From 17a65430cda200c871086b4c25af4c48bbe52dde Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Fri, 22 Mar 2024 18:42:54 +0800
Subject: [PATCH 01/32] fix(tsdb): remove assert

---
 source/dnode/vnode/src/tsdb/tsdbRead2.c    | 5 +++++
 source/dnode/vnode/src/tsdb/tsdbReadUtil.c | 3 +--
 2 files changed, 6 insertions(+), 2 deletions(-)

diff --git a/source/dnode/vnode/src/tsdb/tsdbRead2.c b/source/dnode/vnode/src/tsdb/tsdbRead2.c
index b05ce0680b..cda8330282 100644
--- a/source/dnode/vnode/src/tsdb/tsdbRead2.c
+++ b/source/dnode/vnode/src/tsdb/tsdbRead2.c
@@ -3101,6 +3101,7 @@ static void initBlockDumpInfo(STsdbReader* pReader, SDataBlockIter* pBlockIter)
   SFileBlockDumpInfo* pDumpInfo = &pStatus->fBlockDumpInfo;
 
   if (pBlockInfo) {
+    // todo handle
 //    STableBlockScanInfo* pScanInfo = tSimpleHashGet(pBlockIter->pTableMap, &pBlockInfo->uid, sizeof(pBlockInfo->uid));
 //    if (pScanInfo) {
 //      tsdbRowKeyAssign(&pDumpInfo->lastKey, &pScanInfo->lastProcKey);
@@ -4666,9 +4667,13 @@ int32_t tsdbNextDataBlock2(STsdbReader* pReader, bool* hasNext) {
       tsdbReleaseReader(pReader);
     }
 
+    tsdbReaderSuspend2(pReader);
+    tsdbReaderResume2(pReader);
+
     return code;
   }
 
+
   if (pReader->step == EXTERNAL_ROWS_MAIN && pReader->innerReader[1] != NULL) {
     // prepare for the next row scan
     int32_t step = -1;
diff --git a/source/dnode/vnode/src/tsdb/tsdbReadUtil.c b/source/dnode/vnode/src/tsdb/tsdbReadUtil.c
index c7077a2e71..93bc5ca573 100644
--- a/source/dnode/vnode/src/tsdb/tsdbReadUtil.c
+++ b/source/dnode/vnode/src/tsdb/tsdbReadUtil.c
@@ -229,8 +229,7 @@ void resetAllDataBlockScanInfo(SSHashObj* pTableMap, int64_t ts, int32_t step) {
 
     pInfo->delSkyline = taosArrayDestroy(pInfo->delSkyline);
     pInfo->lastProcKey.ts = ts;
-    ASSERT(0);
-
+    // todo check the nextProcKey info
     pInfo->sttKeyInfo.nextProcKey = ts + step;
   }
 }

From c35e834141d722d3f4162ed5684eecea3f5c5442 Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Sat, 23 Mar 2024 00:11:21 +0800
Subject: [PATCH 02/32] fix(tsdb): fix error.

---
 source/dnode/vnode/src/tsdb/tsdbRead2.c | 44 +++++++++++++++----------
 1 file changed, 27 insertions(+), 17 deletions(-)

diff --git a/source/dnode/vnode/src/tsdb/tsdbRead2.c b/source/dnode/vnode/src/tsdb/tsdbRead2.c
index cda8330282..7cd624d7c7 100644
--- a/source/dnode/vnode/src/tsdb/tsdbRead2.c
+++ b/source/dnode/vnode/src/tsdb/tsdbRead2.c
@@ -148,6 +148,7 @@ static void tRowGetKeyDeepCopy(SRow* pRow, SRowKey* pKey) {
   }
 
   SPrimaryKeyIndex indices[TD_MAX_PK_COLS];
+  ASSERT(pKey->numOfPKs <= TD_MAX_PK_COLS);
 
   uint8_t *data = pRow->data;
   for (int32_t i = 0; i < pRow->numOfPKs; i++) {
@@ -686,13 +687,14 @@ static int32_t doLoadFileBlock(STsdbReader* pReader, SArray* pIndexList, SBlockN
     STableBlockScanInfo* pScanInfo = getTableBlockScanInfo(pReader->status.pTableMap, uid, pReader->idStr);
 
     // todo: here we should find the first timestamp that is greater than the lastProcKey
+    // the window is an open interval NOW.
     if (asc) {
-      w.skey = pScanInfo->lastProcKey.ts + step;
+      w.skey = pScanInfo->lastProcKey.ts;
     } else {
-      w.ekey = pScanInfo->lastProcKey.ts + step;
+      w.ekey = pScanInfo->lastProcKey.ts;
     }
 
-    if (isEmptyQueryTimeWindow(&w)) {
+    if (/*isEmptyQueryTimeWindow(&w)*/ w.ekey - w.skey < 2) {  // NOTE: specialized for open interval
       k += 1;
 
       if (k >= numOfTables) {
@@ -707,7 +709,11 @@ static int32_t doLoadFileBlock(STsdbReader* pReader, SArray* pIndexList, SBlockN
       continue;
     }
 
-    // 2. version range check
+    if (pkCompEx(pReader->pkComparFn, &pRecord->lastKey.key, &pScanInfo->lastProcKey) <= 0) {
+      continue;
+    }
+
+    // 2. version range check, version range is an CLOSED interval
     if (pRecord->minVer > pReader->info.verRange.maxVer || pRecord->maxVer < pReader->info.verRange.minVer) {
       continue;
     }
@@ -724,6 +730,7 @@ static int32_t doLoadFileBlock(STsdbReader* pReader, SArray* pIndexList, SBlockN
     if (pScanInfo->filesetWindow.skey > pRecord->firstKey.key.ts) {
       pScanInfo->filesetWindow.skey = pRecord->firstKey.key.ts;
     }
+
     if (pScanInfo->filesetWindow.ekey < pRecord->lastKey.key.ts) {
       pScanInfo->filesetWindow.ekey = pRecord->lastKey.key.ts;
     }
@@ -1637,9 +1644,11 @@ static int32_t doMergeBufAndFileRows(STsdbReader* pReader, STableBlockScanInfo*
   __compar_fn_t       compFn = pReader->pkComparFn;
   int32_t             pkSrcSlot = pReader->suppInfo.pkSrcSlot;
 
-  SRowKey* pSttKey = NULL;
+  SRowKey* pSttKey = &(SRowKey){0};
   if (hasDataInSttBlock(pBlockScanInfo) && (!pBlockScanInfo->cleanSttBlocks)) {
-    pSttKey = getCurrentKeyInSttBlock(pSttBlockReader);
+    tRowKeyAssign(pSttKey, getCurrentKeyInSttBlock(pSttBlockReader));
+  } else {
+    pSttKey = NULL;
   }
 
   SRowKey k;
@@ -1675,21 +1684,21 @@ static int32_t doMergeBufAndFileRows(STsdbReader* pReader, STableBlockScanInfo*
   if (pReader->info.order == TSDB_ORDER_ASC) {
     minKey = k;  // chosen the minimum value
 
-    if (pkCompEx(compFn, pfKey, &minKey) < 0) {
+    if (pfKey != NULL && pkCompEx(compFn, pfKey, &minKey) < 0) {
       minKey = *pfKey;
     }
 
-    if (pkCompEx(compFn, pSttKey, &minKey) < 0) {
+    if (pSttKey != NULL && pkCompEx(compFn, pSttKey, &minKey) < 0) {
       minKey = *pSttKey;
     }
   } else {
     minKey = k;
 
-    if (pkCompEx(compFn, pfKey, &minKey) > 0) {
+    if (pfKey != NULL && pkCompEx(compFn, pfKey, &minKey) > 0) {
       minKey = *pfKey;
     }
 
-    if (pkCompEx(compFn, pSttKey, &minKey) > 0) {
+    if (pSttKey != NULL && pkCompEx(compFn, pSttKey, &minKey) > 0) {
       minKey = *pSttKey;
     }
   }
@@ -1888,7 +1897,7 @@ static int32_t doMergeMultiLevelRows(STsdbReader* pReader, STableBlockScanInfo*
       minKey = ik;
     }
 
-    if (hasDataInFileBlock(pBlockData, pDumpInfo) && (pkCompEx(compFn, pfKey, &minKey) < 0)) {
+    if (pfKey != NULL && (pkCompEx(compFn, pfKey, &minKey) < 0)) {
       minKey = *pfKey;
     }
 
@@ -1901,7 +1910,7 @@ static int32_t doMergeMultiLevelRows(STsdbReader* pReader, STableBlockScanInfo*
       minKey = ik;
     }
 
-    if (hasDataInFileBlock(pBlockData, pDumpInfo) && (pkCompEx(compFn, pfKey, &minKey) > 0)) {
+    if (pfKey != NULL && (pkCompEx(compFn, pfKey, &minKey) > 0)) {
       minKey = *pfKey;
     }
 
@@ -3659,7 +3668,7 @@ int32_t doMergeMemTableMultiRows(TSDBROW* pRow, SRowKey* pKey, uint64_t uid, SIt
 
       SRowKey nextRowKey = {0};
       tRowGetKeyEx(pNextRow, &nextRowKey);
-      if (pkCompEx(pReader->pkComparFn, pKey, &nextRowKey) != 0) {
+      if (pKey->numOfPKs > 0 && pkCompEx(pReader->pkComparFn, pKey, &nextRowKey) != 0) {
         *pResRow = current;
         *freeTSRow = false;
         return TSDB_CODE_SUCCESS;
@@ -3966,6 +3975,9 @@ int32_t buildDataBlockFromBufImpl(STableBlockScanInfo* pBlockScanInfo, int64_t e
 
     if (row.type == TSDBROW_ROW_FMT) {
       code = doAppendRowFromTSRow(pBlock, pReader, row.pTSRow, pBlockScanInfo);
+      if (code == TSDB_CODE_SUCCESS) {
+        tRowGetKeyDeepCopy(row.pTSRow, &pBlockScanInfo->lastProcKey);
+      }
 
       if (freeTSRow) {
         taosMemoryFree(row.pTSRow);
@@ -3974,8 +3986,6 @@ int32_t buildDataBlockFromBufImpl(STableBlockScanInfo* pBlockScanInfo, int64_t e
       if (code) {
         return code;
       }
-
-      tRowGetKeyDeepCopy(row.pTSRow, &pBlockScanInfo->lastProcKey);
     } else {
       code = doAppendRowFromFileBlock(pBlock, pReader, row.pBlockData, row.iRow);
       if (code) {
@@ -4667,8 +4677,8 @@ int32_t tsdbNextDataBlock2(STsdbReader* pReader, bool* hasNext) {
       tsdbReleaseReader(pReader);
     }
 
-    tsdbReaderSuspend2(pReader);
-    tsdbReaderResume2(pReader);
+//    tsdbReaderSuspend2(pReader);
+//    tsdbReaderResume2(pReader);
 
     return code;
   }

From 447b8ddd32c5ba8002f8daf84efc43a05b2ff311 Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Sat, 23 Mar 2024 23:33:40 +0800
Subject: [PATCH 03/32] fix(tsdb): remove invalid assert.

---
 source/dnode/vnode/src/tsdb/tsdbRead2.c | 1 -
 1 file changed, 1 deletion(-)

diff --git a/source/dnode/vnode/src/tsdb/tsdbRead2.c b/source/dnode/vnode/src/tsdb/tsdbRead2.c
index 7cd624d7c7..731442af67 100644
--- a/source/dnode/vnode/src/tsdb/tsdbRead2.c
+++ b/source/dnode/vnode/src/tsdb/tsdbRead2.c
@@ -3628,7 +3628,6 @@ int32_t doMergeRowsInSttBlock(SSttBlockReader* pSttBlockReader, STableBlockScanI
       TSDBROW* pRow1 = tMergeTreeGetRow(&pSttBlockReader->mergeTree);
       tsdbRowMergerAdd(pMerger, pRow1, NULL);
     } else {
-      ASSERT(ret < 0);
       tsdbTrace("uid:%" PRIu64 " last del index:%d, del range:%d, lastKeyInStt:%" PRId64 ", %s", pScanInfo->uid,
                 pScanInfo->sttBlockDelIndex, (int32_t)taosArrayGetSize(pScanInfo->delSkyline),
                 pScanInfo->sttKeyInfo.nextProcKey, idStr);

From f9a6243859032e210f9b927abea175a683604096 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Mon, 25 Mar 2024 09:38:50 +0800
Subject: [PATCH 04/32] fix: add pk col after ts col to make operator of show
 table distributed super_table work

---
 source/libs/planner/src/planLogicCreater.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/source/libs/planner/src/planLogicCreater.c b/source/libs/planner/src/planLogicCreater.c
index c1793bfb90..c79fbd01f0 100644
--- a/source/libs/planner/src/planLogicCreater.c
+++ b/source/libs/planner/src/planLogicCreater.c
@@ -341,10 +341,11 @@ static int32_t addDefaultScanCol(const STableMeta* pMeta, SNodeList** pCols) {
   if (TSDB_SYSTEM_TABLE == pMeta->tableType) {
     return addSystableFirstCol(pMeta->uid, pMeta->schema, pCols, pMeta);
   }
-  if (hasPkInTable(pMeta)) {
-    addPkCol(pMeta->uid, pMeta->schema + 1, pCols, pMeta);
+  int32_t code = addPrimaryKeyCol(pMeta->uid, pMeta->schema, pCols, pMeta);
+  if (code == TSDB_CODE_SUCCESS && hasPkInTable(pMeta)) {
+    code = addPkCol(pMeta->uid, pMeta->schema + 1, pCols, pMeta);
   }
-  return addPrimaryKeyCol(pMeta->uid, pMeta->schema, pCols, pMeta);
+  return code;
 }
 
 static int32_t makeScanLogicNode(SLogicPlanContext* pCxt, SRealTableNode* pRealTable, bool hasRepeatScanFuncs,

From 073e4d838916cc0450220d6585a79c9ad3f7c986 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Mon, 25 Mar 2024 09:48:54 +0800
Subject: [PATCH 05/32] fix: increase last_row result size since SFirstLastRes
 has pk info

---
 tests/script/tsim/query/cache_last.sim | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tests/script/tsim/query/cache_last.sim b/tests/script/tsim/query/cache_last.sim
index 65eb46de69..50199117b0 100644
--- a/tests/script/tsim/query/cache_last.sim
+++ b/tests/script/tsim/query/cache_last.sim
@@ -89,7 +89,7 @@ if $data10 != @   -> Merge (columns=3 width=24 input_order=unknown output_order=
   return -1
 endi
 sql explain select count(*), last_row(f1), min(f1) from sta interval(1s);
-if $data10 != @   -> Merge (columns=4 width=66 input_order=asc output_order=asc mode=sort)@ then
+if $data10 != @   -> Merge (columns=4 width=82 input_order=asc output_order=asc mode=sort)@ then
   return -1
 endi
 sql explain select distinct count(*), last_row(f1), min(f1) from tba1;

From 75d6b1edde9d09aceb04d1c30948057169465453 Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Mon, 25 Mar 2024 09:59:10 +0800
Subject: [PATCH 06/32] fix(tsdb): fix the bug caused by sttKey changing.

