, false);
@@ -303,6 +308,11 @@ SOperatorInfo* createExchangeOperatorInfo(void* pTransporter, SExchangePhysiNode
                   pTaskInfo);
   pOperator->exprSupp.numOfExprs = taosArrayGetSize(pInfo->pDummyBlock->pDataBlock);
 
+  code = filterInitFromNode((SNode*)pExNode->node.pConditions, &pOperator->exprSupp.pFilterInfo, 0);
+  if (code != TSDB_CODE_SUCCESS) {
+    goto _error;
+  }
+
   pOperator->fpSet =
       createOperatorFpSet(prepareLoadRemoteData, loadRemoteData, NULL, destroyExchangeOperatorInfo, optrDefaultBufFn, NULL);
   return pOperator;

source/libs/executor/src/executil.c

@@ -34,7 +34,7 @@ struct STableListInfo {
   int32_t   numOfOuputGroups;  // the data block will be generated one by one
   int32_t*  groupOffset;       // keep the offset value for each group in the tableList
   SArray*   pTableList;
-  SHashObj* map;  // speedup acquire the tableQueryInfo by table uid
+  SHashObj* map;               // speedup acquire the tableQueryInfo by table uid
   uint64_t  suid;
 };
 
@@ -571,6 +571,10 @@ int32_t getColInfoResultForGroupby(void* metaHandle, SNodeList* group, STableLis
           memcpy(pStart, data, len);
           pStart += len;
         } else if (IS_VAR_DATA_TYPE(pValue->info.type)) {
+          if (varDataTLen(data) > pValue->info.bytes) {
+            code = TSDB_CODE_TDB_INVALID_TABLE_SCHEMA_VER;
+            goto end;
+          }
           memcpy(pStart, data, varDataTLen(data));
           pStart += varDataTLen(data);
         } else {
@@ -1800,6 +1804,21 @@ STableKeyInfo* tableListGetInfo(const STableListInfo* pTableList, int32_t index)
   return taosArrayGet(pTableList->pTableList, index);
 }
 
+int32_t tableListFind(const STableListInfo* pTableList, uint64_t uid, int32_t startIndex) {
+  int32_t numOfTables = taosArrayGetSize(pTableList->pTableList);
+  if (startIndex >= numOfTables) {
+    return -1;
+  }
+
+  for (int32_t i = startIndex; i < numOfTables; ++i) {
+    STableKeyInfo* p = taosArrayGet(pTableList->pTableList, i);
+    if (p->uid == uid) {
+      return i;
+    }
+  }
+  return -1;
+}
+
 uint64_t getTableGroupId(const STableListInfo* pTableList, uint64_t tableUid) {
   int32_t* slot = taosHashGet(pTableList->map, &tableUid, sizeof(tableUid));
   ASSERT(pTableList->map != NULL && slot != NULL);
@@ -2042,7 +2061,7 @@ int32_t createScanTableListInfo(SScanPhysiNode* pScanNode, SNodeList* pGroupTags
 
 void printDataBlock(SSDataBlock* pBlock, const char* flag) {
   if (!pBlock || pBlock->info.rows == 0) {
-    qDebug("===stream===printDataBlock: Block is Null or Empty");
+    qDebug("===stream===%s: Block is Null or Empty", flag);
     return;
   }
   char* pBuf = NULL;

source/libs/executor/src/executor.c

@@ -242,30 +242,28 @@ int32_t qSetSMAInput(qTaskInfo_t tinfo, const void* pBlocks, size_t numOfBlocks,
   return code;
 }
 
-qTaskInfo_t qCreateQueueExecTaskInfo(void* msg, SReadHandle* pReaderHandle, int32_t vgId, int32_t* numOfCols, SSchemaWrapper** pSchema) {
+qTaskInfo_t qCreateQueueExecTaskInfo(void* msg, SReadHandle* pReaderHandle, int32_t vgId, int32_t* numOfCols, uint64_t id) {
-  if (msg == NULL) {
+  if (msg == NULL) { // create raw scan
-    // create raw scan
+    SExecTaskInfo* pTaskInfo = doCreateExecTaskInfo(0, id, vgId, OPTR_EXEC_MODEL_QUEUE, "");
-    SExecTaskInfo* pTaskInfo = taosMemoryCalloc(1, sizeof(SExecTaskInfo));
     if (NULL == pTaskInfo) {
       terrno = TSDB_CODE_OUT_OF_MEMORY;
       return NULL;
     }
 
-    setTaskStatus(pTaskInfo, TASK_NOT_COMPLETED);
-
-    pTaskInfo->cost.created = taosGetTimestampUs();
-    pTaskInfo->execModel = OPTR_EXEC_MODEL_QUEUE;
     pTaskInfo->pRoot = createRawScanOperatorInfo(pReaderHandle, pTaskInfo);
     if (NULL == pTaskInfo->pRoot) {
       terrno = TSDB_CODE_OUT_OF_MEMORY;
       taosMemoryFree(pTaskInfo);
       return NULL;
     }
+
+    qDebug("create raw scan task info completed, vgId:%d, %s", vgId, GET_TASKID(pTaskInfo));
     return pTaskInfo;
   }
 
   struct SSubplan* pPlan = NULL;
-  int32_t          code = qStringToSubplan(msg, &pPlan);
+
+  int32_t code = qStringToSubplan(msg, &pPlan);
   if (code != TSDB_CODE_SUCCESS) {
     terrno = code;
     return NULL;
@@ -292,9 +290,6 @@ qTaskInfo_t qCreateQueueExecTaskInfo(void* msg, SReadHandle* pReaderHandle, int3
     }
   }
 
-  if (pSchema) {
-    *pSchema = tCloneSSchemaWrapper(((SExecTaskInfo*)pTaskInfo)->schemaInfo.qsw);
-  }
   return pTaskInfo;
 }
 
@@ -410,6 +405,7 @@ int32_t qUpdateQualifiedTableId(qTaskInfo_t tinfo, const SArray* tableIdList, bo
     }
 
     STableListInfo* pTableListInfo = pTaskInfo->pTableInfoList;
+    taosWLockLatch(&pTaskInfo->lock);
 
     for (int32_t i = 0; i < numOfQualifiedTables; ++i) {
       uint64_t*     uid = taosArrayGet(qa, i);
@@ -424,6 +420,7 @@ int32_t qUpdateQualifiedTableId(qTaskInfo_t tinfo, const SArray* tableIdList, bo
           if (code != TSDB_CODE_SUCCESS) {
             taosMemoryFree(keyBuf);
             taosArrayDestroy(qa);
+            taosWUnLockLatch(&pTaskInfo->lock);
             return code;
           }
         }
@@ -445,6 +442,7 @@ int32_t qUpdateQualifiedTableId(qTaskInfo_t tinfo, const SArray* tableIdList, bo
       tableListAddTableInfo(pTableListInfo, keyInfo.uid, keyInfo.groupId);
     }
 
+    taosWUnLockLatch(&pTaskInfo->lock);
     if (keyBuf != NULL) {
       taosMemoryFree(keyBuf);
     }
@@ -452,7 +450,9 @@ int32_t qUpdateQualifiedTableId(qTaskInfo_t tinfo, const SArray* tableIdList, bo
     taosArrayDestroy(qa);
   } else {  // remove the table id in current list
     qDebug(" %d remove child tables from the stream scanner", (int32_t)taosArrayGetSize(tableIdList));
+    taosWLockLatch(&pTaskInfo->lock);
     code = tqReaderRemoveTbUidList(pScanInfo->tqReader, tableIdList);
+    taosWUnLockLatch(&pTaskInfo->lock);
   }
 
   return code;
@@ -749,6 +749,23 @@ int32_t qAsyncKillTask(qTaskInfo_t qinfo, int32_t rspCode) {
   return TSDB_CODE_SUCCESS;
 }
 
+int32_t qKillTask(qTaskInfo_t tinfo, int32_t rspCode) {
+  SExecTaskInfo* pTaskInfo = (SExecTaskInfo*)tinfo;
+  if (pTaskInfo == NULL) {
+    return TSDB_CODE_QRY_INVALID_QHANDLE;
+  }
+
+  qDebug("%s execTask async killed", GET_TASKID(pTaskInfo));
+  setTaskKilled(pTaskInfo, rspCode);
+
+  while(qTaskIsExecuting(pTaskInfo)) {
+    taosMsleep(10);
+  }
+
+  pTaskInfo->code = rspCode;
+  return TSDB_CODE_SUCCESS;
+}
+
 bool qTaskIsExecuting(qTaskInfo_t qinfo) {
   SExecTaskInfo* pTaskInfo = (SExecTaskInfo*)qinfo;
   if (NULL == pTaskInfo) {
@@ -888,7 +905,8 @@ int32_t qStreamSetParamForRecover(qTaskInfo_t tinfo) {
       pInfo->twAggSup.deleteMarkSaved = pInfo->twAggSup.deleteMark;
       pInfo->twAggSup.calTrigger = STREAM_TRIGGER_AT_ONCE;
       pInfo->twAggSup.deleteMark = INT64_MAX;
-
+      pInfo->ignoreExpiredDataSaved = pInfo->ignoreExpiredData;
+      pInfo->ignoreExpiredData = false;
     } else if (pOperator->operatorType == QUERY_NODE_PHYSICAL_PLAN_STREAM_SESSION ||
                pOperator->operatorType == QUERY_NODE_PHYSICAL_PLAN_STREAM_SEMI_SESSION ||
                pOperator->operatorType == QUERY_NODE_PHYSICAL_PLAN_STREAM_FINAL_SESSION) {
@@ -904,6 +922,8 @@ int32_t qStreamSetParamForRecover(qTaskInfo_t tinfo) {
       pInfo->twAggSup.deleteMarkSaved = pInfo->twAggSup.deleteMark;
       pInfo->twAggSup.calTrigger = STREAM_TRIGGER_AT_ONCE;
       pInfo->twAggSup.deleteMark = INT64_MAX;
+      pInfo->ignoreExpiredDataSaved = pInfo->ignoreExpiredData;
+      pInfo->ignoreExpiredData = false;
     } else if (pOperator->operatorType == QUERY_NODE_PHYSICAL_PLAN_STREAM_STATE) {
       SStreamStateAggOperatorInfo* pInfo = pOperator->info;
       ASSERT(pInfo->twAggSup.calTrigger == STREAM_TRIGGER_AT_ONCE ||
@@ -917,6 +937,8 @@ int32_t qStreamSetParamForRecover(qTaskInfo_t tinfo) {
       pInfo->twAggSup.deleteMarkSaved = pInfo->twAggSup.deleteMark;
       pInfo->twAggSup.calTrigger = STREAM_TRIGGER_AT_ONCE;
       pInfo->twAggSup.deleteMark = INT64_MAX;
+      pInfo->ignoreExpiredDataSaved = pInfo->ignoreExpiredData;
+      pInfo->ignoreExpiredData = false;
     }
 
     // iterate operator tree
@@ -944,35 +966,23 @@ int32_t qStreamRestoreParam(qTaskInfo_t tinfo) {
         pOperator->operatorType == QUERY_NODE_PHYSICAL_PLAN_STREAM_SEMI_INTERVAL ||
         pOperator->operatorType == QUERY_NODE_PHYSICAL_PLAN_STREAM_FINAL_INTERVAL) {
       SStreamIntervalOperatorInfo* pInfo = pOperator->info;
-      /*ASSERT(pInfo->twAggSup.calTrigger == STREAM_TRIGGER_AT_ONCE);*/
-      /*ASSERT(pInfo->twAggSup.deleteMark == INT64_MAX);*/
-
       pInfo->twAggSup.calTrigger = pInfo->twAggSup.calTriggerSaved;
       pInfo->twAggSup.deleteMark = pInfo->twAggSup.deleteMarkSaved;
-      /*ASSERT(pInfo->twAggSup.calTrigger == STREAM_TRIGGER_AT_ONCE ||*/
+      pInfo->ignoreExpiredData = pInfo->ignoreExpiredDataSaved;
-      /*pInfo->twAggSup.calTrigger == STREAM_TRIGGER_WINDOW_CLOSE);*/
       qInfo("restore stream param for interval: %d,  %" PRId64, pInfo->twAggSup.calTrigger, pInfo->twAggSup.deleteMark);
     } else if (pOperator->operatorType == QUERY_NODE_PHYSICAL_PLAN_STREAM_SESSION ||
                pOperator->operatorType == QUERY_NODE_PHYSICAL_PLAN_STREAM_SEMI_SESSION ||
                pOperator->operatorType == QUERY_NODE_PHYSICAL_PLAN_STREAM_FINAL_SESSION) {
       SStreamSessionAggOperatorInfo* pInfo = pOperator->info;
-      /*ASSERT(pInfo->twAggSup.calTrigger == STREAM_TRIGGER_AT_ONCE);*/
-      /*ASSERT(pInfo->twAggSup.deleteMark == INT64_MAX);*/
-
       pInfo->twAggSup.calTrigger = pInfo->twAggSup.calTriggerSaved;
       pInfo->twAggSup.deleteMark = pInfo->twAggSup.deleteMarkSaved;
-      /*ASSERT(pInfo->twAggSup.calTrigger == STREAM_TRIGGER_AT_ONCE ||*/
+      pInfo->ignoreExpiredData = pInfo->ignoreExpiredDataSaved;
-      /*pInfo->twAggSup.calTrigger == STREAM_TRIGGER_WINDOW_CLOSE);*/
       qInfo("restore stream param for session: %d,  %" PRId64, pInfo->twAggSup.calTrigger, pInfo->twAggSup.deleteMark);
     } else if (pOperator->operatorType == QUERY_NODE_PHYSICAL_PLAN_STREAM_STATE) {
       SStreamStateAggOperatorInfo* pInfo = pOperator->info;
-      /*ASSERT(pInfo->twAggSup.calTrigger == STREAM_TRIGGER_AT_ONCE);*/
-      /*ASSERT(pInfo->twAggSup.deleteMark == INT64_MAX);*/
-
       pInfo->twAggSup.calTrigger = pInfo->twAggSup.calTriggerSaved;
       pInfo->twAggSup.deleteMark = pInfo->twAggSup.deleteMarkSaved;
-      /*ASSERT(pInfo->twAggSup.calTrigger == STREAM_TRIGGER_AT_ONCE ||*/
+      pInfo->ignoreExpiredData = pInfo->ignoreExpiredDataSaved;
-      /*pInfo->twAggSup.calTrigger == STREAM_TRIGGER_WINDOW_CLOSE);*/
       qInfo("restore stream param for state: %d,  %" PRId64, pInfo->twAggSup.calTrigger, pInfo->twAggSup.deleteMark);
     }
 
@@ -990,6 +1000,7 @@ int32_t qStreamRestoreParam(qTaskInfo_t tinfo) {
   }
   return 0;
 }
+
 bool qStreamRecoverScanFinished(qTaskInfo_t tinfo) {
   SExecTaskInfo* pTaskInfo = (SExecTaskInfo*)tinfo;
   return pTaskInfo->streamInfo.recoverScanFinished;
@@ -1070,8 +1081,11 @@ int32_t qStreamSetScanMemData(qTaskInfo_t tinfo, SPackedData submit) {
 }
 
 int32_t qStreamPrepareScan(qTaskInfo_t tinfo, STqOffsetVal* pOffset, int8_t subType) {
-  SExecTaskInfo* pTaskInfo = (SExecTaskInfo*)tinfo;
+  SExecTaskInfo*  pTaskInfo = (SExecTaskInfo*)tinfo;
-  SOperatorInfo* pOperator = pTaskInfo->pRoot;
+  SOperatorInfo*  pOperator = pTaskInfo->pRoot;
+  STableListInfo* pTableListInfo = pTaskInfo->pTableInfoList;
+  const char*     id = GET_TASKID(pTaskInfo);
+
   pTaskInfo->streamInfo.prepareStatus = *pOffset;
   pTaskInfo->streamInfo.returned = 0;
 
@@ -1084,21 +1098,24 @@ int32_t qStreamPrepareScan(qTaskInfo_t tinfo, STqOffsetVal* pOffset, int8_t subT
 
     // TODO add more check
     if (pOperator->operatorType != QUERY_NODE_PHYSICAL_PLAN_STREAM_SCAN) {
-      if(pOperator->numOfDownstream != 1){
+      if (pOperator->numOfDownstream != 1) {
-        qError("pOperator->numOfDownstream != 1:%d", pOperator->numOfDownstream);
+        qError("invalid operator, number of downstream:%d, %s", pOperator->numOfDownstream, id);
         return -1;
       }
       pOperator = pOperator->pDownstream[0];
     }
 