---
 source/dnode/vnode/src/tsdb/tsdbRead2.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/source/dnode/vnode/src/tsdb/tsdbRead2.c b/source/dnode/vnode/src/tsdb/tsdbRead2.c
index 731442af67..6e134698b5 100644
--- a/source/dnode/vnode/src/tsdb/tsdbRead2.c
+++ b/source/dnode/vnode/src/tsdb/tsdbRead2.c
@@ -1801,6 +1801,7 @@ static int32_t mergeFileBlockAndSttBlock(STsdbReader* pReader, SSttBlockReader*
       }
     }
 
+    // pKey == pSttKey
     tRowKeyAssign(&pBlockScanInfo->lastProcKey, pKey);
 
     // the following for key == sttKey->key.ts
@@ -1819,7 +1820,8 @@ static int32_t mergeFileBlockAndSttBlock(STsdbReader* pReader, SSttBlockReader*
       return code;
     }
 
-    doMergeRowsInSttBlock(pSttBlockReader, pBlockScanInfo, pSttKey, pMerger, pkSrcSlot, &pReader->info.verRange,
+    // pSttKey will be changed when sttBlockReader iterates to the next row, so use pKey instead.
+    doMergeRowsInSttBlock(pSttBlockReader, pBlockScanInfo, pKey, pMerger, pkSrcSlot, &pReader->info.verRange,
                           pReader->idStr);
 
     code = tsdbRowMergerGetRow(pMerger, &pTSRow);
@@ -4675,10 +4677,6 @@ int32_t tsdbNextDataBlock2(STsdbReader* pReader, bool* hasNext) {
       qTrace("tsdb/read: %p, unlock read mutex", pReader);
       tsdbReleaseReader(pReader);
     }
-
-//    tsdbReaderSuspend2(pReader);
-//    tsdbReaderResume2(pReader);
-
     return code;
   }
 
@@ -4904,6 +4902,8 @@ SSDataBlock* tsdbRetrieveDataBlock2(STsdbReader* pReader, SArray* pIdList) {
   qTrace("tsdb/read-retrieve: %p, unlock read mutex", pReader);
   tsdbReleaseReader(pReader);
 
+//  tsdbReaderSuspend2(pReader);
+//  tsdbReaderResume2(pReader);
   return ret;
 }
 

From 9d2c310616aeb6ce7930968c3eea06eac9905533 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Mon, 25 Mar 2024 10:15:50 +0800
Subject: [PATCH 07/32] fix: remove the extract space after column type for
 none pk column

---
 source/libs/command/src/command.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/source/libs/command/src/command.c b/source/libs/command/src/command.c
index 2136ef7d6d..dc8fa49e27 100644
--- a/source/libs/command/src/command.c
+++ b/source/libs/command/src/command.c
@@ -501,8 +501,12 @@ void appendColumnFields(char* buf, int32_t* len, STableCfg* pCfg) {
     } else if (TSDB_DATA_TYPE_NCHAR == pSchema->type) {
       sprintf(type + strlen(type), "(%d)", (int32_t)((pSchema->bytes - VARSTR_HEADER_SIZE) / TSDB_NCHAR_SIZE));
     }
-    char* pk = (pSchema->flags & COL_IS_KEY) ? "PRIMARY KEY" : "";
-    *len += sprintf(buf + VARSTR_HEADER_SIZE + *len, "%s`%s` %s %s", ((i > 0) ? ", " : ""), pSchema->name, type, pk);
+    if (!(pSchema->flags & COL_IS_KEY)) {
+      *len += sprintf(buf + VARSTR_HEADER_SIZE + *len, "%s`%s` %s", ((i > 0) ? ", " : ""), pSchema->name, type);
+    } else {
+      char* pk = "PRIMARY KEY";
+      *len += sprintf(buf + VARSTR_HEADER_SIZE + *len, "%s`%s` %s %s", ((i > 0) ? ", " : ""), pSchema->name, type, pk);
+    }
   }
 }
 

From 4d3eb6824c5444b4c83dec982f27167687e87d4a Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Mon, 25 Mar 2024 13:38:34 +0800
Subject: [PATCH 08/32] feat: add pk support for last/first/last_row
 dynDataRequired

---
 include/libs/function/functionMgt.h     |  2 +-
 source/common/src/tdatablock.c          |  4 +-
 source/libs/executor/src/scanoperator.c |  2 +-
 source/libs/function/inc/builtins.h     |  2 +-
 source/libs/function/inc/builtinsimpl.h |  4 +-
 source/libs/function/src/builtinsimpl.c | 50 ++++++++++++++++++++++---
 source/libs/function/src/functionMgt.c  |  4 +-
 7 files changed, 53 insertions(+), 15 deletions(-)

diff --git a/include/libs/function/functionMgt.h b/include/libs/function/functionMgt.h
index f126bb587a..72e2815ade 100644
--- a/include/libs/function/functionMgt.h
+++ b/include/libs/function/functionMgt.h
@@ -256,7 +256,7 @@ typedef enum EFuncDataRequired {
 } EFuncDataRequired;
 
 EFuncDataRequired fmFuncDataRequired(SFunctionNode* pFunc, STimeWindow* pTimeWindow);
-EFuncDataRequired fmFuncDynDataRequired(int32_t funcId, void* pRes, STimeWindow* pTimeWindow);
+EFuncDataRequired fmFuncDynDataRequired(int32_t funcId, void* pRes, SDataBlockInfo* pBlockInfo);
 
 int32_t fmGetFuncExecFuncs(int32_t funcId, SFuncExecFuncs* pFpSet);
 int32_t fmGetScalarFuncExecFuncs(int32_t funcId, SScalarFuncExecFuncs* pFpSet);
diff --git a/source/common/src/tdatablock.c b/source/common/src/tdatablock.c
index 0d805cb8a4..d2164b024b 100644
--- a/source/common/src/tdatablock.c
+++ b/source/common/src/tdatablock.c
@@ -546,8 +546,8 @@ int32_t blockDataUpdatePkRange(SSDataBlock* pDataBlock, int32_t pkColumnIndex, b
     }
   } else {
     if (IS_NUMERIC_TYPE(pColInfoData->info.type)) {
-      pDataBlock->info.pks[0].val = *(int64_t*) ekey;
-      pDataBlock->info.pks[1].val = *(int64_t*) skey;
+      pDataBlock->info.pks[0].val = *(int32_t*) ekey;
+      pDataBlock->info.pks[1].val = *(int32_t*) skey;
     } else {  // todo refactor
       memcpy(pDataBlock->info.pks[0].pData, varDataVal(ekey), varDataLen(ekey));
       pDataBlock->info.pks[0].nData = varDataLen(ekey);
diff --git a/source/libs/executor/src/scanoperator.c b/source/libs/executor/src/scanoperator.c
index 1e5fa065de..a51e627272 100644
--- a/source/libs/executor/src/scanoperator.c
+++ b/source/libs/executor/src/scanoperator.c
@@ -201,7 +201,7 @@ static int32_t doDynamicPruneDataBlock(SOperatorInfo* pOperator, SDataBlockInfo*
 
     SResultRowEntryInfo* pEntry = getResultEntryInfo(pRow, i, pTableScanInfo->base.pdInfo.pExprSup->rowEntryInfoOffset);
 
-    int32_t reqStatus = fmFuncDynDataRequired(functionId, pEntry, &pBlockInfo->window);
+    int32_t reqStatus = fmFuncDynDataRequired(functionId, pEntry, pBlockInfo);
     if (reqStatus != FUNC_DATA_REQUIRED_NOT_LOAD) {
       notLoadBlock = false;
       break;
diff --git a/source/libs/function/inc/builtins.h b/source/libs/function/inc/builtins.h
index 4c1e46dbba..af33c3ffa2 100644
--- a/source/libs/function/inc/builtins.h
+++ b/source/libs/function/inc/builtins.h
@@ -25,7 +25,7 @@ extern "C" {
 typedef int32_t (*FTranslateFunc)(SFunctionNode* pFunc, char* pErrBuf, int32_t len);
 typedef EFuncDataRequired (*FFuncDataRequired)(SFunctionNode* pFunc, STimeWindow* pTimeWindow);
 typedef int32_t (*FCreateMergeFuncParameters)(SNodeList* pRawParameters, SNode* pPartialRes, SNodeList** pParameters);
-typedef EFuncDataRequired (*FFuncDynDataRequired)(void* pRes, STimeWindow* pTimeWindow);
+typedef EFuncDataRequired (*FFuncDynDataRequired)(void* pRes, SDataBlockInfo* pBlocInfo);
 typedef EFuncReturnRows (*FEstimateReturnRows)(SFunctionNode* pFunc);
 
 typedef struct SBuiltinFuncDefinition {
diff --git a/source/libs/function/inc/builtinsimpl.h b/source/libs/function/inc/builtinsimpl.h
index ccb50ff2a8..54bcd26bc3 100644
--- a/source/libs/function/inc/builtinsimpl.h
+++ b/source/libs/function/inc/builtinsimpl.h
@@ -158,8 +158,8 @@ int32_t           firstLastPartialFinalize(SqlFunctionCtx* pCtx, SSDataBlock* pB
 int32_t           firstCombine(SqlFunctionCtx* pDestCtx, SqlFunctionCtx* pSourceCtx);
 int32_t           lastCombine(SqlFunctionCtx* pDestCtx, SqlFunctionCtx* pSourceCtx);
 int32_t           getFirstLastInfoSize(int32_t resBytes, int32_t pkBytes);
-EFuncDataRequired firstDynDataReq(void* pRes, STimeWindow* pTimeWindow);
-EFuncDataRequired lastDynDataReq(void* pRes, STimeWindow* pTimeWindow);
+EFuncDataRequired firstDynDataReq(void* pRes, SDataBlockInfo* pBlockInfo);
+EFuncDataRequired lastDynDataReq(void* pRes, SDataBlockInfo* pBlockInfo);
 
 int32_t lastRowFunction(SqlFunctionCtx* pCtx);
 
diff --git a/source/libs/function/src/builtinsimpl.c b/source/libs/function/src/builtinsimpl.c
index 097b263b8e..83f9dcff08 100644
--- a/source/libs/function/src/builtinsimpl.c
+++ b/source/libs/function/src/builtinsimpl.c
@@ -2290,7 +2290,31 @@ int32_t apercentileCombine(SqlFunctionCtx* pDestCtx, SqlFunctionCtx* pSourceCtx)
   return TSDB_CODE_SUCCESS;
 }
 
-EFuncDataRequired firstDynDataReq(void* pRes, STimeWindow* pTimeWindow) {
+// TODO: change this function when block data info pks changed
+static int32_t comparePkDataWithSValue(int8_t pkType, char* pkData, SValue* pVal, int32_t order) {
+  char numVal[8] = {0};
+  switch (pkType) {
+    case TSDB_DATA_TYPE_INT:
+      *(int32_t*)numVal = (int32_t)pVal->val;
+      break;
+    case TSDB_DATA_TYPE_UINT:
+      *(uint32_t*)numVal = (uint32_t)pVal->val;
+      break;
+    case TSDB_DATA_TYPE_BIGINT:
+      *(int64_t*)numVal = (int64_t)pVal->val;
+      break;
+    case TSDB_DATA_TYPE_UBIGINT:
+      *(uint64_t*)numVal = (uint64_t)pVal->val;
+      break;
+    default:
+      break;
+  }
+  char*         blockData = (IS_NUMERIC_TYPE(pkType)) ? (char*) numVal : (char*)pVal->pData;
+  __compar_fn_t fn = getKeyComparFunc(pkType, order);
+  return fn(pkData, blockData);
+}
+
+EFuncDataRequired firstDynDataReq(void* pRes, SDataBlockInfo* pBlockInfo) {
   SResultRowEntryInfo* pEntry = (SResultRowEntryInfo*)pRes;
 