     SStreamScanInfo* pInfo = pOperator->info;
+    STableScanInfo*  pScanInfo = pInfo->pTableScanOp->info;
+    STableScanBase*  pScanBaseInfo = &pScanInfo->base;
+
     if (pOffset->type == TMQ_OFFSET__LOG) {
-      STableScanInfo* pTSInfo = pInfo->pTableScanOp->info;
+      tsdbReaderClose(pScanBaseInfo->dataReader);
-      tsdbReaderClose(pTSInfo->base.dataReader);
+      pScanBaseInfo->dataReader = NULL;
-      pTSInfo->base.dataReader = NULL;
+
       // let's seek to the next version in wal file
       if (tqSeekVer(pInfo->tqReader, pOffset->version + 1, pTaskInfo->id.str) < 0) {
-        qError("tqSeekVer failed ver:%" PRId64, pOffset->version + 1);
+        qError("tqSeekVer failed ver:%"PRId64", %s", pOffset->version + 1, id);
         return -1;
       }
     } else if (pOffset->type == TMQ_OFFSET__SNAPSHOT_DATA) {
@@ -1106,121 +1123,125 @@ int32_t qStreamPrepareScan(qTaskInfo_t tinfo, STqOffsetVal* pOffset, int8_t subT
       // those data are from the snapshot in tsdb, besides the data in the wal file.
       int64_t uid = pOffset->uid;
       int64_t ts = pOffset->ts;
+      int32_t index = 0;
+
+      // this value may be changed if new tables are created
+      taosRLockLatch(&pTaskInfo->lock);
+      int32_t numOfTables = tableListGetSize(pTableListInfo);
 
       if (uid == 0) {
-        if (tableListGetSize(pTaskInfo->pTableInfoList) != 0) {
+        if (numOfTables != 0) {
-          STableKeyInfo* pTableInfo = tableListGetInfo(pTaskInfo->pTableInfoList, 0);
+          STableKeyInfo* pTableInfo = tableListGetInfo(pTableListInfo, 0);
           uid = pTableInfo->uid;
           ts = INT64_MIN;
+          pScanInfo->currentTable = 0;
         } else {
-          qError("uid == 0 and tablelist size is 0");
+          taosRUnLockLatch(&pTaskInfo->lock);
+          qError("no table in table list, %s", id);
           return -1;
         }
       }
 
-      /*if (pTaskInfo->streamInfo.lastStatus.type != TMQ_OFFSET__SNAPSHOT_DATA ||*/
-      /*pTaskInfo->streamInfo.lastStatus.uid != uid || pTaskInfo->streamInfo.lastStatus.ts != ts) {*/
-      STableScanInfo* pTableScanInfo = pInfo->pTableScanOp->info;
-      int32_t         numOfTables = tableListGetSize(pTaskInfo->pTableInfoList);
-
-      qDebug("switch to table uid:%" PRId64 " ts:%" PRId64 "% "PRId64 " rows returned", uid, ts, pInfo->pTableScanOp->resultInfo.totalRows);
       pInfo->pTableScanOp->resultInfo.totalRows = 0;
 
-      bool found = false;
+      // start from current accessed position
-      for (int32_t i = 0; i < numOfTables; i++) {
+      // we cannot start from the pScanInfo->currentTable, since the commit offset may cause the rollback of the start
-        STableKeyInfo* pTableInfo = tableListGetInfo(pTaskInfo->pTableInfoList, i);
+      // position, let's find it from the beginning.
-        if (pTableInfo->uid == uid) {
+      index = tableListFind(pTableListInfo, uid, 0);
-          found = true;
+      taosRUnLockLatch(&pTaskInfo->lock);
-          pTableScanInfo->currentTable = i;
-          break;
-        }
-      }
 
-      // TODO after dropping table, table may not found
+      if (index >= 0) {
-      if(!found){
+        pScanInfo->currentTable = index;
-        qError("uid not found in tablelist %" PRId64, uid);
+      } else {
+        qError("vgId:%d uid:%" PRIu64 " not found in table list, total:%d, index:%d %s", pTaskInfo->id.vgId, uid,
+               numOfTables, pScanInfo->currentTable, id);
         return -1;
       }
 
-      if (pTableScanInfo->base.dataReader == NULL) {
+      STableKeyInfo keyInfo = {.uid = uid};
-        STableKeyInfo* pList = tableListGetInfo(pTaskInfo->pTableInfoList, 0);
+      int64_t oldSkey = pScanBaseInfo->cond.twindows.skey;
-        int32_t        num = tableListGetSize(pTaskInfo->pTableInfoList);
 
-        if (tsdbReaderOpen(pTableScanInfo->base.readHandle.vnode, &pTableScanInfo->base.cond, pList, num,
+      // let's start from the next ts that returned to consumer.
-                           pTableScanInfo->pResBlock, &pTableScanInfo->base.dataReader, NULL) < 0 ||
+      pScanBaseInfo->cond.twindows.skey = ts + 1;
-            pTableScanInfo->base.dataReader == NULL) {
+      pScanInfo->scanTimes = 0;
-          qError("tsdbReaderOpen failed. uid:%" PRIi64, pOffset->uid);
+
+      if (pScanBaseInfo->dataReader == NULL) {
+        int32_t code = tsdbReaderOpen(pScanBaseInfo->readHandle.vnode, &pScanBaseInfo->cond, &keyInfo, 1,
+                                      pScanInfo->pResBlock, &pScanBaseInfo->dataReader, id, false);
+        if (code != TSDB_CODE_SUCCESS) {
+          qError("prepare read tsdb snapshot failed, uid:%" PRId64 ", code:%s %s", pOffset->uid, tstrerror(code), id);
+          terrno = code;
           return -1;
         }
+
+        qDebug("tsdb reader created with offset(snapshot) uid:%" PRId64 " ts:%" PRId64 " table index:%d, total:%d, %s",
+               uid, pScanBaseInfo->cond.twindows.skey, pScanInfo->currentTable, numOfTables, id);
+      } else {
+        tsdbSetTableList(pScanBaseInfo->dataReader, &keyInfo, 1);
+        tsdbReaderReset(pScanBaseInfo->dataReader, &pScanBaseInfo->cond);
+        qDebug("tsdb reader offset seek snapshot to uid:%" PRId64 " ts %" PRId64 "  table index:%d numOfTable:%d, %s",
+               uid, pScanBaseInfo->cond.twindows.skey, pScanInfo->currentTable, numOfTables, id);
       }
 
-      STableKeyInfo tki = {.uid = uid};
+      // restore the key value
-      tsdbSetTableList(pTableScanInfo->base.dataReader, &tki, 1);
+      pScanBaseInfo->cond.twindows.skey = oldSkey;
-      int64_t oldSkey = pTableScanInfo->base.cond.twindows.skey;
-      pTableScanInfo->base.cond.twindows.skey = ts + 1;
-      tsdbReaderReset(pTableScanInfo->base.dataReader, &pTableScanInfo->base.cond);
-      pTableScanInfo->base.cond.twindows.skey = oldSkey;
-      pTableScanInfo->scanTimes = 0;
-
-      qDebug("tsdb reader offset seek snapshot to uid:%" PRId64 " ts %" PRId64 ", table cur set to %d , all table num %d", uid,
-             ts, pTableScanInfo->currentTable, numOfTables);
     } else {
-      qError("invalid pOffset->type:%d", pOffset->type);
+      qError("invalid pOffset->type:%d, %s", pOffset->type, id);
-      return -1;
-    }
-  } else if (pOffset->type == TMQ_OFFSET__SNAPSHOT_DATA) {
-    SStreamRawScanInfo* pInfo = pOperator->info;
-    SSnapContext*       sContext = pInfo->sContext;
-    if (setForSnapShot(sContext, pOffset->uid) != 0) {
-      qError("setDataForSnapShot error. uid:%" PRIi64, pOffset->uid);
       return -1;
     }
 
-    SMetaTableInfo mtInfo = getUidfromSnapShot(sContext);
+  } else {  // subType == TOPIC_SUB_TYPE__TABLE/TOPIC_SUB_TYPE__DB
-    tsdbReaderClose(pInfo->dataReader);
+    if (pOffset->type == TMQ_OFFSET__SNAPSHOT_DATA) {
-    pInfo->dataReader = NULL;
+      SStreamRawScanInfo* pInfo = pOperator->info;
+      SSnapContext*       sContext = pInfo->sContext;
+      if (setForSnapShot(sContext, pOffset->uid) != 0) {
+        qError("setDataForSnapShot error. uid:%" PRId64" , %s", pOffset->uid, id);
+        return -1;
+      }
 
-    cleanupQueryTableDataCond(&pTaskInfo->streamInfo.tableCond);
+      SMetaTableInfo mtInfo = getUidfromSnapShot(sContext);
-    tableListClear(pTaskInfo->pTableInfoList);
+      tsdbReaderClose(pInfo->dataReader);
+      pInfo->dataReader = NULL;
 
-    if (mtInfo.uid == 0) {
+      cleanupQueryTableDataCond(&pTaskInfo->streamInfo.tableCond);
-      return 0;  // no data
+      tableListClear(pTableListInfo);
+
+      if (mtInfo.uid == 0) {
+        return 0;  // no data
+      }
+
+      initQueryTableDataCondForTmq(&pTaskInfo->streamInfo.tableCond, sContext, &mtInfo);
+      pTaskInfo->streamInfo.tableCond.twindows.skey = pOffset->ts;
+
+      tableListAddTableInfo(pTableListInfo, mtInfo.uid, 0);
+
+      STableKeyInfo* pList = tableListGetInfo(pTableListInfo, 0);
+      int32_t        size = tableListGetSize(pTableListInfo);
+
+      tsdbReaderOpen(pInfo->vnode, &pTaskInfo->streamInfo.tableCond, pList, size, NULL, &pInfo->dataReader, NULL, false);
+
+      cleanupQueryTableDataCond(&pTaskInfo->streamInfo.tableCond);
+      strcpy(pTaskInfo->streamInfo.tbName, mtInfo.tbName);
+      tDeleteSSchemaWrapper(pTaskInfo->streamInfo.schema);
+      pTaskInfo->streamInfo.schema = mtInfo.schema;
+
+      qDebug("tmqsnap qStreamPrepareScan snapshot data uid:%" PRId64 " ts %" PRId64" %s", mtInfo.uid, pOffset->ts, id);
+    } else if (pOffset->type == TMQ_OFFSET__SNAPSHOT_META) {
+      SStreamRawScanInfo* pInfo = pOperator->info;
+      SSnapContext*       sContext = pInfo->sContext;
+      if (setForSnapShot(sContext, pOffset->uid) != 0) {
+        qError("setForSnapShot error. uid:%" PRIu64 " ,version:%" PRId64, pOffset->uid, pOffset->version);
+        return -1;
+      }
+      qDebug("tmqsnap qStreamPrepareScan snapshot meta uid:%" PRId64 " ts %" PRId64 " %s", pOffset->uid, pOffset->ts, id);
+    } else if (pOffset->type == TMQ_OFFSET__LOG) {
+      SStreamRawScanInfo* pInfo = pOperator->info;
+      tsdbReaderClose(pInfo->dataReader);
+      pInfo->dataReader = NULL;
+      qDebug("tmqsnap qStreamPrepareScan snapshot log, %s", id);
     }
-
-    initQueryTableDataCondForTmq(&pTaskInfo->streamInfo.tableCond, sContext, &mtInfo);
-    pTaskInfo->streamInfo.tableCond.twindows.skey = pOffset->ts;
-
-    if (pTaskInfo->pTableInfoList == NULL) {
-      pTaskInfo->pTableInfoList = tableListCreate();
-    }
-
-    tableListAddTableInfo(pTaskInfo->pTableInfoList, mtInfo.uid, 0);
-
-    STableKeyInfo* pList = tableListGetInfo(pTaskInfo->pTableInfoList, 0);
-    int32_t        size = tableListGetSize(pTaskInfo->pTableInfoList);
-
-    tsdbReaderOpen(pInfo->vnode, &pTaskInfo->streamInfo.tableCond, pList, size, NULL, &pInfo->dataReader, NULL);
-
-    cleanupQueryTableDataCond(&pTaskInfo->streamInfo.tableCond);
-    strcpy(pTaskInfo->streamInfo.tbName, mtInfo.tbName);
-    tDeleteSSchemaWrapper(pTaskInfo->streamInfo.schema);
-    pTaskInfo->streamInfo.schema = mtInfo.schema;
-
-    qDebug("tmqsnap qStreamPrepareScan snapshot data uid:%" PRId64 " ts %" PRId64, mtInfo.uid, pOffset->ts);
-  } else if (pOffset->type == TMQ_OFFSET__SNAPSHOT_META) {
-    SStreamRawScanInfo* pInfo = pOperator->info;
-    SSnapContext*       sContext = pInfo->sContext;
-    if (setForSnapShot(sContext, pOffset->uid) != 0) {
-      qError("setForSnapShot error. uid:%" PRIu64 " ,version:%" PRId64, pOffset->uid, pOffset->version);
-      return -1;
-    }
-    qDebug("tmqsnap qStreamPrepareScan snapshot meta uid:%" PRId64 " ts %" PRId64, pOffset->uid, pOffset->ts);
-  } else if (pOffset->type == TMQ_OFFSET__LOG) {
-    SStreamRawScanInfo* pInfo = pOperator->info;
-    tsdbReaderClose(pInfo->dataReader);
-    pInfo->dataReader = NULL;
-    qDebug("tmqsnap qStreamPrepareScan snapshot log");
   }
+
   return 0;
 }
 

source/libs/executor/src/executorimpl.c

@@ -633,7 +633,7 @@ void setBlockSMAInfo(SqlFunctionCtx* pCtx, SExprInfo* pExprInfo, SSDataBlock* pB
   }
 }
 
-bool isTaskKilled(SExecTaskInfo* pTaskInfo) { return (0 != pTaskInfo->code) ? true : false; }
+bool isTaskKilled(SExecTaskInfo* pTaskInfo) { return (0 != pTaskInfo->code);}
 
 void setTaskKilled(SExecTaskInfo* pTaskInfo, int32_t rspCode) { pTaskInfo->code = rspCode; }
 
@@ -1201,7 +1201,7 @@ int32_t doCopyToSDataBlock(SExecTaskInfo* pTaskInfo, SSDataBlock* pBlock, SExprS
 
     if (pBlock->info.rows + pRow->numOfRows > pBlock->info.capacity) {
       blockDataEnsureCapacity(pBlock, pBlock->info.rows + pRow->numOfRows);
-      qDebug("datablock capacity not sufficient, expand to required:%d, current capacity:%d, %s",
+      qDebug("datablock capacity not sufficient, expand to required:%" PRId64 ", current capacity:%d, %s",
              (pRow->numOfRows+pBlock->info.rows),
              pBlock->info.capacity, GET_TASKID(pTaskInfo));
         // todo set the pOperator->resultInfo size
@@ -1214,7 +1214,7 @@ int32_t doCopyToSDataBlock(SExecTaskInfo* pTaskInfo, SSDataBlock* pBlock, SExprS
     pBlock->info.rows += pRow->numOfRows;
   }
 
-  qDebug("%s result generated, rows:%d, groupId:%" PRIu64, GET_TASKID(pTaskInfo), pBlock->info.rows,
+  qDebug("%s result generated, rows:%" PRId64 ", groupId:%" PRIu64, GET_TASKID(pTaskInfo), pBlock->info.rows,
          pBlock->info.id.groupId);
   pBlock->info.dataLoad = 1;
   blockDataUpdateTsWindow(pBlock, 0);
@@ -1974,7 +1974,7 @@ char* buildTaskId(uint64_t taskId, uint64_t queryId) {
   return p;
 }
 
-static SExecTaskInfo* doCreateExecTaskInfo(uint64_t queryId, uint64_t taskId, int32_t vgId, EOPTR_EXEC_MODEL model, char* dbFName) {
+SExecTaskInfo* doCreateExecTaskInfo(uint64_t queryId, uint64_t taskId, int32_t vgId, EOPTR_EXEC_MODEL model, char* dbFName) {
   SExecTaskInfo* pTaskInfo = taosMemoryCalloc(1, sizeof(SExecTaskInfo));
   if (pTaskInfo == NULL) {
     terrno = TSDB_CODE_OUT_OF_MEMORY;
@@ -1982,6 +1982,7 @@ static SExecTaskInfo* doCreateExecTaskInfo(uint64_t queryId, uint64_t taskId, in
   }
 
   setTaskStatus(pTaskInfo, TASK_NOT_COMPLETED);
+  pTaskInfo->cost.created = taosGetTimestampUs();
 
   pTaskInfo->schemaInfo.dbname = taosStrdup(dbFName);
   pTaskInfo->execModel = model;
@@ -1989,6 +1990,7 @@ static SExecTaskInfo* doCreateExecTaskInfo(uint64_t queryId, uint64_t taskId, in
   pTaskInfo->stopInfo.pStopInfo = taosArrayInit(4, sizeof(SExchangeOpStopInfo));
   pTaskInfo->pResultBlockList = taosArrayInit(128, POINTER_BYTES);
 
+  taosInitRWLatch(&pTaskInfo->lock);
   pTaskInfo->id.vgId = vgId;
   pTaskInfo->id.queryId = queryId;
   pTaskInfo->id.str = buildTaskId(taskId, queryId);
@@ -2464,7 +2466,6 @@ int32_t createExecTaskInfo(SSubplan* pPlan, SExecTaskInfo** pTaskInfo, SReadHand
     goto _complete;
   }
 
-  (*pTaskInfo)->cost.created = taosGetTimestampUs();
   return TSDB_CODE_SUCCESS;
 
 _complete:

source/libs/executor/src/filloperator.c

@@ -447,9 +447,14 @@ void* destroyStreamFillSupporter(SStreamFillSupporter* pFillSup) {
   return NULL;
 }
 
+void destroySPoint(void* ptr) {
+  SPoint* point = (SPoint*) ptr;
+  taosMemoryFreeClear(point->val);
+}
+
 void* destroyStreamFillLinearInfo(SStreamFillLinearInfo* pFillLinear) {
-  taosArrayDestroy(pFillLinear->pDeltaVal);
+  taosArrayDestroyEx(pFillLinear->pEndPoints, destroySPoint);
-  taosArrayDestroy(pFillLinear->pNextDeltaVal);
+  taosArrayDestroyEx(pFillLinear->pNextEndPoints, destroySPoint);
   taosMemoryFree(pFillLinear);
   return NULL;
 }
@@ -611,19 +616,15 @@ static void calcDeltaData(SSDataBlock* pBlock, int32_t rowId, SResultRowData* pR
   }
 }
 