   // not initialized yet, data is required
@@ -2299,14 +2323,21 @@ EFuncDataRequired firstDynDataReq(void* pRes, STimeWindow* pTimeWindow) {
   }
 
   SFirstLastRes* pResult = GET_ROWCELL_INTERBUF(pEntry);
-  if (pResult->hasResult && pResult->ts <= pTimeWindow->skey) {
-    return FUNC_DATA_REQUIRED_NOT_LOAD;
+  if (pResult->hasResult) {
+    if (pResult->ts < pBlockInfo->window.skey) {
+      return FUNC_DATA_REQUIRED_NOT_LOAD;
+    } else if (pResult->ts == pBlockInfo->window.skey && pResult->pkData) {
+      if (comparePkDataWithSValue(pResult->pkType, pResult->pkData, pBlockInfo->pks + 0, TSDB_ORDER_ASC) < 0) {
+        return FUNC_DATA_REQUIRED_NOT_LOAD;
+      }
+    }
+    return FUNC_DATA_REQUIRED_DATA_LOAD;
   } else {
     return FUNC_DATA_REQUIRED_DATA_LOAD;
   }
 }
 
-EFuncDataRequired lastDynDataReq(void* pRes, STimeWindow* pTimeWindow) {
+EFuncDataRequired lastDynDataReq(void* pRes, SDataBlockInfo* pBlockInfo) {
   SResultRowEntryInfo* pEntry = (SResultRowEntryInfo*)pRes;
 
   // not initialized yet, data is required
@@ -2315,8 +2346,15 @@ EFuncDataRequired lastDynDataReq(void* pRes, STimeWindow* pTimeWindow) {
   }
 
   SFirstLastRes* pResult = GET_ROWCELL_INTERBUF(pEntry);
-  if (pResult->hasResult && pResult->ts >= pTimeWindow->ekey) {
-    return FUNC_DATA_REQUIRED_NOT_LOAD;
+  if (pResult->hasResult) {
+    if (pResult->ts > pBlockInfo->window.ekey) {
+      return FUNC_DATA_REQUIRED_NOT_LOAD;
+    } else if (pResult->ts == pBlockInfo->window.ekey && pResult->pkData) {
+      if (comparePkDataWithSValue(pResult->pkType, pResult->pkData, pBlockInfo->pks + 1, TSDB_ORDER_DESC) < 0) {
+        return FUNC_DATA_REQUIRED_NOT_LOAD;
+      }
+    }
+    return FUNC_DATA_REQUIRED_DATA_LOAD;
   } else {
     return FUNC_DATA_REQUIRED_DATA_LOAD;
   }
diff --git a/source/libs/function/src/functionMgt.c b/source/libs/function/src/functionMgt.c
index ae3958647b..697efb63fd 100644
--- a/source/libs/function/src/functionMgt.c
+++ b/source/libs/function/src/functionMgt.c
@@ -115,7 +115,7 @@ EFuncDataRequired fmFuncDataRequired(SFunctionNode* pFunc, STimeWindow* pTimeWin
   return funcMgtBuiltins[pFunc->funcId].dataRequiredFunc(pFunc, pTimeWindow);
 }
 
-EFuncDataRequired fmFuncDynDataRequired(int32_t funcId, void* pRes, STimeWindow* pTimeWindow) {
+EFuncDataRequired fmFuncDynDataRequired(int32_t funcId, void* pRes, SDataBlockInfo* pBlockInfo) {
   if (fmIsUserDefinedFunc(funcId) || funcId < 0 || funcId >= funcMgtBuiltinsNum) {
     return TSDB_CODE_FAILED;
   }
@@ -128,7 +128,7 @@ EFuncDataRequired fmFuncDynDataRequired(int32_t funcId, void* pRes, STimeWindow*
   if (funcMgtBuiltins[funcId].dynDataRequiredFunc == NULL) {
     return FUNC_DATA_REQUIRED_DATA_LOAD;
   } else {
-    return funcMgtBuiltins[funcId].dynDataRequiredFunc(pRes, pTimeWindow);
+    return funcMgtBuiltins[funcId].dynDataRequiredFunc(pRes, pBlockInfo);
   }
 }
 

From 750ea5789f4246be2874d0ac305d7f65a0ed1af0 Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Mon, 25 Mar 2024 14:20:46 +0800
Subject: [PATCH 09/32] fix(tsdb): fix errors identified by CI.

---
 source/dnode/vnode/src/tsdb/tsdbRead2.c | 69 +++++++++++++++++--------
 source/dnode/vnode/src/tsdb/tsdbUtil.c  |  2 +-
 2 files changed, 48 insertions(+), 23 deletions(-)

diff --git a/source/dnode/vnode/src/tsdb/tsdbRead2.c b/source/dnode/vnode/src/tsdb/tsdbRead2.c
index 6e134698b5..891fd823e7 100644
--- a/source/dnode/vnode/src/tsdb/tsdbRead2.c
+++ b/source/dnode/vnode/src/tsdb/tsdbRead2.c
@@ -141,7 +141,7 @@ static int32_t tGetPrimaryKeyIndex(uint8_t *p, SPrimaryKeyIndex *index) {
 }
 
 static void tRowGetKeyDeepCopy(SRow* pRow, SRowKey* pKey) {
-  pKey->ts = pKey->ts;
+  pKey->ts = pRow->ts;
   pKey->numOfPKs = pRow->numOfPKs;
   if (pKey->numOfPKs == 0) {
     return;
@@ -164,7 +164,7 @@ static void tRowGetKeyDeepCopy(SRow* pRow, SRowKey* pKey) {
       pKey->pks[i].pData = memcpy(pKey->pks[i].pData, data + indices[i].offset, pKey->pks[i].nData);
       pKey->pks[i].pData += pKey->pks[i].nData;
     } else {
-      pKey->pks[i].val = *(int64_t*) data + indices[i].offset;
+      pKey->pks[i].val = *(int64_t*) (data + indices[i].offset);
     }
   }
 }
@@ -694,7 +694,7 @@ static int32_t doLoadFileBlock(STsdbReader* pReader, SArray* pIndexList, SBlockN
       w.ekey = pScanInfo->lastProcKey.ts;
     }
 
-    if (/*isEmptyQueryTimeWindow(&w)*/ w.ekey - w.skey < 2) {  // NOTE: specialized for open interval
+    if (/*isEmptyQueryTimeWindow(&w)*/ w.ekey - w.skey < 1) {  // NOTE: specialized for open interval
       k += 1;
 
       if (k >= numOfTables) {
@@ -2012,41 +2012,58 @@ int32_t doInitMemDataIter(STsdbReader* pReader, STbData** pData, STableBlockScan
   return code;
 }
 
+static void doForwardDataIter(SRowKey* pKey, SIterInfo* pIter, STableBlockScanInfo* pBlockScanInfo, STsdbReader* pReader) {
+  SRowKey rowKey;
+
+  while (1) {
+    TSDBROW* pRow = getValidMemRow(pIter, pBlockScanInfo->delSkyline, pReader);
+    if (!pIter->hasVal) {
+      break;
+    }
+
+    tRowGetKeyEx(pRow, &rowKey);
+    int32_t ret = pkCompEx(pReader->pkComparFn, pKey, &rowKey);
+    if (ret == 0) {
+      pIter->hasVal = tsdbTbDataIterNext(pIter->iter);
+    } else {
+      break;
+    }
+  }
+}
+
+// handle the open interval issue. Find the first row key that is greater than the given one.
+static int32_t forwardDataIter(SRowKey* pKey, STableBlockScanInfo* pBlockScanInfo, STsdbReader* pReader) {
+  doForwardDataIter(pKey, &pBlockScanInfo->iter, pBlockScanInfo, pReader);
+  doForwardDataIter(pKey, &pBlockScanInfo->iiter, pBlockScanInfo, pReader);
+  return TSDB_CODE_SUCCESS;
+}
+
 static int32_t initMemDataIterator(STableBlockScanInfo* pBlockScanInfo, STsdbReader* pReader) {
+  STbData*       d = NULL;
+  STbData*       di = NULL;
+  bool           asc = ASCENDING_TRAVERSE(pReader->info.order);
+  STsdbReadSnap* pSnap = pReader->pReadSnap;
+
   if (pBlockScanInfo->iterInit) {
     return TSDB_CODE_SUCCESS;
   }
 
-  STbData*    d = NULL;
   STsdbRowKey startKey = {0};
-  if (ASCENDING_TRAVERSE(pReader->info.order)) {
-    startKey = (STsdbRowKey){.version = pReader->info.verRange.minVer,
-                             .key = {
-                                 .ts = pBlockScanInfo->lastProcKey.ts + 1,
-                                 .numOfPKs = pReader->suppInfo.numOfPks,
-                             }};
-  } else {
-    startKey = (STsdbRowKey){.version = pReader->info.verRange.maxVer,
-                             .key = {
-                                 .ts = pBlockScanInfo->lastProcKey.ts - 1,
-                                 .numOfPKs = pReader->suppInfo.numOfPks,
-                             }};
-  }
+  tRowKeyAssign(&startKey.key, &pBlockScanInfo->lastProcKey);
+  startKey.version = asc ? pReader->info.verRange.minVer : pReader->info.verRange.maxVer;
 
-  int32_t code =
-      doInitMemDataIter(pReader, &d, pBlockScanInfo, &startKey, pReader->pReadSnap->pMem, &pBlockScanInfo->iter, "mem");
+  int32_t code = doInitMemDataIter(pReader, &d, pBlockScanInfo, &startKey, pSnap->pMem, &pBlockScanInfo->iter, "mem");
   if (code != TSDB_CODE_SUCCESS) {
     return code;
   }
 
-  STbData* di = NULL;
-  code = doInitMemDataIter(pReader, &di, pBlockScanInfo, &startKey, pReader->pReadSnap->pIMem, &pBlockScanInfo->iiter,
-                           "imem");
+  code = doInitMemDataIter(pReader, &di, pBlockScanInfo, &startKey, pSnap->pIMem, &pBlockScanInfo->iiter, "imem");
   if (code != TSDB_CODE_SUCCESS) {
     return code;
   }
 
   loadMemTombData(&pBlockScanInfo->pMemDelData, d, di, pReader->info.verRange.maxVer);
+  forwardDataIter(&startKey.key, pBlockScanInfo, pReader);
 
   pBlockScanInfo->iterInit = true;
   return TSDB_CODE_SUCCESS;
@@ -4128,6 +4145,9 @@ int32_t tsdbReaderOpen2(void* pVnode, SQueryTableDataCond* pCond, void* pTableLi
     blockDataEnsureCapacity(pResBlock, capacity);
   }
 
+  // for debug purpose
+//  capacity = 7;
+
   int32_t code = tsdbReaderCreate(pVnode, pCond, ppReader, capacity, pResBlock, idstr);
   if (code != TSDB_CODE_SUCCESS) {
     goto _err;
@@ -4894,6 +4914,10 @@ SSDataBlock* tsdbRetrieveDataBlock2(STsdbReader* pReader, SArray* pIdList) {
 
   SReaderStatus* pStatus = &pTReader->status;
   if (pStatus->composedDataBlock || pReader->info.execMode == READER_EXEC_ROWS) {
+
+//    tsdbReaderSuspend2(pReader);
+//    tsdbReaderResume2(pReader);
+
     return pTReader->resBlockInfo.pResBlock;
   }
 
@@ -4904,6 +4928,7 @@ SSDataBlock* tsdbRetrieveDataBlock2(STsdbReader* pReader, SArray* pIdList) {
 
 //  tsdbReaderSuspend2(pReader);
 //  tsdbReaderResume2(pReader);
+
   return ret;
 }
 
diff --git a/source/dnode/vnode/src/tsdb/tsdbUtil.c b/source/dnode/vnode/src/tsdb/tsdbUtil.c
index fe7a3457bd..59452ebb9d 100644
--- a/source/dnode/vnode/src/tsdb/tsdbUtil.c
+++ b/source/dnode/vnode/src/tsdb/tsdbUtil.c
@@ -633,7 +633,7 @@ int32_t tRowKeyAssign(SRowKey *pDst, SRowKey* pSrc) {
   pDst->numOfPKs = pSrc->numOfPKs;
 
   if (pSrc->numOfPKs > 0) {
-    for (int32_t i = 0; i < pDst->numOfPKs; ++i) {
+    for (int32_t i = 0; i < pSrc->numOfPKs; ++i) {
       SValue *pVal = &pDst->pks[i];
       pVal->type = pSrc->pks[i].type;
 

From a5ba546bd4d54a06913fdfb354e890a0e7717ca0 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Mon, 25 Mar 2024 15:10:25 +0800
Subject: [PATCH 10/32] fix: add pk to funcInputGetNextRow

---
 include/libs/function/function.h        |  5 ++-
 source/libs/function/src/builtinsimpl.c | 53 +++++++++++++++----------
 2 files changed, 37 insertions(+), 21 deletions(-)

diff --git a/include/libs/function/function.h b/include/libs/function/function.h
index 4ee0989118..78bd1d807b 100644
--- a/include/libs/function/function.h
+++ b/include/libs/function/function.h
@@ -185,6 +185,7 @@ typedef struct SFuncInputRow {
   TSKEY ts;
   bool isDataNull;
   char* pData;
+  char* pPk;
 
   SSDataBlock* block; // prev row block or src block
   int32_t rowIndex; // prev row block ? 0 : rowIndex in srcBlock
@@ -198,7 +199,8 @@ typedef struct SFuncInputRowIter {
   bool  hasPrev;
  
   SInputColumnInfoData* pInput;
-  SColumnInfoData* pData;
+  SColumnInfoData* pDataCol;
+  SColumnInfoData* pPkCol;
   TSKEY* tsList;
   int32_t rowIndex;
   int32_t inputEndIndex;
@@ -207,6 +209,7 @@ typedef struct SFuncInputRowIter {
   TSKEY prevBlockTsEnd;
   bool prevIsDataNull;
   char* pPrevData;
+  char* pPrevPk;
   SSDataBlock* pPrevRowBlock; // pre one row block
 
   //TODO:
diff --git a/source/libs/function/src/builtinsimpl.c b/source/libs/function/src/builtinsimpl.c
index 83f9dcff08..2907c096fe 100644
--- a/source/libs/function/src/builtinsimpl.c
+++ b/source/libs/function/src/builtinsimpl.c
@@ -424,7 +424,8 @@ int32_t funcInputUpdate(SqlFunctionCtx* pCtx) {
   if (!pCtx->bInputFinished) {
     pIter->pInput = &pCtx->input;
     pIter->tsList = (TSKEY*)pIter->pInput->pPTS->pData;
-    pIter->pData = pIter->pInput->pData[0];
+    pIter->pDataCol = pIter->pInput->pData[0];
+    pIter->pPkCol = pIter->pInput->pPrimaryKey;
     pIter->rowIndex = pIter->pInput->startRowIndex;
     pIter->inputEndIndex = pIter->rowIndex + pIter->pInput->numOfRows - 1;
     pIter->pSrcBlock = pCtx->pSrcBlock; 
@@ -454,10 +455,16 @@ bool funcInputGetNextRowDescPk(SFuncInputRowIter* pIter, SFuncInputRow* pRow) {
     if (pIter->prevBlockTsEnd == pIter->tsList[pIter->inputEndIndex]) {
       blockDataDestroy(pIter->pPrevRowBlock);
       pIter->pPrevRowBlock = blockDataExtractBlock(pIter->pSrcBlock, pIter->inputEndIndex, 1);
-      pIter->prevIsDataNull = colDataIsNull_f(pIter->pData->nullbitmap, pIter->inputEndIndex);
-      pIter->pPrevData = taosMemoryMalloc(pIter->pData->info.bytes);
-      char* srcData = colDataGetData(pIter->pData, pIter->inputEndIndex);
-      memcpy(pIter->pPrevData, srcData, pIter->pData->info.bytes);
+      pIter->prevIsDataNull = colDataIsNull_f(pIter->pDataCol->nullbitmap, pIter->inputEndIndex);
+
+      pIter->pPrevData = taosMemoryMalloc(pIter->pDataCol->info.bytes);
+      char* srcData = colDataGetData(pIter->pDataCol, pIter->inputEndIndex);
+      memcpy(pIter->pPrevData, srcData, pIter->pDataCol->info.bytes);
+
+      pIter->pPrevPk = taosMemoryMalloc(pIter->pPkCol->info.bytes);
+      char* pkData = colDataGetData(pIter->pPkCol, pIter->inputEndIndex);
+      memcpy(pIter->pPrevPk, pkData, pIter->pPkCol->info.bytes);
+
       pIter->pPrevRowBlock = blockDataExtractBlock(pIter->pSrcBlock, pIter->inputEndIndex, 1);
 
       pIter->hasPrev = true;
@@ -475,8 +482,9 @@ bool funcInputGetNextRowDescPk(SFuncInputRowIter* pIter, SFuncInputRow* pRow) {
         pRow->rowIndex = 0;
       } else {
         pRow->ts = pIter->tsList[idx - 1];
-        pRow->isDataNull = colDataIsNull_f(pIter->pData->nullbitmap, idx - 1);
-        pRow->pData = colDataGetData(pIter->pData, idx - 1);
+        pRow->isDataNull = colDataIsNull_f(pIter->pDataCol->nullbitmap, idx - 1);
+        pRow->pData = colDataGetData(pIter->pDataCol, idx - 1);
+        pRow->pPk = colDataGetData(pIter->pPkCol, idx - 1);
         pRow->block = pIter->pSrcBlock;
         pRow->rowIndex = idx - 1;
       }
@@ -492,20 +500,22 @@ bool funcInputGetNextRowDescPk(SFuncInputRowIter* pIter, SFuncInputRow* pRow) {
         ++idx;
       }
       pRow->ts = pIter->tsList[idx];
-      pRow->isDataNull = colDataIsNull_f(pIter->pData->nullbitmap, idx);
-      pRow->pData = colDataGetData(pIter->pData, pIter->rowIndex);
+      pRow->isDataNull = colDataIsNull_f(pIter->pDataCol->nullbitmap, idx);
+      pRow->pData = colDataGetData(pIter->pDataCol, idx);
+      pRow->pPk = colDataGetData(pIter->pPkCol, idx);
       pRow->block = pIter->pSrcBlock;
-      pRow->rowIndex = idx;
 
       pIter->rowIndex = idx + 1;
       return true;
     } else {
       pIter->hasPrev = true;
       pIter->prevBlockTsEnd = tsEnd;
-      // TODO
-      pIter->prevIsDataNull = colDataIsNull_f(pIter->pData->nullbitmap, pIter->inputEndIndex);
-      pIter->pPrevData = taosMemoryMalloc(pIter->pData->info.bytes);
-      memcpy(pIter->pPrevData, colDataGetData(pIter->pData, pIter->inputEndIndex), pIter->pData->info.bytes);
+      pIter->prevIsDataNull = colDataIsNull_f(pIter->pDataCol->nullbitmap, pIter->inputEndIndex);
+      pIter->pPrevData = taosMemoryMalloc(pIter->pDataCol->info.bytes);
+      memcpy(pIter->pPrevData, colDataGetData(pIter->pDataCol, pIter->inputEndIndex), pIter->pDataCol->info.bytes);
+      pIter->pPrevPk = taosMemoryMalloc(pIter->pPkCol->info.bytes);
+      memcpy(pIter->pPrevPk, colDataGetData(pIter->pPkCol, pIter->inputEndIndex), pIter->pPkCol->info.bytes);
+
       pIter->pPrevRowBlock = blockDataExtractBlock(pIter->pSrcBlock, pIter->inputEndIndex, 1);
       return false;
     }
@@ -523,8 +533,9 @@ bool funcInputGetNextRowAscPk(SFuncInputRowIter *pIter, SFuncInputRow* pRow) {
         ++idx;
       }
       pRow->ts = pIter->tsList[idx];
-      pRow->isDataNull = colDataIsNull_f(pIter->pData->nullbitmap, idx);
-      pRow->pData = colDataGetData(pIter->pData, idx);
+      pRow->isDataNull = colDataIsNull_f(pIter->pDataCol->nullbitmap, idx);
+      pRow->pData = colDataGetData(pIter->pDataCol, idx);
+      pRow->pPk = colDataGetData(pIter->pPkCol, idx);
       pRow->block = pIter->pSrcBlock;
       pRow->rowIndex = idx;
 
@@ -535,8 +546,9 @@ bool funcInputGetNextRowAscPk(SFuncInputRowIter *pIter, SFuncInputRow* pRow) {
   } else {
     if (pIter->rowIndex <= pIter->inputEndIndex) { 
       pRow->ts = pIter->tsList[pIter->rowIndex];
-      pRow->isDataNull = colDataIsNull_f(pIter->pData->nullbitmap, pIter->rowIndex);
-      pRow->pData = colDataGetData(pIter->pData, pIter->rowIndex);
+      pRow->isDataNull = colDataIsNull_f(pIter->pDataCol->nullbitmap, pIter->rowIndex);
+      pRow->pData = colDataGetData(pIter->pDataCol, pIter->rowIndex);
+      pRow->pPk = colDataGetData(pIter->pPkCol, pIter->rowIndex);
       pRow->block = pIter->pSrcBlock;
       pRow->rowIndex = pIter->rowIndex;
 
@@ -563,8 +575,9 @@ bool funcInputGetNextRowAscPk(SFuncInputRowIter *pIter, SFuncInputRow* pRow) {
 bool funcInputGetNextRowNoPk(SFuncInputRowIter *pIter, SFuncInputRow* pRow) {
   if (pIter->rowIndex <= pIter->inputEndIndex) {
     pRow->ts = pIter->tsList[pIter->rowIndex];
-    pRow->isDataNull = colDataIsNull_f(pIter->pData->nullbitmap, pIter->rowIndex);
-    pRow->pData = colDataGetData(pIter->pData, pIter->rowIndex);
+    pRow->isDataNull = colDataIsNull_f(pIter->pDataCol->nullbitmap, pIter->rowIndex);
+    pRow->pData = colDataGetData(pIter->pDataCol, pIter->rowIndex);
+    pRow->pPk = NULL;
     pRow->block = pIter->pSrcBlock;
     pRow->rowIndex = pIter->rowIndex;
 

From 5c729cc836f2d9eac23bceb83c9b26c43dd541a4 Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Mon, 25 Mar 2024 17:57:59 +0800
Subject: [PATCH 11/32] refactor: do some internal refactor.

---
 source/dnode/vnode/src/inc/tsdb.h           |  6 +-
 source/dnode/vnode/src/tsdb/tsdbMergeTree.c | 37 ++++++++++--
 source/dnode/vnode/src/tsdb/tsdbRead2.c     | 62 ++++++++++++++-------
 source/dnode/vnode/src/tsdb/tsdbReadUtil.c  | 47 +++++++++++-----
 source/dnode/vnode/src/tsdb/tsdbReadUtil.h  |  5 +-
 5 files changed, 117 insertions(+), 40 deletions(-)

diff --git a/source/dnode/vnode/src/inc/tsdb.h b/source/dnode/vnode/src/inc/tsdb.h
index 2b806e95a5..761e74b9cf 100644
--- a/source/dnode/vnode/src/inc/tsdb.h
+++ b/source/dnode/vnode/src/inc/tsdb.h
@@ -836,6 +836,8 @@ struct SLDataIter {
   STimeWindow            timeWindow;
   SVersionRange          verRange;
   SSttBlockLoadInfo     *pBlockLoadInfo;
+  SRowKey                startRowKey;        // current row key
+  __compar_fn_t          comparFn;
   bool                   ignoreEarlierTs;
   struct SSttFileReader *pReader;
 };
@@ -846,7 +848,7 @@ struct SSttFileReader;
 typedef int32_t (*_load_tomb_fn)(STsdbReader *pReader, struct SSttFileReader *pSttFileReader,
                                  SSttBlockLoadInfo *pLoadInfo);
 
-typedef struct {
+typedef struct SMergeTreeConf {
   int8_t        backward;
   STsdb        *pTsdb;
   uint64_t      suid;
@@ -859,7 +861,9 @@ typedef struct {
   STSchema     *pSchema;
   int16_t      *pCols;
   int32_t       numOfCols;
+  SRowKey      *pCurRowKey;
   _load_tomb_fn loadTombFn;
+  __compar_fn_t comparFn;
   void         *pReader;
   void         *idstr;
   bool          rspRows;  // response the rows in stt-file, if possible
diff --git a/source/dnode/vnode/src/tsdb/tsdbMergeTree.c b/source/dnode/vnode/src/tsdb/tsdbMergeTree.c
index 61c610538e..97194acb00 100644
--- a/source/dnode/vnode/src/tsdb/tsdbMergeTree.c
+++ b/source/dnode/vnode/src/tsdb/tsdbMergeTree.c
@@ -479,6 +479,9 @@ int32_t tLDataIterOpen2(SLDataIter *pIter, SSttFileReader *pSttFileReader, int32
   pIter->verRange.maxVer = pConf->verRange.maxVer;
   pIter->timeWindow.skey = pConf->timewindow.skey;
   pIter->timeWindow.ekey = pConf->timewindow.ekey;
+  pIter->comparFn = pConf->comparFn;
+
+  tRowKeyAssign(&pIter->startRowKey, pConf->pCurRowKey);
   pIter->pReader = pSttFileReader;
   pIter->pBlockLoadInfo = pBlockLoadInfo;
 
@@ -618,17 +621,39 @@ static void findNextValidRow(SLDataIter *pIter, const char *idStr) {
     }
 
     int64_t ts = pData->aTSKEY[i];
-    if (!pIter->backward) {               // asc
+    if (!pIter->backward) {              // asc
       if (ts > pIter->timeWindow.ekey) {  // no more data
         break;
-      } else if (ts < pIter->timeWindow.skey) {
-        continue;
+      } else {
+        if (ts < pIter->timeWindow.skey) {
+          continue;
+        }
+
+        if (ts == pIter->timeWindow.skey && pIter->startRowKey.numOfPKs > 0) {
+          SRowKey key;
+          tColRowGetKey(pData, i, &key);
+          int32_t ret = pkCompEx(pIter->comparFn, &key, &pIter->startRowKey);
+          if (ret < 0) {
+            continue;
+          }
+        }
       }
     } else {
       if (ts < pIter->timeWindow.skey) {
         break;
-      } else if (ts > pIter->timeWindow.ekey) {
-        continue;
+      } else {
+        if (ts > pIter->timeWindow.ekey) {
+          continue;
+        }
+
+        if (ts == pIter->timeWindow.ekey && pIter->startRowKey.numOfPKs > 0) {
+          SRowKey key;
+          tColRowGetKey(pData, i, &key);
+          int32_t ret = pkCompEx(pIter->comparFn, &key, &pIter->startRowKey);
+          if (ret > 0) {
+            continue;
+          }
+        }
       }
     }
 