-static void calcRowDeltaData(SResultRowData* pStartRow, SResultRowData* pEndRow, SArray* pDelta, SFillColInfo* pFillCol,
+static void calcRowDeltaData(SResultRowData* pEndRow, SArray* pEndPoins, SFillColInfo* pFillCol,
-                             int32_t numOfCol, int32_t winCount) {
+                             int32_t numOfCol) {
   for (int32_t i = 0; i < numOfCol; i++) {
     if (!pFillCol[i].notFillCol) {
       int32_t          slotId = GET_DEST_SLOT_ID(pFillCol + i);
-      SResultCellData* pSCell = getResultCell(pStartRow, slotId);
-      double           start = 0.0;
-      GET_TYPED_DATA(start, double, pSCell->type, pSCell->pData);
       SResultCellData* pECell = getResultCell(pEndRow, slotId);
-      double           end = 0.0;
+      SPoint* pPoint = taosArrayGet(pEndPoins, slotId);
-      GET_TYPED_DATA(end, double, pECell->type, pECell->pData);
+      pPoint->key = pEndRow->key;
-      double delta = (end - start) / winCount;
+      memcpy(pPoint->val, pECell->pData, pECell->bytes);
-      taosArraySet(pDelta, slotId, &delta);
     }
   }
 }
@@ -674,10 +675,8 @@ void setDeleteFillValueInfo(TSKEY start, TSKEY end, SStreamFillSupporter* pFillS
       setFillKeyInfo(pFillSup->prev.key, pFillSup->next.key, &pFillSup->interval, pFillInfo);
       pFillInfo->pLinearInfo->hasNext = false;
       pFillInfo->pLinearInfo->nextEnd = INT64_MIN;
-      int32_t numOfWins = taosTimeCountInterval(pFillSup->prev.key, pFillSup->next.key, pFillSup->interval.sliding,
+      calcRowDeltaData(&pFillSup->next, pFillInfo->pLinearInfo->pEndPoints, pFillSup->pAllColInfo,
-                                                pFillSup->interval.slidingUnit, pFillSup->interval.precision);
+                       pFillSup->numOfAllCols);
-      calcRowDeltaData(&pFillSup->prev, &pFillSup->next, pFillInfo->pLinearInfo->pDeltaVal, pFillSup->pAllColInfo,
-                       pFillSup->numOfAllCols, numOfWins);
       pFillInfo->pResRow = &pFillSup->prev;
       pFillInfo->pLinearInfo->winIndex = 0;
     } break;
@@ -780,25 +779,19 @@ void setFillValueInfo(SSDataBlock* pBlock, TSKEY ts, int32_t rowId, SStreamFillS
         setFillKeyInfo(prevWKey, ts, &pFillSup->interval, pFillInfo);
         pFillInfo->pos = FILL_POS_MID;
         pFillInfo->pLinearInfo->nextEnd = nextWKey;
-        int32_t numOfWins = taosTimeCountInterval(prevWKey, ts, pFillSup->interval.sliding,
+        calcRowDeltaData(&pFillSup->cur, pFillInfo->pLinearInfo->pEndPoints, pFillSup->pAllColInfo,
-                                                  pFillSup->interval.slidingUnit, pFillSup->interval.precision);
+                         pFillSup->numOfAllCols);
-        calcRowDeltaData(&pFillSup->prev, &pFillSup->cur, pFillInfo->pLinearInfo->pDeltaVal, pFillSup->pAllColInfo,
-                         pFillSup->numOfAllCols, numOfWins);
         pFillInfo->pResRow = &pFillSup->prev;
 
-        numOfWins = taosTimeCountInterval(ts, nextWKey, pFillSup->interval.sliding, pFillSup->interval.slidingUnit,
+        calcRowDeltaData(&pFillSup->next, pFillInfo->pLinearInfo->pNextEndPoints, pFillSup->pAllColInfo,
-                                          pFillSup->interval.precision);
+                         pFillSup->numOfAllCols);
-        calcRowDeltaData(&pFillSup->cur, &pFillSup->next, pFillInfo->pLinearInfo->pNextDeltaVal, pFillSup->pAllColInfo,
-                         pFillSup->numOfAllCols, numOfWins);
         pFillInfo->pLinearInfo->hasNext = true;
       } else if (hasPrevWindow(pFillSup)) {
         setFillKeyInfo(prevWKey, ts, &pFillSup->interval, pFillInfo);
         pFillInfo->pos = FILL_POS_END;
         pFillInfo->pLinearInfo->nextEnd = INT64_MIN;
-        int32_t numOfWins = taosTimeCountInterval(prevWKey, ts, pFillSup->interval.sliding,
+        calcRowDeltaData(&pFillSup->cur, pFillInfo->pLinearInfo->pEndPoints, pFillSup->pAllColInfo,
-                                                  pFillSup->interval.slidingUnit, pFillSup->interval.precision);
+                         pFillSup->numOfAllCols);
-        calcRowDeltaData(&pFillSup->prev, &pFillSup->cur, pFillInfo->pLinearInfo->pDeltaVal, pFillSup->pAllColInfo,
-                         pFillSup->numOfAllCols, numOfWins);
         pFillInfo->pResRow = &pFillSup->prev;
         pFillInfo->pLinearInfo->hasNext = false;
       } else {
@@ -806,10 +799,8 @@ void setFillValueInfo(SSDataBlock* pBlock, TSKEY ts, int32_t rowId, SStreamFillS
         setFillKeyInfo(ts, nextWKey, &pFillSup->interval, pFillInfo);
         pFillInfo->pos = FILL_POS_START;
         pFillInfo->pLinearInfo->nextEnd = INT64_MIN;
-        int32_t numOfWins = taosTimeCountInterval(ts, nextWKey, pFillSup->interval.sliding,
+        calcRowDeltaData(&pFillSup->next, pFillInfo->pLinearInfo->pEndPoints, pFillSup->pAllColInfo,
-                                                  pFillSup->interval.slidingUnit, pFillSup->interval.precision);
+                         pFillSup->numOfAllCols);
-        calcRowDeltaData(&pFillSup->cur, &pFillSup->next, pFillInfo->pLinearInfo->pDeltaVal, pFillSup->pAllColInfo,
-                         pFillSup->numOfAllCols, numOfWins);
         pFillInfo->pResRow = &pFillSup->cur;
         pFillInfo->pLinearInfo->hasNext = false;
       }
@@ -906,13 +897,18 @@ static void doStreamFillLinear(SStreamFillSupporter* pFillSup, SStreamFillInfo*
           colDataSetNULL(pColData, index);
           continue;
         }
-        double* pDelta = taosArrayGet(pFillInfo->pLinearInfo->pDeltaVal, slotId);
+        SPoint* pEnd = taosArrayGet(pFillInfo->pLinearInfo->pEndPoints, slotId);
         double  vCell = 0;
-        GET_TYPED_DATA(vCell, double, pCell->type, pCell->pData);
+        SPoint start = {0};
-        vCell += (*pDelta) * pFillInfo->pLinearInfo->winIndex;
+        start.key = pFillInfo->pResRow->key;
-        int64_t result = 0;
+        start.val = pCell->pData;
-        SET_TYPED_DATA(&result, pCell->type, vCell);
+
-        colDataSetVal(pColData, index, (const char*)&result, false);
+        SPoint cur = {0};
+        cur.key = pFillInfo->current;
+        cur.val = taosMemoryCalloc(1, pCell->bytes);
+        taosGetLinearInterpolationVal(&cur, pCell->type, &start, pEnd, pCell->type);
+        colDataSetVal(pColData, index, (const char*)cur.val, false);
+        destroySPoint(&cur);
       }
     }
     pFillInfo->current = taosTimeAdd(pFillInfo->current, pFillSup->interval.sliding, pFillSup->interval.slidingUnit,
@@ -953,8 +949,7 @@ static void doStreamFillRange(SStreamFillInfo* pFillInfo, SStreamFillSupporter*
     if (pFillInfo->current > pFillInfo->end && pFillInfo->pLinearInfo->hasNext) {
       pFillInfo->pLinearInfo->hasNext = false;
       pFillInfo->pLinearInfo->winIndex = 0;
-      taosArrayClear(pFillInfo->pLinearInfo->pDeltaVal);
+      taosArraySwap(pFillInfo->pLinearInfo->pEndPoints, pFillInfo->pLinearInfo->pNextEndPoints);
-      taosArrayAddAll(pFillInfo->pLinearInfo->pDeltaVal, pFillInfo->pLinearInfo->pNextDeltaVal);
       pFillInfo->pResRow = &pFillSup->cur;
       setFillKeyInfo(pFillSup->cur.key, pFillInfo->pLinearInfo->nextEnd, &pFillSup->interval, pFillInfo);
       doStreamFillLinear(pFillSup, pFillInfo, pRes);
@@ -1359,15 +1354,19 @@ SStreamFillInfo* initStreamFillInfo(SStreamFillSupporter* pFillSup, SSDataBlock*
   pFillInfo->pLinearInfo = taosMemoryCalloc(1, sizeof(SStreamFillLinearInfo));
   pFillInfo->pLinearInfo->hasNext = false;
   pFillInfo->pLinearInfo->nextEnd = INT64_MIN;
-  pFillInfo->pLinearInfo->pDeltaVal = NULL;
+  pFillInfo->pLinearInfo->pEndPoints = NULL;
-  pFillInfo->pLinearInfo->pNextDeltaVal = NULL;
+  pFillInfo->pLinearInfo->pNextEndPoints = NULL;
   if (pFillSup->type == TSDB_FILL_LINEAR) {
-    pFillInfo->pLinearInfo->pDeltaVal = taosArrayInit(pFillSup->numOfAllCols, sizeof(double));
+    pFillInfo->pLinearInfo->pEndPoints = taosArrayInit(pFillSup->numOfAllCols, sizeof(SPoint));
-    pFillInfo->pLinearInfo->pNextDeltaVal = taosArrayInit(pFillSup->numOfAllCols, sizeof(double));
+    pFillInfo->pLinearInfo->pNextEndPoints = taosArrayInit(pFillSup->numOfAllCols, sizeof(SPoint));
     for (int32_t i = 0; i < pFillSup->numOfAllCols; i++) {
-      double value = 0.0;
+      SColumnInfoData* pColData = taosArrayGet(pRes->pDataBlock, i);
-      taosArrayPush(pFillInfo->pLinearInfo->pDeltaVal, &value);
+      SPoint value = {0};
-      taosArrayPush(pFillInfo->pLinearInfo->pNextDeltaVal, &value);
+      value.val = taosMemoryCalloc(1, pColData->info.bytes);
+      taosArrayPush(pFillInfo->pLinearInfo->pEndPoints, &value);
+
+      value.val = taosMemoryCalloc(1, pColData->info.bytes);
+      taosArrayPush(pFillInfo->pLinearInfo->pNextEndPoints, &value);
     }
   }
   pFillInfo->pLinearInfo->winIndex = 0;

source/libs/executor/src/projectoperator.c

@@ -271,7 +271,7 @@ SSDataBlock* doProjectOperation(SOperatorInfo* pOperator) {
             return NULL;
           }
         }
-        qDebug("set op close, exec %d, status %d rows %d", pTaskInfo->execModel, pOperator->status,
+        qDebug("set op close, exec %d, status %d rows %" PRId64 , pTaskInfo->execModel, pOperator->status,
                pFinalRes->info.rows);
         setOperatorCompleted(pOperator);
         break;
@@ -337,7 +337,7 @@ SSDataBlock* doProjectOperation(SOperatorInfo* pOperator) {
 
       // when apply the limit/offset for each group, pRes->info.rows may be 0, due to limit constraint.
       if (pFinalRes->info.rows > 0 || (pOperator->status == OP_EXEC_DONE)) {
-        qDebug("project return %d rows, status %d", pFinalRes->info.rows, pOperator->status);
+        qDebug("project return %" PRId64 " rows, status %d", pFinalRes->info.rows, pOperator->status);
         break;
       }
     } else {

source/libs/executor/src/scanoperator.c

@@ -31,6 +31,9 @@
 #include "thash.h"
 #include "ttypes.h"
 
+int32_t scanDebug = 0;
+
+
 #define MULTI_READER_MAX_TABLE_NUM 5000
 #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))
@@ -308,14 +311,14 @@ static int32_t loadDataBlock(SOperatorInfo* pOperator, STableScanBase* pTableSca
   taosMemoryFreeClear(pBlock->pBlockAgg);
 
   if (*status == FUNC_DATA_REQUIRED_FILTEROUT) {
-    qDebug("%s data block filter out, brange:%" PRId64 "-%" PRId64 ", rows:%d", GET_TASKID(pTaskInfo),
+    qDebug("%s data block filter out, brange:%" PRId64 "-%" PRId64 ", rows:%" PRId64 , GET_TASKID(pTaskInfo),
            pBlockInfo->window.skey, pBlockInfo->window.ekey, pBlockInfo->rows);
     pCost->filterOutBlocks += 1;
     pCost->totalRows += pBlock->info.rows;
     tsdbReleaseDataBlock(pTableScanInfo->dataReader);
     return TSDB_CODE_SUCCESS;
   } else if (*status == FUNC_DATA_REQUIRED_NOT_LOAD) {
-    qDebug("%s data block skipped, brange:%" PRId64 "-%" PRId64 ", rows:%d, uid:%" PRIu64, GET_TASKID(pTaskInfo),
+    qDebug("%s data block skipped, brange:%" PRId64 "-%" PRId64 ", rows:%" PRId64 ", uid:%" PRIu64, GET_TASKID(pTaskInfo),
            pBlockInfo->window.skey, pBlockInfo->window.ekey, pBlockInfo->rows, pBlockInfo->id.uid);
     doSetTagColumnData(pTableScanInfo, pBlock, pTaskInfo, 1);
     pCost->skipBlocks += 1;
@@ -326,7 +329,7 @@ static int32_t loadDataBlock(SOperatorInfo* pOperator, STableScanBase* pTableSca
     loadSMA = true;  // mark the operation of load sma;
     bool success = doLoadBlockSMA(pTableScanInfo, pBlock, pTaskInfo);
     if (success) {  // failed to load the block sma data, data block statistics does not exist, load data block instead
-      qDebug("%s data block SMA loaded, brange:%" PRId64 "-%" PRId64 ", rows:%d", GET_TASKID(pTaskInfo),
+      qDebug("%s data block SMA loaded, brange:%" PRId64 "-%" PRId64 ", rows:%" PRId64 , GET_TASKID(pTaskInfo),
              pBlockInfo->window.skey, pBlockInfo->window.ekey, pBlockInfo->rows);
       doSetTagColumnData(pTableScanInfo, pBlock, pTaskInfo, 1);
       tsdbReleaseDataBlock(pTableScanInfo->dataReader);
@@ -346,7 +349,7 @@ static int32_t loadDataBlock(SOperatorInfo* pOperator, STableScanBase* pTableSca
       size_t size = taosArrayGetSize(pBlock->pDataBlock);
       bool   keep = doFilterByBlockSMA(pOperator->exprSupp.pFilterInfo, pBlock->pBlockAgg, size, pBlockInfo->rows);
       if (!keep) {
-        qDebug("%s data block filter out by block SMA, brange:%" PRId64 "-%" PRId64 ", rows:%d", GET_TASKID(pTaskInfo),
+        qDebug("%s data block filter out by block SMA, brange:%" PRId64 "-%" PRId64 ", rows:%" PRId64 , GET_TASKID(pTaskInfo),
                pBlockInfo->window.skey, pBlockInfo->window.ekey, pBlockInfo->rows);
         pCost->filterOutBlocks += 1;
         (*status) = FUNC_DATA_REQUIRED_FILTEROUT;
@@ -363,7 +366,7 @@ static int32_t loadDataBlock(SOperatorInfo* pOperator, STableScanBase* pTableSca
   // try to filter data block according to current results
   doDynamicPruneDataBlock(pOperator, pBlockInfo, status);
   if (*status == FUNC_DATA_REQUIRED_NOT_LOAD) {
-    qDebug("%s data block skipped due to dynamic prune, brange:%" PRId64 "-%" PRId64 ", rows:%d", GET_TASKID(pTaskInfo),
+    qDebug("%s data block skipped due to dynamic prune, brange:%" PRId64 "-%" PRId64 ", rows:%" PRId64 , GET_TASKID(pTaskInfo),
            pBlockInfo->window.skey, pBlockInfo->window.ekey, pBlockInfo->rows);
     pCost->skipBlocks += 1;
     tsdbReleaseDataBlock(pTableScanInfo->dataReader);
@@ -394,7 +397,7 @@ static int32_t loadDataBlock(SOperatorInfo* pOperator, STableScanBase* pTableSca
 
     if (pBlock->info.rows == 0) {
       pCost->filterOutBlocks += 1;
-      qDebug("%s data block filter out, brange:%" PRId64 "-%" PRId64 ", rows:%d, elapsed time:%.2f ms",
+      qDebug("%s data block filter out, brange:%" PRId64 "-%" PRId64 ", rows:%" PRId64 ", elapsed time:%.2f ms",
              GET_TASKID(pTaskInfo), pBlockInfo->window.skey, pBlockInfo->window.ekey, pBlockInfo->rows, el);
     } else {
       qDebug("%s data block filter applied, elapsed time:%.2f ms", GET_TASKID(pTaskInfo), el);
@@ -584,10 +587,16 @@ int32_t addTagPseudoColumnData(SReadHandle* pHandle, const SExprInfo* pExpr, int
       if (isNullVal) {
         colDataSetNNULL(pColInfoData, 0, pBlock->info.rows);
       } else if (pColInfoData->info.type != TSDB_DATA_TYPE_JSON) {
-        colDataSetNItems(pColInfoData, 0, data, pBlock->info.rows);
+        code = colDataSetNItems(pColInfoData, 0, data, pBlock->info.rows, false);
         if (IS_VAR_DATA_TYPE(((const STagVal*)p)->type)) {
           taosMemoryFree(data);
         }
+        if (code) {
+          if (freeReader) {
+            metaReaderClear(&mr);
+          }
+          return code;
+        }
       } else {  // todo opt for json tag
         for (int32_t i = 0; i < pBlock->info.rows; ++i) {
           colDataSetVal(pColInfoData, i, data, false);
@@ -634,10 +643,22 @@ static SSDataBlock* doTableScanImpl(SOperatorInfo* pOperator) {
   STableScanInfo* pTableScanInfo = pOperator->info;
   SExecTaskInfo*  pTaskInfo = pOperator->pTaskInfo;
   SSDataBlock*    pBlock = pTableScanInfo->pResBlock;
+  bool            hasNext = false;
+  int32_t         code = TSDB_CODE_SUCCESS;
 
   int64_t st = taosGetTimestampUs();
 