@@ -802,8 +827,8 @@ int32_t tMergeTreeOpen2(SMergeTree *pMTree, SMergeTreeConf *pConf, SSttDataInfoF
 
       STimeWindow w = {0};
       int64_t     numOfRows = 0;
+      int64_t     cid = pSttLevel->fobjArr->data[i]->f->cid;
 
-      int64_t cid = pSttLevel->fobjArr->data[i]->f->cid;
       code = tLDataIterOpen2(pIter, pSttFileReader, cid, pMTree->backward, pConf, pLoadInfo, &w, &numOfRows,
                              pMTree->idStr);
       if (code != TSDB_CODE_SUCCESS) {
diff --git a/source/dnode/vnode/src/tsdb/tsdbRead2.c b/source/dnode/vnode/src/tsdb/tsdbRead2.c
index 891fd823e7..80c4514579 100644
--- a/source/dnode/vnode/src/tsdb/tsdbRead2.c
+++ b/source/dnode/vnode/src/tsdb/tsdbRead2.c
@@ -89,7 +89,7 @@ static bool outOfTimeWindow(int64_t ts, STimeWindow* pWindow) { return (ts > pWi
 
 static void resetPreFilesetMemTableListIndex(SReaderStatus* pStatus);
 
-static int32_t pkCompEx(__compar_fn_t comparFn, SRowKey* p1, SRowKey* p2) {
+int32_t pkCompEx(__compar_fn_t comparFn, SRowKey* p1, SRowKey* p2) {
   if (p2 == NULL) {
     return 1;
   }
@@ -118,12 +118,13 @@ static void tColRowGetKeyDeepCopy(SBlockData* pBlock, int32_t irow, int32_t slot
     return;
   }
 
-  pKey->numOfPKs = 1;
-
   SColData* pColData = &pBlock->aColData[slotId];
   SColVal   cv;
   tColDataGetValue(pColData, irow, &cv);
 
+  pKey->numOfPKs = 1;
+  pKey->pks[0].type = cv.value.type;
+
   if (IS_NUMERIC_TYPE(cv.value.type)) {
     pKey->pks[0].val = cv.value.val;
   } else {
@@ -1394,7 +1395,7 @@ static int64_t getBoarderKeyInFiles(SFileDataBlockInfo* pBlock, STableBlockScanI
 
   int64_t key = 0;
   if (pScanInfo->sttKeyInfo.status == STT_FILE_HAS_DATA) {
-    int64_t keyInStt = pScanInfo->sttKeyInfo.nextProcKey;
+    int64_t keyInStt = pScanInfo->sttKeyInfo.nextProcKey.ts;
     key = ascScan ? TMIN(pBlock->firstKey, keyInStt) : TMAX(pBlock->lastKey, keyInStt);
   } else {
     key = ascScan ? pBlock->firstKey : pBlock->lastKey;
@@ -1437,9 +1438,10 @@ static void getBlockToLoadInfo(SDataBlockToLoadInfo* pInfo, SFileDataBlockInfo*
   pInfo->hasDupTs = (pBlockInfo->numRow > pBlockInfo->count) || (pBlockInfo->count <= 0);
   pInfo->overlapWithDelInfo = overlapWithDelSkyline(pScanInfo, &pRecord, pReader->info.order);
 
+  // todo handle the primary key overlap case
   ASSERT(pScanInfo->sttKeyInfo.status != STT_FILE_READER_UNINIT);
   if (pScanInfo->sttKeyInfo.status == STT_FILE_HAS_DATA) {
-    int64_t nextProcKeyInStt = pScanInfo->sttKeyInfo.nextProcKey;
+    int64_t nextProcKeyInStt = pScanInfo->sttKeyInfo.nextProcKey.ts;
     pInfo->overlapWithSttBlock = !(pBlockInfo->lastKey < nextProcKeyInStt || pBlockInfo->firstKey > nextProcKeyInStt);
   }
 
@@ -1536,8 +1538,9 @@ static bool tryCopyDistinctRowFromFileBlock(STsdbReader* pReader, SBlockData* pB
 
 static bool nextRowFromSttBlocks(SSttBlockReader* pSttBlockReader, STableBlockScanInfo* pScanInfo, int32_t pkSrcSlot,
                                  SVersionRange* pVerRange) {
-  int32_t order = pSttBlockReader->order;
-  int32_t step = ASCENDING_TRAVERSE(order) ? 1 : -1;
+  int32_t  order = pSttBlockReader->order;
+  int32_t  step = ASCENDING_TRAVERSE(order) ? 1 : -1;
+  SRowKey* pNextProc = &pScanInfo->sttKeyInfo.nextProcKey;
 
   while (1) {
     bool hasVal = tMergeTreeNext(&pSttBlockReader->mergeTree);
@@ -1545,7 +1548,14 @@ static bool nextRowFromSttBlocks(SSttBlockReader* pSttBlockReader, STableBlockSc
       pScanInfo->sttKeyInfo.status = STT_FILE_NO_DATA;
 
       // next file, the timestamps in the next file must be greater than those in current
-      pScanInfo->sttKeyInfo.nextProcKey += step;
+      pNextProc->ts += step;
+      if (pSttBlockReader->numOfPks > 0) {
+        if (IS_NUMERIC_TYPE(pNextProc->pks[0].type)) {
+          pNextProc->pks[0].val = INT64_MIN;
+        } else {
+          memset(pNextProc->pks[0].pData, 0, pNextProc->pks[0].nData);
+        }
+      }
       return false;
     }
 
@@ -1559,7 +1569,8 @@ static bool nextRowFromSttBlocks(SSttBlockReader* pSttBlockReader, STableBlockSc
       tColRowGetKeyDeepCopy(pRow->pBlockData, pRow->iRow, pkSrcSlot, &pSttBlockReader->currentKey);
     }
 
-    pScanInfo->sttKeyInfo.nextProcKey = key;
+    tColRowGetKeyDeepCopy(pRow->pBlockData, pRow->iRow, pkSrcSlot, pNextProc);
+
     if (pScanInfo->delSkyline != NULL && TARRAY_SIZE(pScanInfo->delSkyline) > 0) {
       if (!hasBeenDropped(pScanInfo->delSkyline, &pScanInfo->sttBlockDelIndex, key, ver, order, pVerRange)) {
         pScanInfo->sttKeyInfo.status = STT_FILE_HAS_DATA;
@@ -2042,7 +2053,9 @@ static int32_t initMemDataIterator(STableBlockScanInfo* pBlockScanInfo, STsdbRea
   STbData*       d = NULL;
   STbData*       di = NULL;
   bool           asc = ASCENDING_TRAVERSE(pReader->info.order);
+  bool           forward = true;
   STsdbReadSnap* pSnap = pReader->pReadSnap;
+  STimeWindow*   pWindow = &pReader->info.window;
 
   if (pBlockScanInfo->iterInit) {
     return TSDB_CODE_SUCCESS;
@@ -2051,6 +2064,10 @@ static int32_t initMemDataIterator(STableBlockScanInfo* pBlockScanInfo, STsdbRea
   STsdbRowKey startKey = {0};
   tRowKeyAssign(&startKey.key, &pBlockScanInfo->lastProcKey);
   startKey.version = asc ? pReader->info.verRange.minVer : pReader->info.verRange.maxVer;
+  if ((asc && (startKey.key.ts < pWindow->skey)) || ((!asc) && startKey.key.ts > pWindow->ekey)) {
+    startKey.key.ts = asc? pWindow->skey:pWindow->ekey;
+    forward = false;
+  }
 
   int32_t code = doInitMemDataIter(pReader, &d, pBlockScanInfo, &startKey, pSnap->pMem, &pBlockScanInfo->iter, "mem");
   if (code != TSDB_CODE_SUCCESS) {
@@ -2063,7 +2080,10 @@ static int32_t initMemDataIterator(STableBlockScanInfo* pBlockScanInfo, STsdbRea
   }
 
   loadMemTombData(&pBlockScanInfo->pMemDelData, d, di, pReader->info.verRange.maxVer);
-  forwardDataIter(&startKey.key, pBlockScanInfo, pReader);
+
+  if (forward) {
+    forwardDataIter(&startKey.key, pBlockScanInfo, pReader);
+  }
 
   pBlockScanInfo->iterInit = true;
   return TSDB_CODE_SUCCESS;
@@ -2115,6 +2135,7 @@ static bool isValidFileBlockRow(SBlockData* pBlockData, int32_t rowIndex, STable
 
 static bool initSttBlockReader(SSttBlockReader* pSttBlockReader, STableBlockScanInfo* pScanInfo, STsdbReader* pReader) {
   bool hasData = true;
+  bool asc = ASCENDING_TRAVERSE(pReader->info.order);
 
   // the stt block reader has been initialized for this table.
   if (pSttBlockReader->uid == pScanInfo->uid) {
@@ -2133,10 +2154,10 @@ static bool initSttBlockReader(SSttBlockReader* pSttBlockReader, STableBlockScan
   }
 
   STimeWindow w = pSttBlockReader->window;
-  if (ASCENDING_TRAVERSE(pSttBlockReader->order)) {
-    w.skey = pScanInfo->sttKeyInfo.nextProcKey;
+  if (asc) {
+    w.skey = pScanInfo->sttKeyInfo.nextProcKey.ts;
   } else {
-    w.ekey = pScanInfo->sttKeyInfo.nextProcKey;
+    w.ekey = pScanInfo->sttKeyInfo.nextProcKey.ts;
   }
 
   int64_t st = taosGetTimestampUs();
@@ -2157,6 +2178,8 @@ static bool initSttBlockReader(SSttBlockReader* pSttBlockReader, STableBlockScan
       .pCols = pReader->suppInfo.colId,
       .numOfCols = pReader->suppInfo.numOfCols,
       .loadTombFn = loadSttTombDataForAll,
+      .pCurRowKey = &pScanInfo->sttKeyInfo.nextProcKey,
+      .comparFn = pReader->pkComparFn,
       .pReader = pReader,
       .idstr = pReader->idStr,
       .rspRows = (pReader->info.execMode == READER_EXEC_ROWS),
@@ -2193,8 +2216,9 @@ static bool initSttBlockReader(SSttBlockReader* pSttBlockReader, STableBlockScan
       }
 
       pScanInfo->sttKeyInfo.status = taosArrayGetSize(info.pTimeWindowList) ? STT_FILE_HAS_DATA : STT_FILE_NO_DATA;
-      pScanInfo->sttKeyInfo.nextProcKey =
-          ASCENDING_TRAVERSE(pReader->info.order) ? pScanInfo->sttWindow.skey : pScanInfo->sttWindow.ekey;
+
+      // todo set the primary key value
+      pScanInfo->sttKeyInfo.nextProcKey.ts = asc ? pScanInfo->sttWindow.skey : pScanInfo->sttWindow.ekey;
       hasData = (pScanInfo->sttKeyInfo.status == STT_FILE_HAS_DATA);
     } else {  // not clean stt blocks
       INIT_TIMEWINDOW(&pScanInfo->sttWindow);  //reset the time window
@@ -2755,7 +2779,7 @@ static void buildCleanBlockFromSttFiles(STsdbReader* pReader, STableBlockScanInf
 
   setComposedBlockFlag(pReader, true);
 
-  pScanInfo->sttKeyInfo.nextProcKey = asc ? pScanInfo->sttWindow.ekey + 1 : pScanInfo->sttWindow.skey - 1;
+  pScanInfo->sttKeyInfo.nextProcKey.ts = asc ? pScanInfo->sttWindow.ekey + 1 : pScanInfo->sttWindow.skey - 1;
   pScanInfo->sttKeyInfo.status = STT_FILE_NO_DATA;
 
   pScanInfo->lastProcKey.ts = asc ? pScanInfo->sttWindow.ekey : pScanInfo->sttWindow.skey;
@@ -2915,7 +2939,7 @@ static bool notOverlapWithFiles(SFileDataBlockInfo* pBlockInfo, STableBlockScanI
   if ((!hasDataInSttBlock(pScanInfo)) || (pScanInfo->cleanSttBlocks == true)) {
     return true;
   } else {
-    int64_t keyInStt = pScanInfo->sttKeyInfo.nextProcKey;
+    int64_t keyInStt = pScanInfo->sttKeyInfo.nextProcKey.ts;
     return (asc && pBlockInfo->lastKey < keyInStt) || (!asc && pBlockInfo->firstKey > keyInStt);
   }
 }
@@ -2963,7 +2987,7 @@ static int32_t doBuildDataBlock(STsdbReader* pReader) {
     code = buildDataBlockFromBuf(pReader, pScanInfo, endKey);
   } else {
     if (notOverlapWithFiles(pBlockInfo, pScanInfo, asc)) {
-      int64_t keyInStt = pScanInfo->sttKeyInfo.nextProcKey;
+      int64_t keyInStt = pScanInfo->sttKeyInfo.nextProcKey.ts;
 
       if ((!hasDataInSttBlock(pScanInfo)) || (asc && pBlockInfo->lastKey < keyInStt) ||
           (!asc && pBlockInfo->firstKey > keyInStt)) {
@@ -3649,7 +3673,7 @@ int32_t doMergeRowsInSttBlock(SSttBlockReader* pSttBlockReader, STableBlockScanI
     } else {
       tsdbTrace("uid:%" PRIu64 " last del index:%d, del range:%d, lastKeyInStt:%" PRId64 ", %s", pScanInfo->uid,
                 pScanInfo->sttBlockDelIndex, (int32_t)taosArrayGetSize(pScanInfo->delSkyline),
-                pScanInfo->sttKeyInfo.nextProcKey, idStr);
+                pScanInfo->sttKeyInfo.nextProcKey.ts, idStr);
       break;
     }
   }
diff --git a/source/dnode/vnode/src/tsdb/tsdbReadUtil.c b/source/dnode/vnode/src/tsdb/tsdbReadUtil.c
index 93bc5ca573..99f0e42261 100644
--- a/source/dnode/vnode/src/tsdb/tsdbReadUtil.c
+++ b/source/dnode/vnode/src/tsdb/tsdbReadUtil.c
@@ -130,25 +130,46 @@ STableBlockScanInfo* getTableBlockScanInfo(SSHashObj* pTableMap, uint64_t uid, c
   return *p;
 }
 