-  while (tsdbNextDataBlock(pTableScanInfo->base.dataReader)) {
+  while (true) {
+    code = tsdbNextDataBlock(pTableScanInfo->base.dataReader, &hasNext);
+    if (code) {
+      tsdbReleaseDataBlock(pTableScanInfo->base.dataReader);
+      T_LONG_JMP(pTaskInfo->env, code);
+    }
+
+    if (!hasNext) {
+      break;
+    }
+    
     if (isTaskKilled(pTaskInfo)) {
       tsdbReleaseDataBlock(pTableScanInfo->base.dataReader);
       T_LONG_JMP(pTaskInfo->env, pTaskInfo->code);
@@ -654,9 +675,10 @@ static SSDataBlock* doTableScanImpl(SOperatorInfo* pOperator) {
       continue;
     }
 
-    ASSERT(pBlock->info.id.uid != 0);
+    if (pBlock->info.id.uid) {
-    pBlock->info.id.groupId = getTableGroupId(pTaskInfo->pTableInfoList, pBlock->info.id.uid);
+      pBlock->info.id.groupId = getTableGroupId(pTaskInfo->pTableInfoList, pBlock->info.id.uid);
-
+    }
+    
     uint32_t status = 0;
     int32_t  code = loadDataBlock(pOperator, &pTableScanInfo->base, pBlock, &status);
     if (code != TSDB_CODE_SUCCESS) {
@@ -680,7 +702,6 @@ static SSDataBlock* doTableScanImpl(SOperatorInfo* pOperator) {
     pTaskInfo->streamInfo.lastStatus.uid = pBlock->info.id.uid;
     pTaskInfo->streamInfo.lastStatus.ts = pBlock->info.window.ekey;
 
-    ASSERT(pBlock->info.id.uid != 0);
     return pBlock;
   }
   return NULL;
@@ -761,13 +782,14 @@ static SSDataBlock* doTableScan(SOperatorInfo* pOperator) {
       // if no data, switch to next table and continue scan
       pInfo->currentTable++;
       if (pInfo->currentTable >= numOfTables) {
+        qDebug("all table checked in table list, total:%d, return NULL, %s", numOfTables, GET_TASKID(pTaskInfo));
         return NULL;
       }
 
       STableKeyInfo* pTableInfo = tableListGetInfo(pTaskInfo->pTableInfoList, pInfo->currentTable);
       tsdbSetTableList(pInfo->base.dataReader, pTableInfo, 1);
-      qDebug("set uid:%" PRIu64 " into scanner, total tables:%d, index:%d %s", pTableInfo->uid, numOfTables,
+      qDebug("set uid:%" PRIu64 " into scanner, total tables:%d, index:%d/%d %s", pTableInfo->uid, numOfTables,
-             pInfo->currentTable, pTaskInfo->id.str);
+             pInfo->currentTable, numOfTables, GET_TASKID(pTaskInfo));
 
       tsdbReaderReset(pInfo->base.dataReader, &pInfo->base.cond);
       pInfo->scanTimes = 0;
@@ -785,7 +807,7 @@ static SSDataBlock* doTableScan(SOperatorInfo* pOperator) {
       ASSERT(pInfo->base.dataReader == NULL);
 
       int32_t code = tsdbReaderOpen(pInfo->base.readHandle.vnode, &pInfo->base.cond, pList, num, pInfo->pResBlock,
-                                    (STsdbReader**)&pInfo->base.dataReader, GET_TASKID(pTaskInfo));
+                                    (STsdbReader**)&pInfo->base.dataReader, GET_TASKID(pTaskInfo), pInfo->countOnly);
       if (code != TSDB_CODE_SUCCESS) {
         T_LONG_JMP(pTaskInfo->env, code);
       }
@@ -797,7 +819,6 @@ static SSDataBlock* doTableScan(SOperatorInfo* pOperator) {
 
     SSDataBlock* result = doGroupedTableScan(pOperator);
     if (result != NULL) {
-      ASSERT(result->info.id.uid != 0);
       return result;
     }
 
@@ -917,6 +938,10 @@ SOperatorInfo* createTableScanOperatorInfo(STableScanPhysiNode* pTableScanNode,
     goto _error;
   }
 
+  if (scanDebug) {
+    pInfo->countOnly = true;
+  }
+
   taosLRUCacheSetStrictCapacity(pInfo->base.metaCache.pTableMetaEntryCache, false);
   pOperator->fpSet = createOperatorFpSet(optrDummyOpenFn, doTableScan, NULL, destroyTableScanOperatorInfo,
                                          optrDefaultBufFn, getTableScannerExecInfo);
@@ -1008,21 +1033,29 @@ static SSDataBlock* readPreVersionData(SOperatorInfo* pTableScanOp, uint64_t tbU
   SSDataBlock* pBlock = pTableScanInfo->pResBlock;
   STsdbReader* pReader = NULL;
   int32_t      code = tsdbReaderOpen(pTableScanInfo->base.readHandle.vnode, &cond, &tblInfo, 1, pBlock,
-                                     (STsdbReader**)&pReader, GET_TASKID(pTaskInfo));
+                                     (STsdbReader**)&pReader, GET_TASKID(pTaskInfo), false);
   if (code != TSDB_CODE_SUCCESS) {
     terrno = code;
     T_LONG_JMP(pTaskInfo->env, code);
     return NULL;
   }
 
-  if (tsdbNextDataBlock(pReader)) {
+  bool hasNext = false;
+  code = tsdbNextDataBlock(pReader, &hasNext);
+  if (code != TSDB_CODE_SUCCESS) {
+    terrno = code;
+    T_LONG_JMP(pTaskInfo->env, code);
+    return NULL;
+  }
+
+  if (hasNext) {
     /*SSDataBlock* p = */ tsdbRetrieveDataBlock(pReader, NULL);
     doSetTagColumnData(&pTableScanInfo->base, pBlock, pTaskInfo, pBlock->info.rows);
     pBlock->info.id.groupId = getTableGroupId(pTaskInfo->pTableInfoList, pBlock->info.id.uid);
   }
 
   tsdbReaderClose(pReader);
-  qDebug("retrieve prev rows:%d, skey:%" PRId64 ", ekey:%" PRId64 " uid:%" PRIu64 ", max ver:%" PRId64
+  qDebug("retrieve prev rows:%" PRId64 ", skey:%" PRId64 ", ekey:%" PRId64 " uid:%" PRIu64 ", max ver:%" PRId64
          ", suid:%" PRIu64,
          pBlock->info.rows, startTs, endTs, tbUid, maxVersion, cond.suid);
 
@@ -1569,19 +1602,16 @@ static int32_t setBlockIntoRes(SStreamScanInfo* pInfo, const SSDataBlock* pBlock
 static SSDataBlock* doQueueScan(SOperatorInfo* pOperator) {
   SExecTaskInfo*   pTaskInfo = pOperator->pTaskInfo;
   SStreamScanInfo* pInfo = pOperator->info;
+  const char* id = GET_TASKID(pTaskInfo);
 
-  qDebug("start to exec queue scan");
+  qDebug("start to exec queue scan, %s", id);
 
   if (pTaskInfo->streamInfo.submit.msgStr != NULL) {
-    if (pInfo->tqReader->msg2.msgStr == NULL) {
-      /*pInfo->tqReader->pMsg = pTaskInfo->streamInfo.pReq;*/
 
-      /*const SSubmitReq* pSubmit = pInfo->tqReader->pMsg;*/
+    if (pInfo->tqReader->msg2.msgStr == NULL) {
-      /*if (tqReaderSetDataMsg(pInfo->tqReader, pSubmit, 0) < 0) {*/
-      /*void* msgStr = pTaskInfo->streamInfo.*/
       SPackedData submit = pTaskInfo->streamInfo.submit;
       if (tqReaderSetSubmitReq2(pInfo->tqReader, submit.msgStr, submit.msgLen, submit.ver) < 0) {
-        qError("submit msg messed up when initing stream submit block %p", submit.msgStr);
+        qError("submit msg messed up when initing stream submit block %p, %s", submit.msgStr, id);
         pInfo->tqReader->msg2 = (SPackedData){0};
         pInfo->tqReader->setMsg = 0;
         ASSERT(0);
@@ -1615,18 +1645,20 @@ static SSDataBlock* doQueueScan(SOperatorInfo* pOperator) {
   if (pTaskInfo->streamInfo.prepareStatus.type == TMQ_OFFSET__SNAPSHOT_DATA) {
     SSDataBlock* pResult = doTableScan(pInfo->pTableScanOp);
     if (pResult && pResult->info.rows > 0) {
-      qDebug("queue scan tsdb return %d rows min:%" PRId64 " max:%" PRId64 " wal curVersion:%" PRId64, pResult->info.rows,
+      qDebug("queue scan tsdb return %" PRId64 " rows min:%" PRId64 " max:%" PRId64 " wal curVersion:%" PRId64" %s", pResult->info.rows,
-             pResult->info.window.skey, pResult->info.window.ekey, pInfo->tqReader->pWalReader->curVersion);
+             pResult->info.window.skey, pResult->info.window.ekey, pInfo->tqReader->pWalReader->curVersion, id);
       pTaskInfo->streamInfo.returned = 1;
       return pResult;
     } else {
+      // no data has return already, try to extract data in the WAL
       if (!pTaskInfo->streamInfo.returned) {
         STableScanInfo* pTSInfo = pInfo->pTableScanOp->info;
         tsdbReaderClose(pTSInfo->base.dataReader);
-        qDebug("3");
+
         pTSInfo->base.dataReader = NULL;
         tqOffsetResetToLog(&pTaskInfo->streamInfo.prepareStatus, pTaskInfo->streamInfo.snapshotVer);
-        qDebug("queue scan tsdb over, switch to wal ver %" PRId64 "", pTaskInfo->streamInfo.snapshotVer + 1);
+
+        qDebug("queue scan tsdb over, switch to wal ver:%" PRId64 " %s", pTaskInfo->streamInfo.snapshotVer + 1, id);
         if (tqSeekVer(pInfo->tqReader, pTaskInfo->streamInfo.snapshotVer + 1, pTaskInfo->id.str) < 0) {
           tqOffsetResetToLog(&pTaskInfo->streamInfo.lastStatus, pTaskInfo->streamInfo.snapshotVer);
           return NULL;
@@ -1643,7 +1675,7 @@ static SSDataBlock* doQueueScan(SOperatorInfo* pOperator) {
       if (tqNextBlock(pInfo->tqReader, &ret) < 0) {
         // if the end is reached, terrno is 0
         if (terrno != 0) {
-          qError("failed to get next log block since %s", terrstr());
+          qError("failed to get next log block since %s, %s", terrstr(), id);
         }
       }
 
@@ -1652,15 +1684,12 @@ static SSDataBlock* doQueueScan(SOperatorInfo* pOperator) {
         setBlockIntoRes(pInfo, &ret.data, true);
         if (pInfo->pRes->info.rows > 0) {
           pOperator->status = OP_EXEC_RECV;
-          qDebug("queue scan log return %d rows", pInfo->pRes->info.rows);
+          qDebug("queue scan log return %" PRId64 " rows", pInfo->pRes->info.rows);
           return pInfo->pRes;
         }
       } else if (ret.fetchType == FETCH_TYPE__META) {
         qError("unexpected ret.fetchType:%d", ret.fetchType);
         continue;
-        //        pTaskInfo->streamInfo.lastStatus = ret.offset;
-        //        pTaskInfo->streamInfo.metaBlk = ret.meta;
-        //        return NULL;
       } else if (ret.fetchType == FETCH_TYPE__NONE ||
                  (ret.fetchType == FETCH_TYPE__SEP && pOperator->status == OP_EXEC_RECV)) {
         pTaskInfo->streamInfo.lastStatus = ret.offset;
@@ -1672,7 +1701,7 @@ static SSDataBlock* doQueueScan(SOperatorInfo* pOperator) {
       }
     }
   } else {
-    qError("unexpected streamInfo prepare type: %d", pTaskInfo->streamInfo.prepareStatus.type);
+    qError("unexpected streamInfo prepare type: %d %s", pTaskInfo->streamInfo.prepareStatus.type, id);
     return NULL;
   }
 }
@@ -1803,8 +1832,18 @@ static SSDataBlock* doStreamScan(SOperatorInfo* pOperator) {
         generateScanRange(pInfo, pInfo->pUpdateDataRes, pInfo->pUpdateRes);
         prepareRangeScan(pInfo, pInfo->pUpdateRes, &pInfo->updateResIndex);
         pInfo->scanMode = STREAM_SCAN_FROM_DATAREADER_RANGE;
+        printDataBlock(pInfo->pUpdateRes, "recover update");
         return pInfo->pUpdateRes;
       } break;
+      case STREAM_SCAN_FROM_DELETE_DATA: {
+        generateScanRange(pInfo, pInfo->pUpdateDataRes, pInfo->pUpdateRes);
+        prepareRangeScan(pInfo, pInfo->pUpdateRes, &pInfo->updateResIndex);
+        pInfo->scanMode = STREAM_SCAN_FROM_DATAREADER_RANGE;
+        copyDataBlock(pInfo->pDeleteDataRes, pInfo->pUpdateRes);
+        pInfo->pDeleteDataRes->info.type = STREAM_DELETE_DATA;
+        printDataBlock(pInfo->pDeleteDataRes, "recover delete");
+        return pInfo->pDeleteDataRes;
+      } break;
       case STREAM_SCAN_FROM_DATAREADER_RANGE: {
         SSDataBlock* pSDB = doRangeScan(pInfo, pInfo->pUpdateRes, pInfo->primaryTsIndex, &pInfo->updateResIndex);
         if (pSDB) {
@@ -1813,7 +1852,7 @@ static SSDataBlock* doStreamScan(SOperatorInfo* pOperator) {
           updateInfoSetScanRange(pInfo->pUpdateInfo, &pTableScanInfo->base.cond.twindows, pInfo->groupId, version);
           pSDB->info.type = pInfo->scanMode == STREAM_SCAN_FROM_DATAREADER_RANGE ? STREAM_NORMAL : STREAM_PULL_DATA;
           checkUpdateData(pInfo, true, pSDB, false);
-          // printDataBlock(pSDB, "stream scan update");
+          printDataBlock(pSDB, "scan recover update");
           calBlockTbName(pInfo, pSDB);
           return pSDB;
         }
@@ -1838,9 +1877,10 @@ static SSDataBlock* doStreamScan(SOperatorInfo* pOperator) {
       }
       if (pInfo->pCreateTbRes->info.rows > 0) {
         pInfo->scanMode = STREAM_SCAN_FROM_RES;
+        printDataBlock(pInfo->pCreateTbRes, "recover createTbl");
         return pInfo->pCreateTbRes;
       }
-      qDebug("stream recover scan get block, rows %d", pInfo->pRecoverRes->info.rows);
+      qDebug("stream recover scan get block, rows %" PRId64 , pInfo->pRecoverRes->info.rows);
       printDataBlock(pInfo->pRecoverRes, "scan recover");
       return pInfo->pRecoverRes;
     }
@@ -1996,6 +2036,7 @@ FETCH_NEXT_BLOCK:
       copyDataBlock(pInfo->pUpdateRes, pSup->pScanBlock);
       blockDataCleanup(pSup->pScanBlock);
       prepareRangeScan(pInfo, pInfo->pUpdateRes, &pInfo->updateResIndex);
+      pInfo->pUpdateRes->info.type = STREAM_DELETE_DATA;
       return pInfo->pUpdateRes;
     }
 
@@ -2069,7 +2110,7 @@ FETCH_NEXT_BLOCK:
     pOperator->resultInfo.totalRows += pBlockInfo->rows;
     // printDataBlock(pInfo->pRes, "stream scan");
 
-    qDebug("scan rows: %d", pBlockInfo->rows);
+    qDebug("scan rows: %" PRId64 , pBlockInfo->rows);
     if (pBlockInfo->rows > 0) {
       return pInfo->pRes;
     }
@@ -2102,12 +2143,22 @@ static SSDataBlock* doRawScan(SOperatorInfo* pOperator) {
   // NOTE: this operator does never check if current status is done or not
   SExecTaskInfo*      pTaskInfo = pOperator->pTaskInfo;
   SStreamRawScanInfo* pInfo = pOperator->info;
+  int32_t             code = TSDB_CODE_SUCCESS;
   pTaskInfo->streamInfo.metaRsp.metaRspLen = 0;  // use metaRspLen !=0 to judge if data is meta
   pTaskInfo->streamInfo.metaRsp.metaRsp = NULL;
 
   qDebug("tmqsnap doRawScan called");
   if (pTaskInfo->streamInfo.prepareStatus.type == TMQ_OFFSET__SNAPSHOT_DATA) {
-    if (pInfo->dataReader && tsdbNextDataBlock(pInfo->dataReader)) {
+    bool hasNext = false;
+    if (pInfo->dataReader) {
+      code = tsdbNextDataBlock(pInfo->dataReader, &hasNext);
+      if (code) {
+        tsdbReleaseDataBlock(pInfo->dataReader);
+        longjmp(pTaskInfo->env, code);
+      }
+    }
+    
+    if (pInfo->dataReader && hasNext) {
       if (isTaskKilled(pTaskInfo)) {
         tsdbReleaseDataBlock(pInfo->dataReader);
         longjmp(pTaskInfo->env, pTaskInfo->code);
@@ -2600,7 +2651,7 @@ static SSDataBlock* getTableDataBlockImpl(void* param) {
   SReadHandle* pHandle = &pInfo->base.readHandle;
 
   if (NULL == source->dataReader || !source->multiReader) {
-    code = tsdbReaderOpen(pHandle->vnode, pQueryCond, p, 1, pBlock, &source->dataReader, GET_TASKID(pTaskInfo));
+    code = tsdbReaderOpen(pHandle->vnode, pQueryCond, p, 1, pBlock, &source->dataReader, GET_TASKID(pTaskInfo), false);
     if (code != 0) {
       T_LONG_JMP(pTaskInfo->env, code);
     }
@@ -2608,8 +2659,21 @@ static SSDataBlock* getTableDataBlockImpl(void* param) {
   
   pInfo->base.dataReader = source->dataReader;
   STsdbReader* reader = pInfo->base.dataReader;
+  bool hasNext = false;
   qTrace("tsdb/read-table-data: %p, enter next reader", reader);
-  while (tsdbNextDataBlock(reader)) {
+
+  while (true) {
+    code = tsdbNextDataBlock(reader, &hasNext);
+    if (code != 0) {
+      tsdbReleaseDataBlock(reader);
+      pInfo->base.dataReader = NULL;
+      T_LONG_JMP(pTaskInfo->env, code);
+    }
+
+    if (!hasNext) {
+      break;
+    }
+  
     if (isTaskKilled(pTaskInfo)) {
       tsdbReleaseDataBlock(reader);
       pInfo->base.dataReader = NULL;
@@ -2814,7 +2878,7 @@ SSDataBlock* getSortedTableMergeScanBlockData(SSortHandle* pHandle, SSDataBlock*
   }
 
   bool limitReached = applyLimitOffset(&pInfo->limitInfo, pResBlock, pTaskInfo);
-  qDebug("%s get sorted row block, rows:%d, limit:%" PRId64, GET_TASKID(pTaskInfo), pResBlock->info.rows,
+  qDebug("%s get sorted row block, rows:%" PRId64 ", limit:%" PRId64, GET_TASKID(pTaskInfo), pResBlock->info.rows,
          pInfo->limitInfo.numOfOutputRows);
 
   return (pResBlock->info.rows > 0) ? pResBlock : NULL;

source/libs/executor/src/sortoperator.c

@@ -698,7 +698,7 @@ SSDataBlock* getMultiwaySortedBlockData(SSortHandle* pHandle, SSDataBlock* pData
     pDataBlock->info.dataLoad = 1;
   }
 
-  qDebug("%s get sorted block, groupId:0x%" PRIx64 " rows:%d", GET_TASKID(pTaskInfo), pDataBlock->info.id.groupId,
+  qDebug("%s get sorted block, groupId:0x%" PRIx64 " rows:%" PRId64 , GET_TASKID(pTaskInfo), pDataBlock->info.id.groupId,
          pDataBlock->info.rows);
 
   return (pDataBlock->info.rows > 0) ? pDataBlock : NULL;

source/libs/executor/src/sysscanoperator.c

@@ -1627,7 +1627,7 @@ static void sysTableScanFillTbName(SOperatorInfo* pOperator, const SSysTableScan
     char             varTbName[TSDB_TABLE_FNAME_LEN - 1 + VARSTR_HEADER_SIZE] = {0};
     STR_TO_VARSTR(varTbName, name);
 
-    colDataSetNItems(pColumnInfoData, 0, varTbName, pBlock->info.rows);
+    colDataSetNItems(pColumnInfoData, 0, varTbName, pBlock->info.rows, true);
   }
 
   doFilter(pBlock, pOperator->exprSupp.pFilterInfo, NULL);
@@ -2267,7 +2267,7 @@ SOperatorInfo* createDataBlockInfoScanOperator(SReadHandle* readHandle, SBlockDi
     size_t          num = tableListGetSize(pTableListInfo);
     void*           pList = tableListGetInfo(pTableListInfo, 0);
 
-    code = tsdbReaderOpen(readHandle->vnode, &cond, pList, num, pInfo->pResBlock, &pInfo->pHandle, pTaskInfo->id.str);
+    code = tsdbReaderOpen(readHandle->vnode, &cond, pList, num, pInfo->pResBlock, &pInfo->pHandle, pTaskInfo->id.str, false);
     cleanupQueryTableDataCond(&cond);
     if (code != 0) {
       goto _error;

source/libs/executor/src/timewindowoperator.c

@@ -128,8 +128,9 @@ FORCE_INLINE int32_t getForwardStepsInBlock(int32_t numOfRows, __block_search_fn
     if (end >= 0) {
       forwardRows = end;
 
-      if (pData[end + pos] == ekey) {
+      while (pData[end + pos] == ekey) {
         forwardRows += 1;
+        ++pos;
       }
     }
   } else {
@@ -137,8 +138,9 @@ FORCE_INLINE int32_t getForwardStepsInBlock(int32_t numOfRows, __block_search_fn
     if (end >= 0) {
       forwardRows = end;
 
-      if (pData[end + pos] == ekey) {
+      while (pData[end + pos] == ekey) {
         forwardRows += 1;
+        ++pos;
       }
     }
     //    int32_t end = searchFn((char*)pData, pos + 1, ekey, order);
@@ -2110,10 +2112,12 @@ void compactFunctions(SqlFunctionCtx* pDestCtx, SqlFunctionCtx* pSourceCtx, int3
     } else if (functionNeedToExecute(&pDestCtx[k]) && pDestCtx[k].fpSet.combine != NULL) {
       int32_t code = pDestCtx[k].fpSet.combine(&pDestCtx[k], &pSourceCtx[k]);
       if (code != TSDB_CODE_SUCCESS) {
-        qError("%s apply functions error, code: %s", GET_TASKID(pTaskInfo), tstrerror(code));
+        qError("%s apply combine functions error, code: %s", GET_TASKID(pTaskInfo), tstrerror(code));
-        pTaskInfo->code = code;
-        T_LONG_JMP(pTaskInfo->env, code);
       }
+    } else if (pDestCtx[k].fpSet.combine == NULL) {
+      char* funName = fmGetFuncName(pDestCtx[k].functionId);
+      qError("%s error, combine funcion for %s is not implemented", GET_TASKID(pTaskInfo), funName);
+      taosMemoryFreeClear(funName);
     }
   }
 }
@@ -2769,6 +2773,7 @@ SOperatorInfo* createStreamFinalIntervalOperatorInfo(SOperatorInfo* downstream,
   pInfo->pPullDataMap = taosHashInit(64, hashFn, false, HASH_NO_LOCK);
   pInfo->pPullDataRes = createSpecialDataBlock(STREAM_RETRIEVE);
   pInfo->ignoreExpiredData = pIntervalPhyNode->window.igExpired;
+  pInfo->ignoreExpiredDataSaved = false;
   pInfo->pDelRes = createSpecialDataBlock(STREAM_DELETE_RESULT);
   pInfo->delIndex = 0;
   pInfo->pDelWins = taosArrayInit(4, sizeof(SWinKey));
@@ -3587,6 +3592,7 @@ SOperatorInfo* createStreamSessionAggOperatorInfo(SOperatorInfo* downstream, SPh
   pInfo->isFinal = false;
   pInfo->pPhyNode = pPhyNode;
   pInfo->ignoreExpiredData = pSessionNode->window.igExpired;
+  pInfo->ignoreExpiredDataSaved = false;
   pInfo->pUpdated = NULL;
   pInfo->pStUpdated = NULL;
 
@@ -4112,6 +4118,7 @@ SOperatorInfo* createStreamStateAggOperatorInfo(SOperatorInfo* downstream, SPhys
   pInfo->pDelRes = createSpecialDataBlock(STREAM_DELETE_RESULT);
   pInfo->pChildren = NULL;
   pInfo->ignoreExpiredData = pStateNode->window.igExpired;
+  pInfo->ignoreExpiredDataSaved = false;
   pInfo->pUpdated = NULL;
   pInfo->pSeUpdated = NULL;
 
@@ -4885,6 +4892,7 @@ SOperatorInfo* createStreamIntervalOperatorInfo(SOperatorInfo* downstream, SPhys
   pInfo->interval = interval;
   pInfo->twAggSup = twAggSupp;
   pInfo->ignoreExpiredData = pIntervalPhyNode->window.igExpired;
+  pInfo->ignoreExpiredDataSaved = false;
   pInfo->isFinal = false;
 
   SExprSupp* pSup = &pOperator->exprSupp;

source/libs/function/inc/builtinsimpl.h

@@ -235,6 +235,7 @@ int32_t blockDistFinalize(SqlFunctionCtx* pCtx, SSDataBlock* pBlock);
 bool    getGroupKeyFuncEnv(SFunctionNode* pFunc, SFuncExecEnv* pEnv);
 int32_t groupKeyFunction(SqlFunctionCtx* pCtx);
 int32_t groupKeyFinalize(SqlFunctionCtx* pCtx, SSDataBlock* pBlock);
+int32_t groupKeyCombine(SqlFunctionCtx* pDestCtx, SqlFunctionCtx* pSourceCtx);
 
 #ifdef __cplusplus
 }

source/libs/function/src/builtins.c

@@ -2375,7 +2375,7 @@ const SBuiltinFuncDefinition funcMgtBuiltins[] = {
   {
     .name = "spread",
     .type = FUNCTION_TYPE_SPREAD,
-    .classification = FUNC_MGT_AGG_FUNC,
+    .classification = FUNC_MGT_AGG_FUNC | FUNC_MGT_SPECIAL_DATA_REQUIRED,
     .translateFunc = translateSpread,
     .dataRequiredFunc = statisDataRequired,
     .getEnvFunc   = getSpreadFuncEnv,
@@ -2417,7 +2417,7 @@ const SBuiltinFuncDefinition funcMgtBuiltins[] = {
   {
     .name = "elapsed",
     .type = FUNCTION_TYPE_ELAPSED,
-    .classification = FUNC_MGT_AGG_FUNC | FUNC_MGT_TIMELINE_FUNC | FUNC_MGT_INTERVAL_INTERPO_FUNC | FUNC_MGT_FORBID_STREAM_FUNC,
+    .classification = FUNC_MGT_AGG_FUNC | FUNC_MGT_TIMELINE_FUNC | FUNC_MGT_INTERVAL_INTERPO_FUNC | FUNC_MGT_FORBID_STREAM_FUNC | FUNC_MGT_SPECIAL_DATA_REQUIRED,
     .dataRequiredFunc = statisDataRequired,
     .translateFunc = translateElapsed,
     .getEnvFunc   = getElapsedFuncEnv,
@@ -2457,7 +2457,7 @@ const SBuiltinFuncDefinition funcMgtBuiltins[] = {
     .name = "interp",
     .type = FUNCTION_TYPE_INTERP,
     .classification = FUNC_MGT_TIMELINE_FUNC | FUNC_MGT_INTERVAL_INTERPO_FUNC | FUNC_MGT_IMPLICIT_TS_FUNC |
-                      FUNC_MGT_FORBID_STREAM_FUNC,
+                      FUNC_MGT_FORBID_STREAM_FUNC|FUNC_MGT_KEEP_ORDER_FUNC,
     .translateFunc = translateInterp,
     .getEnvFunc    = getSelectivityFuncEnv,
     .initFunc      = functionSetup,
@@ -2480,7 +2480,7 @@ const SBuiltinFuncDefinition funcMgtBuiltins[] = {
   {
     .name = "irate",
     .type = FUNCTION_TYPE_IRATE,
-    .classification = FUNC_MGT_AGG_FUNC | FUNC_MGT_TIMELINE_FUNC | FUNC_MGT_IMPLICIT_TS_FUNC,
+    .classification = FUNC_MGT_AGG_FUNC | FUNC_MGT_TIMELINE_FUNC | FUNC_MGT_IMPLICIT_TS_FUNC | FUNC_MGT_FORBID_STREAM_FUNC,
     .translateFunc = translateIrate,
     .getEnvFunc   = getIrateFuncEnv,
     .initFunc     = irateFuncSetup,
@@ -2500,7 +2500,8 @@ const SBuiltinFuncDefinition funcMgtBuiltins[] = {
     .sprocessFunc = firstLastScalarFunction,
     .pPartialFunc = "_last_row_partial",
     .pMergeFunc   = "_last_row_merge",
-    .finalizeFunc = firstLastFinalize
+    .finalizeFunc = firstLastFinalize,
+    .combineFunc  = lastCombine
   },
   {
     .name = "_cache_last_row",
@@ -2809,7 +2810,7 @@ const SBuiltinFuncDefinition funcMgtBuiltins[] = {
   {
     .name = "mode",
     .type = FUNCTION_TYPE_MODE,
-    .classification = FUNC_MGT_AGG_FUNC | FUNC_MGT_SELECT_FUNC,
+    .classification = FUNC_MGT_AGG_FUNC | FUNC_MGT_SELECT_FUNC | FUNC_MGT_FORBID_STREAM_FUNC,
     .translateFunc = translateMode,
     .getEnvFunc   = getModeFuncEnv,
     .initFunc     = modeFunctionSetup,
@@ -3212,7 +3213,7 @@ const SBuiltinFuncDefinition funcMgtBuiltins[] = {
   {
     .name = "_block_dist",
     .type = FUNCTION_TYPE_BLOCK_DIST,
-    .classification = FUNC_MGT_AGG_FUNC,
+    .classification = FUNC_MGT_AGG_FUNC | FUNC_MGT_FORBID_STREAM_FUNC,
     .translateFunc = translateBlockDistFunc,
     .getEnvFunc   = getBlockDistFuncEnv,
     .initFunc     = blockDistSetup,
@@ -3234,6 +3235,7 @@ const SBuiltinFuncDefinition funcMgtBuiltins[] = {
     .initFunc     = functionSetup,
     .processFunc  = groupKeyFunction,
     .finalizeFunc = groupKeyFinalize,
+    .combineFunc  = groupKeyCombine,
     .pPartialFunc = "_group_key",
     .pMergeFunc   = "_group_key"
   },
@@ -3276,7 +3278,7 @@ const SBuiltinFuncDefinition funcMgtBuiltins[] = {
   {
     .name = "_irowts",
     .type = FUNCTION_TYPE_IROWTS,
-    .classification = FUNC_MGT_PSEUDO_COLUMN_FUNC | FUNC_MGT_INTERP_PC_FUNC,
+    .classification = FUNC_MGT_PSEUDO_COLUMN_FUNC | FUNC_MGT_INTERP_PC_FUNC|FUNC_MGT_KEEP_ORDER_FUNC,
     .translateFunc = translateTimePseudoColumn,
     .getEnvFunc   = getTimePseudoFuncEnv,
     .initFunc     = NULL,

source/libs/function/src/builtinsimpl.c

@@ -494,8 +494,8 @@ bool getCountFuncEnv(SFunctionNode* UNUSED_PARAM(pFunc), SFuncExecEnv* pEnv) {
   return true;
 }
 
-static int32_t getNumOfElems(SqlFunctionCtx* pCtx) {
+static int64_t getNumOfElems(SqlFunctionCtx* pCtx) {
-  int32_t numOfElem = 0;
+  int64_t numOfElem = 0;
 
   /*
    * 1. column data missing (schema modified) causes pInputCol->hasNull == true. pInput->colDataSMAIsSet == true;
@@ -528,7 +528,7 @@ static int32_t getNumOfElems(SqlFunctionCtx* pCtx) {
  * count function does not use the pCtx->interResBuf to keep the intermediate buffer
  */
 int32_t countFunction(SqlFunctionCtx* pCtx) {
-  int32_t numOfElem = 0;
+  int64_t numOfElem = 0;
 
   SResultRowEntryInfo*  pResInfo = GET_RES_INFO(pCtx);
   SInputColumnInfoData* pInput = &pCtx->input;
@@ -555,7 +555,7 @@ int32_t countFunction(SqlFunctionCtx* pCtx) {
 }
 
 int32_t countInvertFunction(SqlFunctionCtx* pCtx) {
-  int32_t numOfElem = getNumOfElems(pCtx);
+  int64_t numOfElem = getNumOfElems(pCtx);
 
   SResultRowEntryInfo* pResInfo = GET_RES_INFO(pCtx);
   char*                buf = GET_ROWCELL_INTERBUF(pResInfo);
@@ -871,6 +871,12 @@ int32_t setSelectivityValue(SqlFunctionCtx* pCtx, SSDataBlock* pBlock, const STu
       SqlFunctionCtx* pc = pCtx->subsidiaries.pCtx[j];
       int32_t         dstSlotId = pc->pExpr->base.resSchema.slotId;
 