+static int32_t initSRowKey(SRowKey* pKey, int64_t ts, int32_t numOfPks, int32_t type, int32_t len) {
+  pKey->numOfPKs = numOfPks;
+  pKey->ts = ts;
+
+  if (numOfPks > 0) {
+    pKey->pks[0].type = type;
+    if (IS_NUMERIC_TYPE(pKey->pks[0].type)) {
+      pKey->pks[0].val = INT64_MIN;
+    } else {
+      pKey->pks[0].pData = taosMemoryCalloc(1, len);
+      pKey->pks[0].nData = 0;
+
+      if (pKey->pks[0].pData == NULL) {
+        terrno = TSDB_CODE_OUT_OF_MEMORY;
+        return terrno;
+      }
+    }
+  }
+
+  return TSDB_CODE_SUCCESS;
+}
+
 static void initLastProcKey(STableBlockScanInfo *pScanInfo, STsdbReader* pReader) {
+  int32_t numOfPks = pReader->suppInfo.numOfPks;
+
   SRowKey* pRowKey = &pScanInfo->lastProcKey;
   if (ASCENDING_TRAVERSE(pReader->info.order)) {
     int64_t skey = pReader->info.window.skey;
-    pRowKey->ts = (skey > INT64_MIN) ? (skey - 1) : skey;
-    pScanInfo->sttKeyInfo.nextProcKey = skey;
+    int64_t ts = (skey > INT64_MIN) ? (skey - 1) : skey;
+
+    initSRowKey(pRowKey, ts, numOfPks, pReader->suppInfo.pk.type, pReader->suppInfo.pk.bytes);
+    initSRowKey(&pScanInfo->sttKeyInfo.nextProcKey, skey, numOfPks, pReader->suppInfo.pk.type,
+                pReader->suppInfo.pk.bytes);
   } else {
     int64_t ekey = pReader->info.window.ekey;
-    pRowKey->ts = (ekey < INT64_MAX) ? (ekey + 1) : ekey;
-    pScanInfo->sttKeyInfo.nextProcKey = ekey;
-  }
+    int64_t ts = (ekey < INT64_MAX) ? (ekey + 1) : ekey;
 
-  // only handle the first primary key.
-  pRowKey->numOfPKs = pReader->suppInfo.numOfPks;
-  if (pReader->suppInfo.numOfPks > 0) {
-    if (IS_VAR_DATA_TYPE(pReader->suppInfo.pk.type)) {
-      pRowKey->pks[0].pData = taosMemoryCalloc(1, pReader->suppInfo.pk.bytes);
-    }
-    pRowKey->pks[0].type = pReader->suppInfo.pk.type;
+    initSRowKey(pRowKey, ts, numOfPks, pReader->suppInfo.pk.type, pReader->suppInfo.pk.bytes);
+    initSRowKey(&pScanInfo->sttKeyInfo.nextProcKey, ekey, numOfPks, pReader->suppInfo.pk.type,
+                pReader->suppInfo.pk.bytes);
   }
 }
 
@@ -230,7 +251,7 @@ void resetAllDataBlockScanInfo(SSHashObj* pTableMap, int64_t ts, int32_t step) {
     pInfo->delSkyline = taosArrayDestroy(pInfo->delSkyline);
     pInfo->lastProcKey.ts = ts;
     // todo check the nextProcKey info
-    pInfo->sttKeyInfo.nextProcKey = ts + step;
+    pInfo->sttKeyInfo.nextProcKey.ts = ts + step;
   }
 }
 
diff --git a/source/dnode/vnode/src/tsdb/tsdbReadUtil.h b/source/dnode/vnode/src/tsdb/tsdbReadUtil.h
index 4131bb1b86..9cdb36e648 100644
--- a/source/dnode/vnode/src/tsdb/tsdbReadUtil.h
+++ b/source/dnode/vnode/src/tsdb/tsdbReadUtil.h
@@ -77,7 +77,9 @@ typedef enum ESttKeyStatus {
 
 typedef struct SSttKeyInfo {
   ESttKeyStatus status;           // this value should be updated when switch to the next fileset
-  int64_t       nextProcKey;   // todo remove this attribute, since it is impossible to set correct nextProcKey value
+  SRowKey       nextProcKey;
+  //  int64_t       nextProcKey;   // todo remove this attribute, since it is impossible to set correct nextProcKey
+  //  value
 } SSttKeyInfo;
 
 // clean stt file blocks:
@@ -333,6 +335,7 @@ int32_t tsdbGetRowsInSttFiles(STFileSet* pFileSet, SArray* pSttFileBlockIterArra
                               const char* pstr);
 bool    isCleanSttBlock(SArray* pTimewindowList, STimeWindow* pQueryWindow, STableBlockScanInfo* pScanInfo, int32_t order);
 bool    overlapWithDelSkyline(STableBlockScanInfo* pBlockScanInfo, const SBrinRecord* pRecord, int32_t order);
+int32_t pkCompEx(__compar_fn_t comparFn, SRowKey* p1, SRowKey* p2);
 
 typedef struct {
   SArray* pTombData;

From 9638f8d7fcb345d563f605c7d04cf03acc7f2afb Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Tue, 26 Mar 2024 00:14:46 +0800
Subject: [PATCH 12/32] fix(tsdb): set the correct sttkey

---
 source/dnode/vnode/src/tsdb/tsdbRead2.c | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)

diff --git a/source/dnode/vnode/src/tsdb/tsdbRead2.c b/source/dnode/vnode/src/tsdb/tsdbRead2.c
index 80c4514579..27e7da11df 100644
--- a/source/dnode/vnode/src/tsdb/tsdbRead2.c
+++ b/source/dnode/vnode/src/tsdb/tsdbRead2.c
@@ -1861,9 +1861,11 @@ static int32_t doMergeMultiLevelRows(STsdbReader* pReader, STableBlockScanInfo*
   TSDBROW* pRow = getValidMemRow(&pBlockScanInfo->iter, pDelList, pReader);
   TSDBROW* piRow = getValidMemRow(&pBlockScanInfo->iiter, pDelList, pReader);
 
-  SRowKey* pSttKey = NULL;
+  SRowKey* pSttKey = &(SRowKey){0};
   if (hasDataInSttBlock(pBlockScanInfo) && (!pBlockScanInfo->cleanSttBlocks)) {
-    pSttKey = getCurrentKeyInSttBlock(pSttBlockReader);
+    tRowKeyAssign(pSttKey, getCurrentKeyInSttBlock(pSttBlockReader));
+  } else {
+    pSttKey = NULL;
   }
 
   SRowKey* pfKey = &(SRowKey){0};
@@ -1902,7 +1904,7 @@ static int32_t doMergeMultiLevelRows(STsdbReader* pReader, STableBlockScanInfo*
     }
   }
 
-  SRowKey minKey;
+  SRowKey minKey = {0};
   if (ASCENDING_TRAVERSE(pReader->info.order)) {
     minKey = k;  // let's find the minimum
 
@@ -1910,11 +1912,11 @@ static int32_t doMergeMultiLevelRows(STsdbReader* pReader, STableBlockScanInfo*
       minKey = ik;
     }
 
-    if (pfKey != NULL && (pkCompEx(compFn, pfKey, &minKey) < 0)) {
+    if ((pfKey != NULL) && (pkCompEx(compFn, pfKey, &minKey) < 0)) {
       minKey = *pfKey;
     }
 
-    if (pSttKey != NULL && (pkCompEx(compFn, pSttKey, &minKey) < 0)) {
+    if ((pSttKey != NULL) && (pkCompEx(compFn, pSttKey, &minKey) < 0)) {
       minKey = *pSttKey;
     }
   } else {
@@ -1923,11 +1925,11 @@ static int32_t doMergeMultiLevelRows(STsdbReader* pReader, STableBlockScanInfo*
       minKey = ik;
     }
 
-    if (pfKey != NULL && (pkCompEx(compFn, pfKey, &minKey) > 0)) {
+    if ((pfKey != NULL) && (pkCompEx(compFn, pfKey, &minKey) > 0)) {
       minKey = *pfKey;
     }
 
-    if (pSttKey != NULL && (pkCompEx(compFn, pSttKey, &minKey) > 0)) {
+    if ((pSttKey != NULL) && (pkCompEx(compFn, pSttKey, &minKey) > 0)) {
       minKey = *pSttKey;
     }
   }

From 5fe03fe3f4985ba7b9fdbd5fe48f6ea6a8212b26 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Tue, 26 Mar 2024 09:18:45 +0800
Subject: [PATCH 13/32] feat: irate support pk

---
 source/libs/function/inc/builtinsimpl.h |  2 +-
 source/libs/function/src/builtins.c     |  3 +-
 source/libs/function/src/builtinsimpl.c | 86 ++++++++++++++++++-------
 3 files changed, 64 insertions(+), 27 deletions(-)

diff --git a/source/libs/function/inc/builtinsimpl.h b/source/libs/function/inc/builtinsimpl.h
index 54bcd26bc3..ecc64fcd00 100644
--- a/source/libs/function/inc/builtinsimpl.h
+++ b/source/libs/function/inc/builtinsimpl.h
@@ -144,7 +144,7 @@ int32_t irateFunction(SqlFunctionCtx* pCtx);
 int32_t irateFunctionMerge(SqlFunctionCtx* pCtx);
 int32_t irateFinalize(SqlFunctionCtx* pCtx, SSDataBlock* pBlock);
 int32_t iratePartialFinalize(SqlFunctionCtx* pCtx, SSDataBlock* pBlock);
-int32_t getIrateInfoSize();
+int32_t getIrateInfoSize(int32_t pkBytes);
 
 int32_t cachedLastRowFunction(SqlFunctionCtx* pCtx);
 
diff --git a/source/libs/function/src/builtins.c b/source/libs/function/src/builtins.c
index cb2a979909..0f88572dbc 100644
--- a/source/libs/function/src/builtins.c
+++ b/source/libs/function/src/builtins.c
@@ -1638,7 +1638,8 @@ static int32_t translateIrateImpl(SFunctionNode* pFunc, char* pErrBuf, int32_t l
     if (!IS_NUMERIC_TYPE(colType)) {
       return invaildFuncParaTypeErrMsg(pErrBuf, len, pFunc->functionName);
     }
-    pFunc->node.resType = (SDataType){.bytes = getIrateInfoSize() + VARSTR_HEADER_SIZE, .type = TSDB_DATA_TYPE_BINARY};
+    int32_t pkBytes = (pFunc->hasPk) ? pFunc->pkBytes : 0;
+    pFunc->node.resType = (SDataType){.bytes = getIrateInfoSize(pkBytes) + VARSTR_HEADER_SIZE, .type = TSDB_DATA_TYPE_BINARY};
   } else {
     if (1 != LIST_LENGTH(pFunc->pParameterList)) {
       return invaildFuncParaNumErrMsg(pErrBuf, len, pFunc->functionName);
diff --git a/source/libs/function/src/builtinsimpl.c b/source/libs/function/src/builtinsimpl.c
index 2907c096fe..70fe2b19df 100644
--- a/source/libs/function/src/builtinsimpl.c
+++ b/source/libs/function/src/builtinsimpl.c
@@ -269,6 +269,12 @@ typedef struct SRateInfo {
   double lastValue;
   TSKEY  lastKey;
   int8_t hasResult;  // flag to denote has value
+
+  char* firstPk;
+  char* lastPk;
+  int8_t pkType;
+  int32_t pkBytes;
+  char pkData[];
 } SRateInfo;
 
 typedef struct SGroupKeyInfo {
@@ -6165,10 +6171,11 @@ int32_t derivativeFunction(SqlFunctionCtx* pCtx) {
   return TSDB_CODE_SUCCESS;
 }
 
-int32_t getIrateInfoSize() { return (int32_t)sizeof(SRateInfo); }
+int32_t getIrateInfoSize(int32_t pkBytes) { return (int32_t)sizeof(SRateInfo) + 2 * pkBytes; }
 
 bool getIrateFuncEnv(struct SFunctionNode* pFunc, SFuncExecEnv* pEnv) {
-  pEnv->calcMemSize = sizeof(SRateInfo);
+  int32_t pkBytes = (pFunc->hasPk) ? pFunc->pkBytes : 0;
+  pEnv->calcMemSize = getIrateInfoSize(pkBytes);
   return true;
 }
 
@@ -6188,6 +6195,36 @@ bool irateFuncSetup(SqlFunctionCtx* pCtx, SResultRowEntryInfo* pResInfo) {
   return true;
 }
 