+      // group_key function has its own process function
+      // do not process there
+      if (fmIsGroupKeyFunc(pc->functionId)) {
+        continue;
+      }
+
       SColumnInfoData* pDstCol = taosArrayGet(pBlock->pDataBlock, dstSlotId);
       if (nullList[j]) {
         colDataSetNULL(pDstCol, rowIndex);
@@ -1929,7 +1935,7 @@ int32_t apercentileFunctionMerge(SqlFunctionCtx* pCtx) {
 
   SAPercentileInfo* pInfo = GET_ROWCELL_INTERBUF(pResInfo);
 
-  qDebug("%s total %d rows will merge, %p", __FUNCTION__, pInput->numOfRows, pInfo->pHisto);
+  qDebug("%s total %" PRId64 " rows will merge, %p", __FUNCTION__, pInput->numOfRows, pInfo->pHisto);
 
   int32_t start = pInput->startRowIndex;
   for (int32_t i = start; i < start + pInput->numOfRows; ++i) {
@@ -3091,6 +3097,12 @@ void* serializeTupleData(const SSDataBlock* pSrcBlock, int32_t rowIndex, SSubsid
   for (int32_t i = 0; i < pSubsidiaryies->num; ++i) {
     SqlFunctionCtx* pc = pSubsidiaryies->pCtx[i];
 
+    // group_key function has its own process function
+    // do not process there
+    if (fmIsGroupKeyFunc(pc->functionId)) {
+      continue;
+    }
+
     SFunctParam* pFuncParam = &pc->pExpr->base.pParam[0];
     int32_t      srcSlotId = pFuncParam->pCol->slotId;
 
@@ -5900,6 +5912,39 @@ int32_t groupKeyFinalize(SqlFunctionCtx* pCtx, SSDataBlock* pBlock) {
   return pResInfo->numOfRes;
 }
 
+int32_t groupKeyCombine(SqlFunctionCtx* pDestCtx, SqlFunctionCtx* pSourceCtx) {
+  SResultRowEntryInfo* pDResInfo = GET_RES_INFO(pDestCtx);
+  SGroupKeyInfo*       pDBuf = GET_ROWCELL_INTERBUF(pDResInfo);
+
+  SResultRowEntryInfo* pSResInfo = GET_RES_INFO(pSourceCtx);
+  SGroupKeyInfo*       pSBuf = GET_ROWCELL_INTERBUF(pSResInfo);
+
+  // escape rest of data blocks to avoid first entry to be overwritten.
+  if (pDBuf->hasResult) {
+    goto _group_key_over;
+  }
+
+  if (pSBuf->isNull) {
+    pDBuf->isNull = true;
+    pDBuf->hasResult = true;
+    goto _group_key_over;
+  }
+
+  if (IS_VAR_DATA_TYPE(pSourceCtx->resDataInfo.type)) {
+    memcpy(pDBuf->data, pSBuf->data,
+           (pSourceCtx->resDataInfo.type == TSDB_DATA_TYPE_JSON) ? getJsonValueLen(pSBuf->data) : varDataTLen(pSBuf->data));
+  } else {
+    memcpy(pDBuf->data, pSBuf->data, pSourceCtx->resDataInfo.bytes);
+  }
+
+  pDBuf->hasResult = true;
+
+_group_key_over:
+
+  SET_VAL(pDResInfo, 1, 1);
+  return TSDB_CODE_SUCCESS;
+}
+
 int32_t cachedLastRowFunction(SqlFunctionCtx* pCtx) {
   int32_t numOfElems = 0;
 

source/libs/function/src/functionMgt.c

@@ -447,3 +447,10 @@ int32_t fmGetDistMethod(const SFunctionNode* pFunc, SFunctionNode** pPartialFunc
 
   return code;
 }
+
+char* fmGetFuncName(int32_t funcId) {
+  if (fmIsUserDefinedFunc(funcId) || funcId < 0 || funcId >= funcMgtBuiltinsNum) {
+    return taosStrdup("invalid function");
+  }
+  return  taosStrdup(funcMgtBuiltins[funcId].name);
+}

source/libs/parser/src/parCalcConst.c

@@ -183,16 +183,18 @@ static int32_t calcConstProject(SNode* pProject, bool dual, SNode** pNew) {
   } else {
     code = scalarCalculateConstants(pProject, pNew);
   }
-  if (TSDB_CODE_SUCCESS == code && QUERY_NODE_VALUE == nodeType(*pNew) && NULL != pAssociation) {
+  if (TSDB_CODE_SUCCESS == code) {
     strcpy(((SExprNode*)*pNew)->aliasName, aliasName);
-    int32_t size = taosArrayGetSize(pAssociation);
+    if (QUERY_NODE_VALUE == nodeType(*pNew) && NULL != pAssociation) {
-    for (int32_t i = 0; i < size; ++i) {
+      int32_t size = taosArrayGetSize(pAssociation);
-      SNode** pCol = taosArrayGetP(pAssociation, i);
+      for (int32_t i = 0; i < size; ++i) {
-      nodesDestroyNode(*pCol);
+        SNode** pCol = taosArrayGetP(pAssociation, i);
-      *pCol = nodesCloneNode(*pNew);
+        nodesDestroyNode(*pCol);
-      if (NULL == *pCol) {
+        *pCol = nodesCloneNode(*pNew);
-        code = TSDB_CODE_OUT_OF_MEMORY;
+        if (NULL == *pCol) {
-        break;
+          code = TSDB_CODE_OUT_OF_MEMORY;
+          break;
+        }
       }
     }
   }
@@ -383,7 +385,8 @@ static int32_t calcConstSetOpProjections(SCalcConstContext* pCxt, SSetOperator*
   int32_t index = 0;
   SNode*  pProj = NULL;
   WHERE_EACH(pProj, pSetOp->pProjectionList) {
-    if (subquery && notRefByOrderBy((SColumnNode*)pProj, pSetOp->pOrderByList) && isSetUselessCol(pSetOp, index, (SExprNode*)pProj)) {
+    if (subquery && notRefByOrderBy((SColumnNode*)pProj, pSetOp->pOrderByList) &&
+        isSetUselessCol(pSetOp, index, (SExprNode*)pProj)) {
       ERASE_NODE(pSetOp->pProjectionList);
       eraseSetOpChildProjection(pSetOp, index);
       continue;

source/libs/parser/src/parInsertSml.c

@@ -25,6 +25,8 @@ static void clearColValArray(SArray* pCols) {
     if (TSDB_DATA_TYPE_NCHAR == pCol->type) {
       taosMemoryFreeClear(pCol->value.pData);
     }
+    pCol->flag = CV_FLAG_NONE;
+    pCol->value.val = 0;
   }
 }
 

source/libs/parser/src/parInsertUtil.c

@@ -566,53 +566,19 @@ int32_t insBuildVgDataBlocks(SHashObj* pVgroupsHashObj, SArray* pVgDataCxtList,
   return code;
 }
 
-static int bindFileds(SBoundColInfo* pBoundInfo, SSchema* pSchema, TAOS_FIELD* fields, int numFields) {
+static bool findFileds(SSchema* pSchema, TAOS_FIELD* fields, int numFields) {
-  bool* pUseCols = taosMemoryCalloc(pBoundInfo->numOfCols, sizeof(bool));
-  if (NULL == pUseCols) {
-    return TSDB_CODE_OUT_OF_MEMORY;
-  }
-
-  pBoundInfo->numOfBound = 0;
-
-  int16_t lastColIdx = -1;  // last column found
-  int32_t code = TSDB_CODE_SUCCESS;
   for (int i = 0; i < numFields; i++) {
-    SToken token;
+    if(strcmp(pSchema->name, fields[i].name) == 0){
-    token.z = fields[i].name;
+      return true;
-    token.n = strlen(fields[i].name);
-
-    int16_t t = lastColIdx + 1;
-    int16_t index = insFindCol(&token, t, pBoundInfo->numOfCols, pSchema);
-    if (index < 0 && t > 0) {
-      index = insFindCol(&token, 0, t, pSchema);
-    }
-    if (index < 0) {
-      uError("can not find column name:%s", token.z);
-      code = TSDB_CODE_PAR_INVALID_COLUMN;
-      break;
-    } else if (pUseCols[index]) {
-      code = TSDB_CODE_PAR_INVALID_COLUMN;
-      uError("duplicated column name:%s", token.z);
-      break;
-    } else {
-      lastColIdx = index;
-      pUseCols[index] = true;
-      pBoundInfo->pColIndex[pBoundInfo->numOfBound] = index;
-      ++pBoundInfo->numOfBound;
     }
   }
 
-  if (TSDB_CODE_SUCCESS == code && !pUseCols[0]) {
+  return false;
-    uError("primary timestamp column can not be null:");
-    code = TSDB_CODE_PAR_INVALID_COLUMN;
-  }
-
-  taosMemoryFree(pUseCols);
-  return code;
 }
 
 int rawBlockBindData(SQuery* query, STableMeta* pTableMeta, void* data, SVCreateTbReq* pCreateTb, TAOS_FIELD* tFields,
                      int numFields, bool needChangeLength) {
+  void* tmp = taosHashGet(((SVnodeModifyOpStmt*)(query->pRoot))->pTableBlockHashObj, &pTableMeta->uid, sizeof(pTableMeta->uid));
   STableDataCxt* pTableCxt = NULL;
   int            ret = insGetTableDataCxt(((SVnodeModifyOpStmt*)(query->pRoot))->pTableBlockHashObj, &pTableMeta->uid,
                                           sizeof(pTableMeta->uid), pTableMeta, &pCreateTb, &pTableCxt, true);
@@ -620,19 +586,14 @@ int rawBlockBindData(SQuery* query, STableMeta* pTableMeta, void* data, SVCreate
     uError("insGetTableDataCxt error");
     goto end;
   }
-  if (tFields != NULL) {
+
-    ret = bindFileds(&pTableCxt->boundColsInfo, getTableColumnSchema(pTableMeta), tFields, numFields);
+  if(tmp == NULL){
+    ret = initTableColSubmitData(pTableCxt);
     if (ret != TSDB_CODE_SUCCESS) {
-      uError("bindFileds error");
+      uError("initTableColSubmitData error");
       goto end;
     }
   }
-  // no need to bind, because select * get all fields
-  ret = initTableColSubmitData(pTableCxt);
-  if (ret != TSDB_CODE_SUCCESS) {
-    uError("initTableColSubmitData error");
-    goto end;
-  }
 
   char* p = (char*)data;
   // | version | total length | total rows | total columns | flag seg| block group id | column schema | each column
@@ -660,35 +621,83 @@ int rawBlockBindData(SQuery* query, STableMeta* pTableMeta, void* data, SVCreate
   SSchema*       pSchema = getTableColumnSchema(pTableCxt->pMeta);
   SBoundColInfo* boundInfo = &pTableCxt->boundColsInfo;
 
-  if (boundInfo->numOfBound != numOfCols) {
+  if (tFields != NULL && numFields != numOfCols) {
-    uError("boundInfo->numOfBound:%d != numOfCols:%d", boundInfo->numOfBound, numOfCols);
+    uError("numFields:%d != numOfCols:%d", numFields, numOfCols);
     ret = TSDB_CODE_INVALID_PARA;
     goto end;
   }
-  for (int c = 0; c < boundInfo->numOfBound; ++c) {
+  if (tFields != NULL && numFields > boundInfo->numOfBound) {
-    SSchema*  pColSchema = &pSchema[boundInfo->pColIndex[c]];
+    uError("numFields:%d > boundInfo->numOfBound:%d", numFields, boundInfo->numOfBound);
-    SColData* pCol = taosArrayGet(pTableCxt->pData->aCol, c);
+    ret = TSDB_CODE_INVALID_PARA;
+    goto end;
+  }
+  if(tFields == NULL){
+    for (int j = 0; j < boundInfo->numOfBound; j++){
+      SSchema*  pColSchema = &pSchema[j];
+      SColData* pCol = taosArrayGet(pTableCxt->pData->aCol, j);
+      if (*fields != pColSchema->type && *(int32_t*)(fields + sizeof(int8_t)) != pColSchema->bytes) {
+        uError("type or bytes not equal");
+        ret = TSDB_CODE_INVALID_PARA;
+        goto end;
+      }
+
+      int8_t* offset = pStart;
+      if (IS_VAR_DATA_TYPE(pColSchema->type)) {
+        pStart += numOfRows * sizeof(int32_t);
+      } else {
+        pStart += BitmapLen(numOfRows);
+      }
+      char* pData = pStart;
+
+      tColDataAddValueByDataBlock(pCol, pColSchema->type, pColSchema->bytes, numOfRows, offset, pData);
+      fields += sizeof(int8_t) + sizeof(int32_t);
+      if (needChangeLength) {
+        pStart += htonl(colLength[j]);
+      } else {
+        pStart += colLength[j];
+      }
+    }
+  }else{
+    for (int i = 0; i < numFields; i++) {
+      for (int j = 0; j < boundInfo->numOfBound; j++){
+        SSchema*  pColSchema = &pSchema[j];
+        if(strcmp(pColSchema->name, tFields[i].name) == 0){
+          if (*fields != pColSchema->type && *(int32_t*)(fields + sizeof(int8_t)) != pColSchema->bytes) {
+            uError("type or bytes not equal");
+            ret = TSDB_CODE_INVALID_PARA;
+            goto end;
+          }
+
+          int8_t* offset = pStart;
+          if (IS_VAR_DATA_TYPE(pColSchema->type)) {
+            pStart += numOfRows * sizeof(int32_t);
+          } else {
+            pStart += BitmapLen(numOfRows);
+          }
+          char* pData = pStart;
+
+          SColData* pCol = taosArrayGet(pTableCxt->pData->aCol, j);
+          tColDataAddValueByDataBlock(pCol, pColSchema->type, pColSchema->bytes, numOfRows, offset, pData);
+          fields += sizeof(int8_t) + sizeof(int32_t);
+          if (needChangeLength) {
+            pStart += htonl(colLength[i]);
+          } else {
+            pStart += colLength[i];
+          }
+          boundInfo->pColIndex[j] = -1;
+          break;
+        }
+      }
 
-    if (*fields != pColSchema->type && *(int32_t*)(fields + sizeof(int8_t)) != pColSchema->bytes) {
-      uError("type or bytes not equal");
-      ret = TSDB_CODE_INVALID_PARA;
-      goto end;
     }
 
-    int8_t* offset = pStart;
+    for (int c = 0; c < boundInfo->numOfBound; ++c) {
-    if (IS_VAR_DATA_TYPE(pColSchema->type)) {
+      if( boundInfo->pColIndex[c] != -1){
-      pStart += numOfRows * sizeof(int32_t);
+        SColData* pCol = taosArrayGet(pTableCxt->pData->aCol, c);
-    } else {
+        tColDataAddValueByDataBlock(pCol, 0, 0, numOfRows, NULL, NULL);
-      pStart += BitmapLen(numOfRows);
+      }else{
-    }
+        boundInfo->pColIndex[c] = c;  // restore for next block
-    char* pData = pStart;
+      }
-
-    tColDataAddValueByDataBlock(pCol, pColSchema->type, pColSchema->bytes, numOfRows, offset, pData);
-    fields += sizeof(int8_t) + sizeof(int32_t);
-    if (needChangeLength) {
-      pStart += htonl(colLength[c]);
-    } else {
-      pStart += colLength[c];
     }
   }
 

source/libs/parser/src/parTranslater.c

@@ -757,7 +757,7 @@ static bool isPrimaryKeyImpl(SNode* pExpr) {
     if (FUNCTION_TYPE_SELECT_VALUE == pFunc->funcType || FUNCTION_TYPE_GROUP_KEY == pFunc->funcType ||
         FUNCTION_TYPE_FIRST == pFunc->funcType || FUNCTION_TYPE_LAST == pFunc->funcType) {
       return isPrimaryKeyImpl(nodesListGetNode(pFunc->pParameterList, 0));
-    } else if (FUNCTION_TYPE_WSTART == pFunc->funcType || FUNCTION_TYPE_WEND == pFunc->funcType) {
+    } else if (FUNCTION_TYPE_WSTART == pFunc->funcType || FUNCTION_TYPE_WEND == pFunc->funcType || FUNCTION_TYPE_IROWTS == pFunc->funcType) {
       return true;
     }
   }
@@ -1342,8 +1342,8 @@ static bool isCountNotNullValue(SFunctionNode* pFunc) {
 // count(1) is rewritten as count(ts) for scannning optimization
 static int32_t rewriteCountNotNullValue(STranslateContext* pCxt, SFunctionNode* pCount) {
   SValueNode* pValue = (SValueNode*)nodesListGetNode(pCount->pParameterList, 0);
-  STableNode*  pTable = NULL;
+  STableNode* pTable = NULL;
-  int32_t      code = findTable(pCxt, NULL, &pTable);
+  int32_t     code = findTable(pCxt, NULL, &pTable);
   if (TSDB_CODE_SUCCESS == code && QUERY_NODE_REAL_TABLE == nodeType(pTable)) {
     SColumnNode* pCol = (SColumnNode*)nodesMakeNode(QUERY_NODE_COLUMN);
     if (NULL == pCol) {
@@ -1424,7 +1424,7 @@ static int32_t translateAggFunc(STranslateContext* pCxt, SFunctionNode* pFunc) {
   }
   if (isCountNotNullValue(pFunc)) {
     return rewriteCountNotNullValue(pCxt, pFunc);
-  }  
+  }
   if (isCountTbname(pFunc)) {
     return rewriteCountTbname(pCxt, pFunc);
   }
@@ -1634,13 +1634,15 @@ static bool isTableStar(SNode* pNode) {
          (0 == strcmp(((SColumnNode*)pNode)->colName, "*"));
 }
 