+static void doSaveRateInfo(SRateInfo* pRateInfo, bool isFirst, int64_t ts, char* pk, double v) {
+  if (isFirst) {
+    pRateInfo->firstValue = v;
+    pRateInfo->firstKey = ts;
+    if (pRateInfo->firstPk) {
+      int32_t pkBytes = IS_VAR_DATA_TYPE(pRateInfo->pkType) ? varDataTLen(pk) : pRateInfo->pkBytes;
+      memcpy(pRateInfo->firstPk, pk, pkBytes);
+    }
+  } else {
+    pRateInfo->lastValue = v;
+    pRateInfo->lastKey = ts;
+    if (pRateInfo->lastPk) {
+      int32_t pkBytes = IS_VAR_DATA_TYPE(pRateInfo->pkType) ? varDataTLen(pk) : pRateInfo->pkBytes;
+      memcpy(pRateInfo->lastPk, pk, pkBytes);
+    }
+  }
+}
+
+static void initializeRateInfo(SqlFunctionCtx* pCtx, SRateInfo* pRateInfo) {
+  if (pCtx->hasPrimaryKey) {
+    pRateInfo->pkType = pCtx->input.pPrimaryKey->info.type;
+    pRateInfo->pkBytes = pCtx->input.pPrimaryKey->info.bytes;
+    pRateInfo->firstPk = pRateInfo->pkData;
+    pRateInfo->lastPk = pRateInfo->pkData + pRateInfo->pkBytes;
+  } else {
+    pRateInfo->firstPk = NULL;
+    pRateInfo->lastPk = NULL;
+  }  
+}
+
 int32_t irateFunction(SqlFunctionCtx* pCtx) {
   SResultRowEntryInfo* pResInfo = GET_RES_INFO(pCtx);
   SRateInfo*           pRateInfo = GET_ROWCELL_INTERBUF(pResInfo);
@@ -6198,6 +6235,8 @@ int32_t irateFunction(SqlFunctionCtx* pCtx) {
   SColumnInfoData* pOutput = (SColumnInfoData*)pCtx->pOutput;
 
   funcInputUpdate(pCtx);
+  
+  initializeRateInfo(pCtx, pRateInfo);
 
   int32_t numOfElems = 0;
   int32_t type = pInputCol->info.type;
@@ -6212,21 +6251,16 @@ int32_t irateFunction(SqlFunctionCtx* pCtx) {
     GET_TYPED_DATA(v, double, type, data);
 
     if (INT64_MIN == pRateInfo->lastKey) {
-      pRateInfo->lastValue = v;
-      pRateInfo->lastKey = row.ts;
+      doSaveRateInfo(pRateInfo, false, row.ts, row.pPk, v);
       pRateInfo->hasResult = 1;
       continue;
     }
 
     if (row.ts > pRateInfo->lastKey) {
       if ((INT64_MIN == pRateInfo->firstKey) || pRateInfo->lastKey > pRateInfo->firstKey) {
-        pRateInfo->firstValue = pRateInfo->lastValue;
-        pRateInfo->firstKey = pRateInfo->lastKey;
+        doSaveRateInfo(pRateInfo, true, pRateInfo->lastKey, pRateInfo->lastPk, pRateInfo->lastValue);
       }
-
-      pRateInfo->lastValue = v;
-      pRateInfo->lastKey = row.ts;
-
+      doSaveRateInfo(pRateInfo, false, row.ts, row.pPk, v);
       continue;
     } else if (row.ts == pRateInfo->lastKey) {
         return TSDB_CODE_FUNC_DUP_TIMESTAMP;
@@ -6234,8 +6268,7 @@ int32_t irateFunction(SqlFunctionCtx* pCtx) {
     
 
     if ((INT64_MIN == pRateInfo->firstKey) || row.ts > pRateInfo->firstKey) {
-      pRateInfo->firstValue = v;
-      pRateInfo->firstKey = row.ts;
+      doSaveRateInfo(pRateInfo, true, row.ts, row.pPk, v);    
     } else if (row.ts == pRateInfo->firstKey) {
       return TSDB_CODE_FUNC_DUP_TIMESTAMP;
     }
@@ -6271,25 +6304,26 @@ static double doCalcRate(const SRateInfo* pRateInfo, double tickPerSec) {
 
 static void irateTransferInfoImpl(TSKEY inputKey, SRateInfo* pInput, SRateInfo* pOutput, bool isFirstKey) {
   if (inputKey > pOutput->lastKey) {
-    pOutput->firstKey   = pOutput->lastKey;
-    pOutput->firstValue = pOutput->lastValue;
-
-    pOutput->lastKey    = isFirstKey ? pInput->firstKey : pInput->lastKey;
-    pOutput->lastValue  = isFirstKey ? pInput->firstValue : pInput->lastValue;
+    doSaveRateInfo(pOutput, true, pOutput->lastKey, pOutput->lastPk, pOutput->lastValue);
+    if (isFirstKey) {
+      doSaveRateInfo(pOutput, false, pInput->firstKey, pInput->firstPk, pInput->firstValue);
+    } else {
+      doSaveRateInfo(pOutput, false, pInput->lastKey, pInput->lastPk, pInput->lastValue);
+    }
   } else if ((inputKey < pOutput->lastKey) && (inputKey > pOutput->firstKey)) {
-    pOutput->firstKey   = isFirstKey ? pInput->firstKey : pInput->lastKey;
-    pOutput->firstValue = isFirstKey ? pInput->firstValue : pInput->lastValue;
+    if (isFirstKey) {
+      doSaveRateInfo(pOutput, true, pInput->firstKey, pInput->firstPk, pInput->firstValue);
+    } else {
+      doSaveRateInfo(pOutput, true, pInput->lastKey, pInput->lastPk, pInput->lastValue);
+    }
   } else {
     // inputKey < pOutput->firstKey
   }
 }
 
 static void irateCopyInfo(SRateInfo* pInput, SRateInfo* pOutput) {
-  pOutput->firstKey = pInput->firstKey;
-  pOutput->lastKey  = pInput->lastKey;
-
-  pOutput->firstValue = pInput->firstValue;
-  pOutput->lastValue  = pInput->lastValue;
+  doSaveRateInfo(pOutput, true, pInput->firstKey, pInput->firstPk, pInput->firstValue);
+  doSaveRateInfo(pOutput, false, pInput->lastKey, pInput->lastPk, pInput->lastValue);
 }
 
 static int32_t irateTransferInfo(SRateInfo* pInput, SRateInfo* pOutput) {
@@ -6324,11 +6358,13 @@ int32_t irateFunctionMerge(SqlFunctionCtx* pCtx) {
   }
 
   SRateInfo* pInfo = GET_ROWCELL_INTERBUF(GET_RES_INFO(pCtx));
+  initializeRateInfo(pCtx, pInfo);
 
   int32_t start = pInput->startRowIndex;
   for (int32_t i = start; i < start + pInput->numOfRows; ++i) {
     char*        data = colDataGetData(pCol, i);
     SRateInfo*   pInputInfo = (SRateInfo*)varDataVal(data);
+    initializeRateInfo(pCtx, pInfo);
     if (pInputInfo->hasResult) {
       int32_t code = irateTransferInfo(pInputInfo, pInfo);
       if (code != TSDB_CODE_SUCCESS) {
@@ -6347,7 +6383,7 @@ int32_t irateFunctionMerge(SqlFunctionCtx* pCtx) {
 int32_t iratePartialFinalize(SqlFunctionCtx* pCtx, SSDataBlock* pBlock) {
   SResultRowEntryInfo* pResInfo = GET_RES_INFO(pCtx);
   SRateInfo*           pInfo = GET_ROWCELL_INTERBUF(GET_RES_INFO(pCtx));
-  int32_t              resultBytes = getIrateInfoSize();
+  int32_t              resultBytes = getIrateInfoSize(pInfo->pkBytes);
   char*                res = taosMemoryCalloc(resultBytes + VARSTR_HEADER_SIZE, sizeof(char));
 
   memcpy(varDataVal(res), pInfo, resultBytes);

From a70d5a49476a19e45f3717e5f9188ccf13f75c36 Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Tue, 26 Mar 2024 09:50:06 +0800
Subject: [PATCH 14/32] fix(tsdb): fix bug during filtering brin records.

---
 source/dnode/vnode/src/tsdb/tsdbRead2.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/source/dnode/vnode/src/tsdb/tsdbRead2.c b/source/dnode/vnode/src/tsdb/tsdbRead2.c
index 27e7da11df..7505230eca 100644
--- a/source/dnode/vnode/src/tsdb/tsdbRead2.c
+++ b/source/dnode/vnode/src/tsdb/tsdbRead2.c
@@ -650,7 +650,6 @@ static int32_t doLoadFileBlock(STsdbReader* pReader, SArray* pIndexList, SBlockN
   int32_t      k = 0;
   int32_t      numOfTables = tSimpleHashGetSize(pReader->status.pTableMap);
   bool         asc = ASCENDING_TRAVERSE(pReader->info.order);
-  int32_t      step = asc ? 1 : -1;
   STimeWindow  w = pReader->info.window;
   SBrinRecord* pRecord = NULL;
 
@@ -710,7 +709,8 @@ static int32_t doLoadFileBlock(STsdbReader* pReader, SArray* pIndexList, SBlockN
       continue;
     }
 
-    if (pkCompEx(pReader->pkComparFn, &pRecord->lastKey.key, &pScanInfo->lastProcKey) <= 0) {
+    int32_t ret = pkCompEx(pReader->pkComparFn, &pRecord->lastKey.key, &pScanInfo->lastProcKey);
+    if ((asc && ret <= 0) || (!asc && ret >= 0)) {
       continue;
     }
 

From 52a04cd00822576b90dc38ed80c40211285cbd3b Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Tue, 26 Mar 2024 10:55:56 +0800
Subject: [PATCH 15/32] fix(tsdb): fix bug for filter brin record.

---
 source/dnode/vnode/src/tsdb/tsdbRead2.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/source/dnode/vnode/src/tsdb/tsdbRead2.c b/source/dnode/vnode/src/tsdb/tsdbRead2.c
index 7505230eca..c7164d43de 100644
--- a/source/dnode/vnode/src/tsdb/tsdbRead2.c
+++ b/source/dnode/vnode/src/tsdb/tsdbRead2.c
@@ -709,9 +709,14 @@ static int32_t doLoadFileBlock(STsdbReader* pReader, SArray* pIndexList, SBlockN
       continue;
     }
 
-    int32_t ret = pkCompEx(pReader->pkComparFn, &pRecord->lastKey.key, &pScanInfo->lastProcKey);
-    if ((asc && ret <= 0) || (!asc && ret >= 0)) {
-      continue;
+    if (asc) {
+      if (pkCompEx(pReader->pkComparFn, &pRecord->lastKey.key, &pScanInfo->lastProcKey) <= 0) {
+        continue;
+      }
+    } else {
+      if (pkCompEx(pReader->pkComparFn, &pRecord->firstKey.key, &pScanInfo->lastProcKey) >= 0) {
+        continue;
+      }
     }
 
     // 2. version range check, version range is an CLOSED interval

From f778dce52d04220e375c384cf0ce4bcf19e3f202 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Tue, 26 Mar 2024 14:13:00 +0800
Subject: [PATCH 16/32] fix: add pk for merge generated during merge scan node
 split

---
 source/libs/planner/src/planSpliter.c | 59 ++++++++++++++++++++-------
 1 file changed, 45 insertions(+), 14 deletions(-)

diff --git a/source/libs/planner/src/planSpliter.c b/source/libs/planner/src/planSpliter.c
index 34be489333..dd0cb5a964 100644
--- a/source/libs/planner/src/planSpliter.c
+++ b/source/libs/planner/src/planSpliter.c
@@ -690,19 +690,23 @@ static int32_t stbSplCreateExchangeNode(SSplitContext* pCxt, SLogicNode* pParent
   return code;
 }
 
+static int32_t stbSplCreateMergeKeysByExpr(SNode* pExpr, EOrder order, SNodeList** pMergeKeys) {
+  SOrderByExprNode* pOrderByExpr = (SOrderByExprNode*)nodesMakeNode(QUERY_NODE_ORDER_BY_EXPR);
+  if (NULL == pOrderByExpr) {
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
+  pOrderByExpr->pExpr = nodesCloneNode(pExpr);
+  if (NULL == pOrderByExpr->pExpr) {
+    nodesDestroyNode((SNode*)pOrderByExpr);
+    return TSDB_CODE_OUT_OF_MEMORY;
+  }
+  pOrderByExpr->order = order;
+  pOrderByExpr->nullOrder = (order == ORDER_ASC) ? NULL_ORDER_FIRST : NULL_ORDER_LAST;
+  return nodesListMakeStrictAppend(pMergeKeys, (SNode*)pOrderByExpr);
+}
+
 static int32_t stbSplCreateMergeKeysByPrimaryKey(SNode* pPrimaryKey, EOrder order, SNodeList** pMergeKeys) {
-  SOrderByExprNode* pMergeKey = (SOrderByExprNode*)nodesMakeNode(QUERY_NODE_ORDER_BY_EXPR);
-  if (NULL == pMergeKey) {
-    return TSDB_CODE_OUT_OF_MEMORY;
-  }
-  pMergeKey->pExpr = nodesCloneNode(pPrimaryKey);
-  if (NULL == pMergeKey->pExpr) {
-    nodesDestroyNode((SNode*)pMergeKey);
-    return TSDB_CODE_OUT_OF_MEMORY;
-  }
-  pMergeKey->order = order;
-  pMergeKey->nullOrder = NULL_ORDER_FIRST;
-  return nodesListMakeStrictAppend(pMergeKeys, (SNode*)pMergeKey);
+  return stbSplCreateMergeKeysByExpr(pPrimaryKey, order, pMergeKeys);
 }
 
 static int32_t stbSplSplitIntervalForBatch(SSplitContext* pCxt, SStableSplitInfo* pInfo) {
@@ -1357,6 +1361,28 @@ static SNode* stbSplFindPrimaryKeyFromScan(SScanLogicNode* pScan) {
   return pCol;
 }
 