+static bool isStarParam(SNode* pNode) { return isStar(pNode) || isTableStar(pNode); }
+
 static int32_t translateMultiResFunc(STranslateContext* pCxt, SFunctionNode* pFunc) {
   if (!fmIsMultiResFunc(pFunc->funcId)) {
     return TSDB_CODE_SUCCESS;
   }
   if (SQL_CLAUSE_SELECT != pCxt->currClause) {
     SNode* pPara = nodesListGetNode(pFunc->pParameterList, 0);
-    if (isStar(pPara) || isTableStar(pPara)) {
+    if (isStarParam(pPara)) {
       return generateSyntaxErrMsgExt(&pCxt->msgBuf, TSDB_CODE_PAR_NOT_ALLOWED_FUNC,
                                      "%s(*) is only supported in SELECTed list", pFunc->functionName);
     }
@@ -1654,7 +1656,7 @@ static int32_t translateMultiResFunc(STranslateContext* pCxt, SFunctionNode* pFu
 
 static int32_t getMultiResFuncNum(SNodeList* pParameterList) {
   if (1 == LIST_LENGTH(pParameterList)) {
-    return isStar(nodesListGetNode(pParameterList, 0)) ? 2 : 1;
+    return isStarParam(nodesListGetNode(pParameterList, 0)) ? 2 : 1;
   }
   return LIST_LENGTH(pParameterList);
 }
@@ -5923,11 +5925,15 @@ static int32_t addSubtableInfoToCreateStreamQuery(STranslateContext* pCxt, STabl
   return code;
 }
 
+static bool isEventWindowQuery(SSelectStmt* pSelect) {
+  return NULL != pSelect->pWindow && QUERY_NODE_EVENT_WINDOW == nodeType(pSelect->pWindow);
+}
+
 static int32_t checkStreamQuery(STranslateContext* pCxt, SCreateStreamStmt* pStmt) {
   SSelectStmt* pSelect = (SSelectStmt*)pStmt->pQuery;
   if (TSDB_DATA_TYPE_TIMESTAMP != ((SExprNode*)nodesListGetNode(pSelect->pProjectionList, 0))->resType.type ||
       !pSelect->isTimeLineResult || crossTableWithoutAggOper(pSelect) || NULL != pSelect->pOrderByList ||
-      crossTableWithUdaf(pSelect)) {
+      crossTableWithUdaf(pSelect) || isEventWindowQuery(pSelect)) {
     return generateSyntaxErrMsgExt(&pCxt->msgBuf, TSDB_CODE_PAR_INVALID_STREAM_QUERY, "Unsupported stream query");
   }
   if (NULL != pSelect->pSubtable && TSDB_DATA_TYPE_VARCHAR != ((SExprNode*)pSelect->pSubtable)->resType.type) {

source/libs/planner/src/planPhysiCreater.c

@@ -1089,9 +1089,15 @@ static int32_t doCreateExchangePhysiNode(SPhysiPlanContext* pCxt, SExchangeLogic
   pExchange->srcStartGroupId = pExchangeLogicNode->srcStartGroupId;
   pExchange->srcEndGroupId = pExchangeLogicNode->srcEndGroupId;
   pExchange->seqRecvData = pExchangeLogicNode->seqRecvData;
-  *pPhyNode = (SPhysiNode*)pExchange;
 
-  return TSDB_CODE_SUCCESS;
+  int32_t code = setConditionsSlotId(pCxt, (const SLogicNode*)pExchangeLogicNode, (SPhysiNode*)pExchange);
+  if (TSDB_CODE_SUCCESS == code) {
+    *pPhyNode = (SPhysiNode*)pExchange;
+  } else {
+    nodesDestroyNode((SNode*)pExchange);
+  }
+
+  return code;
 }
 
 static int32_t createStreamScanPhysiNodeByExchange(SPhysiPlanContext* pCxt, SExchangeLogicNode* pExchangeLogicNode,
@@ -1119,6 +1125,9 @@ static int32_t createStreamScanPhysiNodeByExchange(SPhysiPlanContext* pCxt, SExc
   if (TSDB_CODE_SUCCESS == code) {
     code = addDataBlockSlots(pCxt, pScan->pScanCols, pScan->node.pOutputDataBlockDesc);
   }
+  if (TSDB_CODE_SUCCESS == code) {
+    code = setConditionsSlotId(pCxt, (const SLogicNode*)pExchangeLogicNode, (SPhysiNode*)pScan);
+  }
 
   if (TSDB_CODE_SUCCESS == code) {
     *pPhyNode = (SPhysiNode*)pScan;

source/libs/planner/src/planSpliter.c

@@ -1365,7 +1365,8 @@ static int32_t unAllSplCreateExchangeNode(SSplitContext* pCxt, int32_t startGrou
   pExchange->srcEndGroupId = pCxt->groupId - 1;
   pExchange->node.precision = pProject->node.precision;
   pExchange->node.pTargets = nodesCloneList(pProject->node.pTargets);
-  if (NULL == pExchange->node.pTargets) {
+  pExchange->node.pConditions = nodesCloneNode(pProject->node.pConditions);
+  if (NULL == pExchange->node.pTargets || (NULL != pProject->node.pConditions && NULL == pExchange->node.pConditions)) {
     return TSDB_CODE_OUT_OF_MEMORY;
   }
   TSWAP(pExchange->node.pLimit, pProject->node.pLimit);

source/libs/qworker/src/qworker.c

@@ -199,7 +199,7 @@ int32_t qwExecTask(QW_FPARAMS_DEF, SQWTaskCtx *ctx, bool *queryStop) {
         QW_ERR_JRET(code);
       }
 
-      QW_TASK_DLOG("data put into sink, rows:%d, continueExecTask:%d", pRes->info.rows, qcontinue);
+      QW_TASK_DLOG("data put into sink, rows:%" PRId64 ", continueExecTask:%d", pRes->info.rows, qcontinue);
     }
 
     if (numOfResBlock == 0 || (hasMore == false)) {

source/libs/scalar/src/sclfunc.c

@@ -1755,7 +1755,11 @@ int32_t winEndTsFunction(SScalarParam *pInput, int32_t inputNum, SScalarParam *p
 int32_t qTbnameFunction(SScalarParam *pInput, int32_t inputNum, SScalarParam *pOutput) {
   char* p = colDataGetVarData(pInput->columnData, 0);
 
-  colDataSetNItems(pOutput->columnData, pOutput->numOfRows, p, pInput->numOfRows);
+  int32_t code = colDataSetNItems(pOutput->columnData, pOutput->numOfRows, p, pInput->numOfRows, true);
+  if (code) {
+    return code;
+  }
+  
   pOutput->numOfRows += pInput->numOfRows;
   return TSDB_CODE_SUCCESS;
 }

source/libs/stream/src/stream.c

@@ -275,7 +275,7 @@ int32_t streamProcessRetrieveReq(SStreamTask* pTask, SStreamRetrieveReq* pReq, S
   return 0;
 }
 
-int32_t streamProcessRetrieveRsp(SStreamTask* pTask, SStreamRetrieveRsp* pRsp) {
+// int32_t streamProcessRetrieveRsp(SStreamTask* pTask, SStreamRetrieveRsp* pRsp) {
-  //
+//   //
-  return 0;
+//   return 0;
-}
+// }

source/libs/stream/src/streamExec.c

@@ -152,8 +152,14 @@ int32_t streamScanExec(SStreamTask* pTask, int32_t batchSz) {
       if (batchCnt >= batchSz) break;
     }
     if (taosArrayGetSize(pRes) == 0) {
-      taosArrayDestroy(pRes);
+      if (finished) {
-      break;
+        taosArrayDestroy(pRes);
+        qDebug("task %d finish recover exec task ", pTask->taskId);
+        break;
+      } else {
+        qDebug("task %d continue recover exec task ", pTask->taskId);
+        continue;
+      }
     }
     SStreamDataBlock* qRes = taosAllocateQitem(sizeof(SStreamDataBlock), DEF_QITEM, 0);
     if (qRes == NULL) {

source/libs/sync/src/syncAppendEntries.c

@@ -105,6 +105,7 @@ int32_t syncNodeOnAppendEntries(SSyncNode* ths, const SRpcMsg* pRpcMsg) {
   SRpcMsg            rpcRsp = {0};
   bool               accepted = false;
   SSyncRaftEntry*    pEntry = NULL;
+  bool               resetElect = false;
 
   // if already drop replica, do not process
   if (!syncNodeInRaftGroup(ths, &(pMsg->srcId))) {
@@ -137,7 +138,7 @@ int32_t syncNodeOnAppendEntries(SSyncNode* ths, const SRpcMsg* pRpcMsg) {
   }
 
   syncNodeStepDown(ths, pMsg->term);
-  syncNodeResetElectTimer(ths);
+  resetElect = true;
 
   if (pMsg->dataLen < sizeof(SSyncRaftEntry)) {
     sError("vgId:%d, incomplete append entries received. prev index:%" PRId64 ", term:%" PRId64 ", datalen:%d",
@@ -184,10 +185,9 @@ _SEND_RESPONSE:
   // commit index, i.e. leader notice me
   if (syncLogBufferCommit(ths->pLogBuf, ths, ths->commitIndex) < 0) {
     sError("vgId:%d, failed to commit raft fsm log since %s.", ths->vgId, terrstr());
-    goto _out;
   }
 
-_out:
+  if (resetElect) syncNodeResetElectTimer(ths);
   return 0;
 
 _IGNORE:

source/libs/sync/src/syncElection.c

@@ -115,6 +115,5 @@ int32_t syncNodeElect(SSyncNode* pSyncNode) {
   ASSERT(ret == 0);
 
   syncNodeResetElectTimer(pSyncNode);
-
   return ret;
 }

source/libs/sync/src/syncMain.c

@@ -182,6 +182,9 @@ int32_t syncProcessMsg(int64_t rid, SRpcMsg* pMsg) {
     case TDMT_SYNC_TIMEOUT:
       code = syncNodeOnTimeout(pSyncNode, pMsg);
       break;
+    case TDMT_SYNC_TIMEOUT_ELECTION:
+      code = syncNodeOnTimeout(pSyncNode, pMsg);
+      break;
     case TDMT_SYNC_CLIENT_REQUEST:
       code = syncNodeOnClientRequest(pSyncNode, pMsg, NULL);
       break;
@@ -1593,8 +1596,8 @@ void syncNodeBecomeFollower(SSyncNode* pSyncNode, const char* debugStr) {
   pSyncNode->state = TAOS_SYNC_STATE_FOLLOWER;
   syncNodeStopHeartbeatTimer(pSyncNode);
 
-  // reset elect timer
+  // trace log
-  syncNodeResetElectTimer(pSyncNode);
+  sNTrace(pSyncNode, "become follower %s", debugStr);
 
   // send rsp to client
   syncNodeLeaderChangeRsp(pSyncNode);
@@ -1610,8 +1613,8 @@ void syncNodeBecomeFollower(SSyncNode* pSyncNode, const char* debugStr) {
   // reset log buffer
   syncLogBufferReset(pSyncNode->pLogBuf, pSyncNode);
 
-  // trace log
+  // reset elect timer
-  sNTrace(pSyncNode, "become follower %s", debugStr);
+  syncNodeResetElectTimer(pSyncNode);
 }
 
 // TLA+ Spec
@@ -2277,6 +2280,7 @@ static int32_t syncNodeAppendNoopOld(SSyncNode* ths) {
 
 int32_t syncNodeOnHeartbeat(SSyncNode* ths, const SRpcMsg* pRpcMsg) {
   SyncHeartbeat* pMsg = pRpcMsg->pCont;
+  bool           resetElect = false;
 
   const STraceId* trace = &pRpcMsg->info.traceId;
   char            tbuf[40] = {0};
@@ -2300,12 +2304,11 @@ int32_t syncNodeOnHeartbeat(SSyncNode* ths, const SRpcMsg* pRpcMsg) {
 
   if (pMsg->term == currentTerm && ths->state != TAOS_SYNC_STATE_LEADER) {
     syncIndexMgrSetRecvTime(ths->pNextIndex, &(pMsg->srcId), tsMs);
+    resetElect = true;
 
-    syncNodeResetElectTimer(ths);
     ths->minMatchIndex = pMsg->minMatchIndex;
 
     if (ths->state == TAOS_SYNC_STATE_FOLLOWER) {
-      // syncNodeFollowerCommit(ths, pMsg->commitIndex);
       SRpcMsg rpcMsgLocalCmd = {0};
       (void)syncBuildLocalCmd(&rpcMsgLocalCmd, ths->vgId);
 
@@ -2328,7 +2331,6 @@ int32_t syncNodeOnHeartbeat(SSyncNode* ths, const SRpcMsg* pRpcMsg) {
   }
 
   if (pMsg->term >= currentTerm && ths->state != TAOS_SYNC_STATE_FOLLOWER) {
-    // syncNodeStepDown(ths, pMsg->term);
     SRpcMsg rpcMsgLocalCmd = {0};
     (void)syncBuildLocalCmd(&rpcMsgLocalCmd, ths->vgId);
 
@@ -2348,15 +2350,10 @@ int32_t syncNodeOnHeartbeat(SSyncNode* ths, const SRpcMsg* pRpcMsg) {
     }
   }
 
-  /*
-    // htonl
-    SMsgHead* pHead = rpcMsg.pCont;
-    pHead->contLen = htonl(pHead->contLen);
-    pHead->vgId = htonl(pHead->vgId);
-  */
-
   // reply
   syncNodeSendMsgById(&pMsgReply->destId, ths, &rpcMsg);
+
+  if (resetElect) syncNodeResetElectTimer(ths);
   return 0;
 }
 

source/libs/sync/src/syncMessage.c

@@ -22,7 +22,7 @@ int32_t syncBuildTimeout(SRpcMsg* pMsg, ESyncTimeoutType timeoutType, uint64_t l
                          SSyncNode* pNode) {
   int32_t bytes = sizeof(SyncTimeout);
   pMsg->pCont = rpcMallocCont(bytes);
-  pMsg->msgType = TDMT_SYNC_TIMEOUT;
+  pMsg->msgType = (timeoutType == SYNC_TIMEOUT_ELECTION) ? TDMT_SYNC_TIMEOUT_ELECTION : TDMT_SYNC_TIMEOUT;
   pMsg->contLen = bytes;
   if (pMsg->pCont == NULL) {
     terrno = TSDB_CODE_OUT_OF_MEMORY;
@@ -31,7 +31,7 @@ int32_t syncBuildTimeout(SRpcMsg* pMsg, ESyncTimeoutType timeoutType, uint64_t l
 
   SyncTimeout* pTimeout = pMsg->pCont;
   pTimeout->bytes = bytes;
-  pTimeout->msgType = TDMT_SYNC_TIMEOUT;
+  pTimeout->msgType = pMsg->msgType;
   pTimeout->vgId = pNode->vgId;
   pTimeout->timeoutType = timeoutType;
   pTimeout->logicClock = logicClock;

source/libs/sync/src/syncRequestVote.c

@@ -89,6 +89,7 @@ static bool syncNodeOnRequestVoteLogOK(SSyncNode* ths, SyncRequestVote* pMsg) {
 int32_t syncNodeOnRequestVote(SSyncNode* ths, const SRpcMsg* pRpcMsg) {
   int32_t          ret = 0;
   SyncRequestVote* pMsg = pRpcMsg->pCont;
+  bool             resetElect = false;
 
   // if already drop replica, do not process
   if (!syncNodeInRaftGroup(ths, &pMsg->srcId)) {
@@ -115,7 +116,7 @@ int32_t syncNodeOnRequestVote(SSyncNode* ths, const SRpcMsg* pRpcMsg) {
     syncNodeStepDown(ths, currentTerm);
 
     // forbid elect for this round
-    syncNodeResetElectTimer(ths);
+    resetElect = true;
   }
 
   // send msg
@@ -134,5 +135,7 @@ int32_t syncNodeOnRequestVote(SSyncNode* ths, const SRpcMsg* pRpcMsg) {
   syncLogRecvRequestVote(ths, pMsg, pReply->voteGranted, "");
   syncLogSendRequestVoteReply(ths, pReply, "");
   syncNodeSendMsgById(&pReply->destId, ths, &rpcMsg);
+
+  if (resetElect) syncNodeResetElectTimer(ths);
   return 0;
 }

source/libs/sync/src/syncSnapshot.c

@@ -798,7 +798,6 @@ int32_t syncNodeOnSnapshot(SSyncNode *pSyncNode, const SRpcMsg *pRpcMsg) {
   if (pMsg->term > raftStoreGetTerm(pSyncNode)) {
     syncNodeStepDown(pSyncNode, pMsg->term);
   }
-  syncNodeResetElectTimer(pSyncNode);
 
   // state, term, seq/ack
   int32_t code = 0;
@@ -840,6 +839,7 @@ int32_t syncNodeOnSnapshot(SSyncNode *pSyncNode, const SRpcMsg *pRpcMsg) {
     code = -1;
   }
 
+  syncNodeResetElectTimer(pSyncNode);
   return code;
 }
 

source/libs/sync/src/syncTimeout.c

@@ -120,9 +120,6 @@ int32_t syncNodeOnTimeout(SSyncNode* ths, const SRpcMsg* pRpc) {
     if (atomic_load_64(&ths->pingTimerLogicClockUser) <= pMsg->logicClock) {
       ++(ths->pingTimerCounter);
 
-      // syncNodePingAll(ths);
-      // syncNodePingPeers(ths);
-
       syncNodeTimerRoutine(ths);
     }
 
@@ -138,8 +135,6 @@ int32_t syncNodeOnTimeout(SSyncNode* ths, const SRpcMsg* pRpc) {
       ++(ths->heartbeatTimerCounter);
       sTrace("vgId:%d, sync timer, type:replicate count:%" PRIu64 ", lc-user:%" PRIu64, ths->vgId,
              ths->heartbeatTimerCounter, ths->heartbeatTimerLogicClockUser);
-
-      // syncNodeReplicate(ths, true);
     }
 
   } else {

source/util/src/tarray.c

@@ -140,7 +140,7 @@ void taosArrayRemoveDuplicate(SArray* pArray, __compar_fn_t comparFn, void (*fp)
 
         taosArraySet(pArray, pos + 1, p2);
         memset(TARRAY_GET_ELEM(pArray, i), 0, pArray->elemSize);
-	pos += 1;
+        pos += 1;
       } else {
         pos += 1;
       }
@@ -157,45 +157,6 @@ void taosArrayRemoveDuplicate(SArray* pArray, __compar_fn_t comparFn, void (*fp)
   pArray->size = pos + 1;
 }
 
-void taosArrayRemoveDuplicateP(SArray* pArray, __compar_fn_t comparFn, void (*fp)(void*)) {
-  size_t size = pArray->size;
-  if (size <= 1) {
-    return;
-  }
-
-  int32_t pos = 0;
-  for (int32_t i = 1; i < size; ++i) {
-    char* p1 = taosArrayGet(pArray, pos);
-    char* p2 = taosArrayGet(pArray, i);
-
-    if (comparFn(p1, p2) == 0) {
-      // do nothing
-    } else {
-      if (pos + 1 != i) {
-        void* p = taosArrayGetP(pArray, pos + 1);
-        if (fp != NULL) {
-          fp(p);
-        }
-
-        taosArraySet(pArray, pos + 1, p2);
-        memset(TARRAY_GET_ELEM(pArray, i), 0, pArray->elemSize);
-	pos += 1;
-      } else {
-        pos += 1;
-      }
-    }
-  }
-
-  if (fp != NULL) {
-    for (int32_t i = pos + 1; i < pArray->size; ++i) {
-      void* p = taosArrayGetP(pArray, i);
-      fp(p);
-    }
-  }
-
-  pArray->size = pos + 1;
-}
-
 void* taosArrayAddAll(SArray* pArray, const SArray* pInput) {
   if (pInput) {
     return taosArrayAddBatch(pArray, pInput->pData, (int32_t)taosArrayGetSize(pInput));
@@ -392,20 +353,6 @@ void taosArrayClearEx(SArray* pArray, void (*fp)(void*)) {
   pArray->size = 0;
 }
 
-void taosArrayClearP(SArray* pArray, FDelete fp) {
-  if (pArray == NULL) return;
-  if (fp == NULL) {
-    pArray->size = 0;
-    return;
-  }
-
-  for (int32_t i = 0; i < pArray->size; ++i) {
-    fp(*(void**)TARRAY_GET_ELEM(pArray, i));
-  }
-
-  pArray->size = 0;
-}
-
 void* taosArrayDestroy(SArray* pArray) {
   if (pArray) {
     taosMemoryFree(pArray->pData);
@@ -495,6 +442,7 @@ static void taosArrayInsertSort(SArray* pArray, __ext_compar_fn_t fn, const void
   if (pArray->size <= 1) {
     return;
   }
+
   for (int32_t i = 1; i <= pArray->size - 1; ++i) {
     for (int32_t j = i; j > 0; --j) {
       if (fn(taosArrayGetP(pArray, j), taosArrayGetP(pArray, j - 1), param) == -1) {
@@ -507,7 +455,6 @@ static void taosArrayInsertSort(SArray* pArray, __ext_compar_fn_t fn, const void
       }
     }
   }
-  return;
 }
 
 int32_t taosEncodeArray(void** buf, const SArray* pArray, FEncode encode) {
@@ -556,4 +503,4 @@ void taosArraySwap(SArray* a, SArray* b) {
   void* data = a->pData;
   a->pData = b->pData;
   b->pData = data;
-}
+}

source/util/src/tconfig.c

@@ -637,6 +637,8 @@ int32_t cfgLoadFromEnvVar(SConfig *pConfig) {
   int32_t code = 0;
   char  **pEnv = environ;
   line[1023] = 0;
+
+  if (pEnv == NULL) return 0;
   while (*pEnv != NULL) {
     name = value = value2 = value3 = NULL;
     olen = vlen = vlen2 = vlen3 = 0;

source/util/src/tsimplehash.c

@@ -28,19 +28,23 @@
 
 #define HASH_INDEX(v, c) ((v) & ((c)-1))
 
-#define FREE_HASH_NODE(_n)   \
+#define FREE_HASH_NODE(_n, fp) \
-  do {                       \
+  do {                         \
-    taosMemoryFreeClear(_n); \
+    if (fp) {                  \
+      fp((_n)->data);          \
+    }                          \
+    taosMemoryFreeClear(_n);   \
   } while (0);
 
 struct SSHashObj {
-  SHNode    **hashList;
+  SHNode        **hashList;
-  size_t      capacity;  // number of slots
+  size_t          capacity;  // number of slots
-  int64_t     size;      // number of elements in hash table
+  int64_t         size;      // number of elements in hash table
-  _hash_fn_t  hashFp;    // hash function
+  _hash_fn_t      hashFp;    // hash function
-  _equal_fn_t equalFp;   // equal function
+  _equal_fn_t     equalFp;   // equal function
-  SArray*     pHashNodeBuf;// hash node allocation buffer, 1k size of each page by default
+  _hash_free_fn_t freeFp;    // free function
-  int32_t     offset;      // allocation offset in current page
+  SArray         *pHashNodeBuf;  // hash node allocation buffer, 1k size of each page by default
+  int32_t         offset;        // allocation offset in current page
 };
 
 static FORCE_INLINE int32_t taosHashCapacity(int32_t length) {
@@ -71,7 +75,7 @@ SSHashObj *tSimpleHashInit(size_t capacity, _hash_fn_t fn) {
   pHashObj->capacity = taosHashCapacity((int32_t)capacity);
   pHashObj->equalFp = memcmp;
 
-  pHashObj->pHashNodeBuf = taosArrayInit(10, sizeof(void*));
+  pHashObj->freeFp = NULL;
   pHashObj->offset = 0;
   pHashObj->size = 0;
   
@@ -92,6 +96,10 @@ int32_t tSimpleHashGetSize(const SSHashObj *pHashObj) {
   return (int32_t) pHashObj->size;
 }
 
+void tSimpleHashSetFreeFp(SSHashObj* pHashObj, _hash_free_fn_t freeFp) {
+  pHashObj->freeFp = freeFp;
+}
+
 static void* doInternalAlloc(SSHashObj* pHashObj, int32_t size) {
 #if 0
   void** p = taosArrayGetLast(pHashObj->pHashNodeBuf);
@@ -306,7 +314,8 @@ int32_t tSimpleHashRemove(SSHashObj *pHashObj, const void *key, size_t keyLen) {
       } else {
         pPrev->next = pNode->next;
       }
-      FREE_HASH_NODE(pNode);
+
+      FREE_HASH_NODE(pNode, pHashObj->freeFp);
       pHashObj->size -= 1;
       code = TSDB_CODE_SUCCESS;
       break;
@@ -341,7 +350,7 @@ int32_t tSimpleHashIterateRemove(SSHashObj *pHashObj, const void *key, size_t ke
         *pIter = pPrev ? GET_SHASH_NODE_DATA(pPrev) : NULL;
       }
 
-      FREE_HASH_NODE(pNode);
+      FREE_HASH_NODE(pNode, pHashObj->freeFp);
       pHashObj->size -= 1;
       break;
     }
@@ -370,14 +379,13 @@ void tSimpleHashClear(SSHashObj *pHashObj) {
 
     while (pNode) {
       pNext = pNode->next;
-      FREE_HASH_NODE(pNode);
+      FREE_HASH_NODE(pNode, pHashObj->freeFp);
       pNode = pNext;
     }
 
     pHashObj->hashList[i] = NULL;
   }
 
-  taosArrayClearEx(pHashObj->pHashNodeBuf, destroyItems);
   pHashObj->offset = 0;
   pHashObj->size = 0;
 }
@@ -388,7 +396,6 @@ void tSimpleHashCleanup(SSHashObj *pHashObj) {
   }
 
   tSimpleHashClear(pHashObj);
-  taosArrayDestroy(pHashObj->pHashNodeBuf);
   taosMemoryFreeClear(pHashObj->hashList);
   taosMemoryFree(pHashObj);
 }

source/util/src/tversion.c

@@ -26,8 +26,9 @@ int32_t taosVersionStrToInt(const char *vstr, int32_t *vint) {
   int32_t vnum[4] = {0};
   int32_t len = strlen(vstr);
   char    tmp[16] = {0};
+  int32_t vpos = 0;
 
-  for (int32_t spos = 0, tpos = 0, vpos = 0; spos < len && vpos < 4; ++spos) {
+  for (int32_t spos = 0, tpos = 0; spos < len && vpos < 4; ++spos) {
     if (vstr[spos] != '.') {
       tmp[spos - tpos] = vstr[spos];
     } else {
@@ -38,6 +39,10 @@ int32_t taosVersionStrToInt(const char *vstr, int32_t *vint) {
     }
   }
 
+  if ('\0' != tmp[0] && vpos < 4) {
+    vnum[vpos] = atoi(tmp);
+  }
+
   if (vnum[0] <= 0) {
     terrno = TSDB_CODE_INVALID_VERSION_STRING;
     return -1;
@@ -66,16 +71,16 @@ int32_t taosCheckVersionCompatible(int32_t clientVer, int32_t serverVer, int32_t
     case 4:
       break;
     case 3:
-      clientVer %= 100;
+      clientVer /= 100;
-      serverVer %= 100;
+      serverVer /= 100;
       break;
     case 2:
-      clientVer %= 10000;
+      clientVer /= 10000;
-      serverVer %= 10000;
+      serverVer /= 10000;
       break;
     case 1:
-      clientVer %= 1000000;
+      clientVer /= 1000000;
-      serverVer %= 1000000;
+      serverVer /= 1000000;
       break;
     default:
       terrno = TSDB_CODE_INVALID_VERSION_NUMBER;

source/util/src/version.c.in

@@ -1,6 +1,7 @@
 char version[64] = "${TD_VER_NUMBER}";
 char compatible_version[12] = "${TD_VER_COMPATIBLE}";
 char gitinfo[48] = "${TD_VER_GIT}";
+char gitinfoOfInternal[48] = "${TD_VER_GIT_INTERNAL}";
 char buildinfo[64] = "Built at ${TD_VER_DATE}";
 
 void libtaos_${TD_LIB_VER_NUMBER}_${TD_VER_OSTYPE}_${TD_VER_CPUTYPE}_${TD_VER_VERTYPE}() {};

tests/develop-test/5-taos-tools/taosbenchmark/json/sml_telnet_tcp.json

@@ -16,7 +16,7 @@
   "num_of_records_per_req": 10,
   "databases": [{
     "dbinfo": {
-      "name": "db",
+      "name": "opentsdb_telnet",
       "drop": "yes"
     },
     "super_tables": [{

tests/parallel_test/cases.task

@@ -10,18 +10,21 @@
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_str.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_math.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_time.py
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_26.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_str.py -Q 2
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_math.py -Q 2
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_time.py -Q 2
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery.py -Q 2
-,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery.py -Q 3
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_26.py -Q 2
-,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_str.py -Q 3
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/columnLenUpdated.py 
-,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_math.py -Q 3
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/columnLenUpdated.py -Q 2
-,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_time.py -Q 3
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/columnLenUpdated.py -Q 3
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/columnLenUpdated.py -Q 4
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery.py -Q 4
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_str.py -Q 4
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_math.py -Q 4
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_time.py -Q 4
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_26.py -Q 4
 ,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/tmqShow.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/tmqDropStb.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/subscribeStb0.py
@@ -43,6 +46,11 @@
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/concat.py -Q 3
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/out_of_order.py -Q 2
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/out_of_order.py -Q 4
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery.py -Q 3
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_str.py -Q 3
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_math.py -Q 3
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_time.py -Q 3
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/nestedQuery_26.py -Q 3
 
 ,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/create_wrong_topic.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/dropDbR3ConflictTransaction.py -N 3
@@ -93,6 +101,7 @@
 ,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/dataFromTsdbNWal.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/dataFromTsdbNWal-multiCtb.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/tmq_taosx.py
+,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/raw_block_interface_test.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/stbTagFilter-multiCtb.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/tmqSubscribeStb-r3.py -N 5
 ,,y,system-test,./pytest.sh python3 ./test.py -f 7-tmq/tmq3mnodeSwitch.py -N 6 -M 3 
@@ -864,6 +873,8 @@
 ,,y,script,./test.sh -f tsim/query/session.sim
 ,,y,script,./test.sh -f tsim/query/udf.sim
 ,,y,script,./test.sh -f tsim/query/udf_with_const.sim
+,,y,script,./test.sh -f tsim/query/join_interval.sim
+,,y,script,./test.sh -f tsim/query/unionall_as_table.sim
 ,,y,script,./test.sh -f tsim/query/sys_tbname.sim
 ,,y,script,./test.sh -f tsim/query/groupby.sim
 ,,y,script,./test.sh -f tsim/query/event.sim
@@ -1102,9 +1113,9 @@
 ,,n,develop-test,python3 ./test.py -f 5-taos-tools/taosbenchmark/json_tag.py
 ,,n,develop-test,python3 ./test.py -f 5-taos-tools/taosbenchmark/query_json.py
 ,,n,develop-test,python3 ./test.py -f 5-taos-tools/taosbenchmark/sample_csv_json.py
-#,,n,develop-test,python3 ./test.py -f 5-taos-tools/taosbenchmark/sml_json_alltypes.py
+,,n,develop-test,python3 ./test.py -f 5-taos-tools/taosbenchmark/sml_json_alltypes.py
 ,,n,develop-test,python3 ./test.py -f 5-taos-tools/taosbenchmark/taosdemoTestQueryWithJson.py -R
-#,,n,develop-test,python3 ./test.py -f 5-taos-tools/taosbenchmark/telnet_tcp.py -R
+,,n,develop-test,python3 ./test.py -f 5-taos-tools/taosbenchmark/telnet_tcp.py -R
 
 #docs-examples test
 ,,n,docs-examples-test,bash python.sh

tests/parallel_test/run.sh

@@ -167,8 +167,10 @@ function run_thread() {
         local case_build_san=`echo "$line"|cut -d, -f3`
         if [ "${case_build_san}" == "y" ]; then
             case_build_san="y"
+            DEBUGPATH="debugSan"
         elif [[ "${case_build_san}" == "n" ]] || [[ "${case_build_san}" == "" ]]; then
             case_build_san="n"
+            DEBUGPATH="debugNoSan"
         else
             usage
             exit 1
@@ -301,10 +303,10 @@ function run_thread() {
             if [ ! -z "$corefile" ]; then
                 echo -e "\e[34m corefiles: $corefile \e[0m"
                 local build_dir=$log_dir/build_${hosts[index]}
-                local remote_build_dir="${workdirs[index]}/TDengine/debug/build"
+                local remote_build_dir="${workdirs[index]}/{DEBUGPATH}/build"
-                if [ $ent -ne 0 ]; then
+                # if [ $ent -ne 0 ]; then
-                    remote_build_dir="${workdirs[index]}/TDinternal/debug/build"
+                #     remote_build_dir="${workdirs[index]}/{DEBUGPATH}/build"
-                fi
+                # fi
                 mkdir $build_dir 2>/dev/null
                 if [ $? -eq 0 ]; then
                     # scp build binary

tests/script/tsim/query/join_interval.sim

@@ -0,0 +1,42 @@
+
+system sh/stop_dnodes.sh
+system sh/deploy.sh -n dnode1 -i 1
+system sh/cfg.sh -n dnode1 -c udf -v 1
+system sh/exec.sh -n dnode1 -s start
+sql connect
+
+print ======== step create databases
+sql create database d1
+sql create database d2
+sql create table d1.t1(ts timestamp, i int) tags(t int);
+sql create table d2.t1(ts timestamp, i int);
+sql insert into d1.t11 using d1.t1 tags(1) values(1500000000000, 0)(1500000000001, 1)(1500000000002,2)(1500000000003,3)(1500000000004,4)
+sql insert into d1.t12 using d1.t1 tags(2) values(1500000000000, 0)(1500000000001, 1)(1500000000002,2)(1500000000003,3)(1500000000004,4)
+sql insert into d1.t13 using d1.t1 tags(3) values(1500000000000, 0)(1500000000001, 1)(1500000000002,2)(1500000000003,3)(1500000000004,4)
+
+sql insert into d2.t1 values(1500000000000,0)(1500000000001,1)(1500000000002,2)
+
+sql select _wstart,_wend,count((a.ts)),count(b.ts) from d1.t1 a, d2.t1 b where a.ts is not null and a.ts = b.ts  interval(1a) ;
+if $data02 != 3 then
+  return -1
+endi
+
+if $data03 != 3 then
+  return -1
+endi
+
+if $data12 != 3 then
+  return -1
+endi
+
+if $data13 != 3 then
+  return -1
+endi
+if $data22 != 3 then
+  return -1
+endi
+
+if $data23 != 3 then
+  return -1
+endi
+system sh/exec.sh -n dnode1 -s stop -x SIGINT