+static SNode* stbSplFindPkFromScan(SScanLogicNode* pScan) {
+  bool   find = false;
+  SNode* pCol = NULL;
+  FOREACH(pCol, pScan->pScanCols) {
+    if (((SColumnNode*)pCol)->isPk) {
+      find = true;
+      break;
+    }
+  }
+  if (!find) {
+    return NULL;
+  }
+  SNode* pTarget = NULL;
+  FOREACH(pTarget, pScan->node.pTargets) {
+    if (nodesEqualNode(pTarget, pCol)) {
+      return pCol;
+    }
+  }
+  nodesListStrictAppend(pScan->node.pTargets, nodesCloneNode(pCol));
+  return pCol;
+}
+
 static int32_t stbSplCreateMergeScanNode(SScanLogicNode* pScan, SLogicNode** pOutputMergeScan,
                                          SNodeList** pOutputMergeKeys) {
   SNodeList* pChildren = pScan->node.pChildren;
@@ -1374,8 +1400,13 @@ static int32_t stbSplCreateMergeScanNode(SScanLogicNode* pScan, SLogicNode** pOu
     pMergeScan->filesetDelimited = true;
     pMergeScan->node.pChildren = pChildren;
     splSetParent((SLogicNode*)pMergeScan);
-    code = stbSplCreateMergeKeysByPrimaryKey(stbSplFindPrimaryKeyFromScan(pMergeScan),
-                                             pMergeScan->scanSeq[0] > 0 ? ORDER_ASC : ORDER_DESC, &pMergeKeys);
+
+    SNode* pTs = stbSplFindPrimaryKeyFromScan(pMergeScan);
+    code = stbSplCreateMergeKeysByPrimaryKey(pTs, pMergeScan->scanSeq[0] > 0 ? ORDER_ASC : ORDER_DESC, &pMergeKeys);
+    SNode* pPk = stbSplFindPkFromScan(pMergeScan);
+    if (TSDB_CODE_SUCCESS == code && NULL != pPk) {
+      code = stbSplCreateMergeKeysByExpr(pPk, pMergeScan->scanSeq[0] > 0 ? ORDER_ASC : ORDER_DESC, &pMergeKeys);
+    }
   }
 
   if (TSDB_CODE_SUCCESS == code) {

From ee4dda56c12d43c8a9b2c9a539fa1575a2884843 Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Tue, 26 Mar 2024 14:43:28 +0800
Subject: [PATCH 17/32] more code

---
 source/common/src/tdataformat.c       | 22 ++++++++--------
 source/dnode/vnode/src/vnd/vnodeSvr.c | 36 ++++++++++++++++++++-------
 2 files changed, 39 insertions(+), 19 deletions(-)

diff --git a/source/common/src/tdataformat.c b/source/common/src/tdataformat.c
index af124f7a02..cf8e28bf9a 100644
--- a/source/common/src/tdataformat.c
+++ b/source/common/src/tdataformat.c
@@ -557,13 +557,10 @@ void tRowDestroy(SRow *pRow) {
 }
 
 static int32_t tRowPCmprFn(const void *p1, const void *p2) {
-  if ((*(SRow **)p1)->ts < (*(SRow **)p2)->ts) {
-    return -1;
-  } else if ((*(SRow **)p1)->ts > (*(SRow **)p2)->ts) {
-    return 1;
-  }
-
-  return 0;
+  SRowKey key1, key2;
+  tRowGetKey(*(SRow **)p1, &key1);
+  tRowGetKey(*(SRow **)p2, &key2);
+  return tRowKeyCompare(&key1, &key2);
 }
 static void    tRowPDestroy(SRow **ppRow) { tRowDestroy(*ppRow); }
 static int32_t tRowMergeImpl(SArray *aRowP, STSchema *pTSchema, int32_t iStart, int32_t iEnd, int8_t flag) {
@@ -645,13 +642,18 @@ int32_t tRowMerge(SArray *aRowP, STSchema *pTSchema, int8_t flag) {
 
   int32_t iStart = 0;
   while (iStart < aRowP->size) {
-    SRow *pRow = (SRow *)taosArrayGetP(aRowP, iStart);
+    SRowKey key1;
+    SRow   *row1 = (SRow *)taosArrayGetP(aRowP, iStart);
+
+    tRowGetKey(row1, &key1);
 
     int32_t iEnd = iStart + 1;
     while (iEnd < aRowP->size) {
-      SRow *pRowT = (SRow *)taosArrayGetP(aRowP, iEnd);
+      SRowKey key2;
+      SRow   *row2 = (SRow *)taosArrayGetP(aRowP, iEnd);
+      tRowGetKey(row2, &key2);
 
-      if (pRow->ts != pRowT->ts) break;
+      if (tRowKeyCompare(&key1, &key2) != 0) break;
 
       iEnd++;
     }
diff --git a/source/dnode/vnode/src/vnd/vnodeSvr.c b/source/dnode/vnode/src/vnd/vnodeSvr.c
index a03adb164b..a4af782ad7 100644
--- a/source/dnode/vnode/src/vnd/vnodeSvr.c
+++ b/source/dnode/vnode/src/vnd/vnodeSvr.c
@@ -32,14 +32,14 @@ static int32_t vnodeProcessAlterTbReq(SVnode *pVnode, int64_t ver, void *pReq, i
 static int32_t vnodeProcessDropTbReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp,
                                      SRpcMsg *pOriginRpc);
 static int32_t vnodeProcessSubmitReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp,
-                                    SRpcMsg *pOriginalMsg);
+                                     SRpcMsg *pOriginalMsg);
 static int32_t vnodeProcessCreateTSmaReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp);
 static int32_t vnodeProcessAlterConfirmReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp);
 static int32_t vnodeProcessAlterConfigReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp);
 static int32_t vnodeProcessDropTtlTbReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp);
 static int32_t vnodeProcessTrimReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp);
 static int32_t vnodeProcessDeleteReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp,
-                                      SRpcMsg *pOriginalMsg);
+                                     SRpcMsg *pOriginalMsg);
 static int32_t vnodeProcessBatchDeleteReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp);
 static int32_t vnodeProcessCreateIndexReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp);
 static int32_t vnodeProcessDropIndexReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp);
@@ -1516,7 +1516,7 @@ static int32_t vnodeRebuildSubmitReqMsg(SSubmitReq2 *pSubmitReq, void **ppMsg) {
 }
 
 static int32_t vnodeProcessSubmitReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp,
-                                    SRpcMsg *pOriginalMsg) {
+                                     SRpcMsg *pOriginalMsg) {
   int32_t code = 0;
   terrno = 0;
 
@@ -1586,12 +1586,26 @@ static int32_t vnodeProcessSubmitReq(SVnode *pVnode, int64_t ver, void *pReq, in
     } else {
       int32_t nRow = TARRAY_SIZE(pSubmitTbData->aRowP);
       SRow  **aRow = (SRow **)TARRAY_DATA(pSubmitTbData->aRowP);
+      SRowKey lastRowKey;
       for (int32_t iRow = 0; iRow < nRow; ++iRow) {
-        if (aRow[iRow]->ts < minKey || aRow[iRow]->ts > maxKey || (iRow > 0 && aRow[iRow]->ts <= aRow[iRow - 1]->ts)) {
+        if (aRow[iRow]->ts < minKey || aRow[iRow]->ts > maxKey) {
           code = TSDB_CODE_INVALID_MSG;
           vError("vgId:%d %s failed since %s, version:%" PRId64, TD_VID(pVnode), __func__, tstrerror(code), ver);
           goto _exit;
         }
+        if (iRow == 0) {
+          tRowGetKey(aRow[iRow], &lastRowKey);
+        } else {
+          SRowKey rowKey;
+          tRowGetKey(aRow[iRow], &rowKey);
+
+          if (tRowKeyCompare(&lastRowKey, &rowKey) >= 0) {
+            code = TSDB_CODE_INVALID_MSG;
+            vError("vgId:%d %s failed since %s, version:%" PRId64, TD_VID(pVnode), __func__, tstrerror(code), ver);
+            goto _exit;
+          }
+          lastRowKey = rowKey;
+        }
       }
     }
   }
@@ -1735,10 +1749,14 @@ _exit:
   atomic_add_fetch_64(&pVnode->statis.nInsertSuccess, pSubmitRsp->affectedRows);
   atomic_add_fetch_64(&pVnode->statis.nBatchInsert, 1);
 
-  if(tsEnableMonitor && pSubmitRsp->affectedRows > 0 && strlen(pOriginalMsg->info.conn.user) > 0){
-    const char *sample_labels[] = {VNODE_METRIC_TAG_VALUE_INSERT_AFFECTED_ROWS, pVnode->monitor.strClusterId,
-                                    pVnode->monitor.strDnodeId, tsLocalEp, pVnode->monitor.strVgId,
-                                    pOriginalMsg->info.conn.user, "Success"};
+  if (tsEnableMonitor && pSubmitRsp->affectedRows > 0 && strlen(pOriginalMsg->info.conn.user) > 0) {
+    const char *sample_labels[] = {VNODE_METRIC_TAG_VALUE_INSERT_AFFECTED_ROWS,
+                                   pVnode->monitor.strClusterId,
+                                   pVnode->monitor.strDnodeId,
+                                   tsLocalEp,
+                                   pVnode->monitor.strVgId,
+                                   pOriginalMsg->info.conn.user,
+                                   "Success"};
     taos_counter_add(pVnode->monitor.insertCounter, pSubmitRsp->affectedRows, sample_labels);
   }
 
@@ -2019,7 +2037,7 @@ static int32_t vnodeProcessBatchDeleteReq(SVnode *pVnode, int64_t ver, void *pRe
 }
 
 static int32_t vnodeProcessDeleteReq(SVnode *pVnode, int64_t ver, void *pReq, int32_t len, SRpcMsg *pRsp,
-                                    SRpcMsg *pOriginalMsg) {
+                                     SRpcMsg *pOriginalMsg) {
   int32_t     code = 0;
   SDecoder   *pCoder = &(SDecoder){0};
   SDeleteRes *pRes = &(SDeleteRes){0};

From 93ca55954fcd6f9d209ea204af8030e09d6ef32f Mon Sep 17 00:00:00 2001
From: Hongze Cheng <hzcheng@taosdata.com>
Date: Tue, 26 Mar 2024 15:11:15 +0800
Subject: [PATCH 18/32] more code change

---
 include/libs/qcom/query.h              |   2 +-
 source/libs/parser/inc/parInsertUtil.h |   4 +-
 source/libs/parser/src/parInsertSml.c  |  31 ++--
 source/libs/parser/src/parInsertSql.c  | 194 ++++++++++++-------------
 source/libs/parser/src/parInsertUtil.c |  60 ++++----
 5 files changed, 148 insertions(+), 143 deletions(-)

diff --git a/include/libs/qcom/query.h b/include/libs/qcom/query.h
index 8d6cd6a3c0..0b192d5593 100644
--- a/include/libs/qcom/query.h
+++ b/include/libs/qcom/query.h
@@ -197,7 +197,7 @@ typedef struct STableDataCxt {
   SBoundColInfo  boundColsInfo;
   SArray*        pValues;
   SSubmitTbData* pData;
-  TSKEY          lastTs;
+  SRowKey        lastKey;
   bool           ordered;
   bool           duplicateTs;
 } STableDataCxt;
diff --git a/source/libs/parser/inc/parInsertUtil.h b/source/libs/parser/inc/parInsertUtil.h
index b20587dd43..1988620539 100644
--- a/source/libs/parser/inc/parInsertUtil.h
+++ b/source/libs/parser/inc/parInsertUtil.h
@@ -45,8 +45,8 @@ int16_t insFindCol(struct SToken *pColname, int16_t start, int16_t end, SSchema
 void    insBuildCreateTbReq(SVCreateTbReq *pTbReq, const char *tname, STag *pTag, int64_t suid, const char *sname,
                             SArray *tagName, uint8_t tagNum, int32_t ttl);
 int32_t insInitBoundColsInfo(int32_t numOfBound, SBoundColInfo *pInfo);
-void    insInitColValues(STableMeta* pTableMeta, SArray* aColValues);
-void    insCheckTableDataOrder(STableDataCxt *pTableCxt, TSKEY tsKey);
+void    insInitColValues(STableMeta *pTableMeta, SArray *aColValues);
+void    insCheckTableDataOrder(STableDataCxt *pTableCxt, SRowKey *rowKey);
 int32_t insGetTableDataCxt(SHashObj *pHash, void *id, int32_t idLen, STableMeta *pTableMeta,
                            SVCreateTbReq **pCreateTbReq, STableDataCxt **pTableCxt, bool colMode, bool ignoreColVals);
 int32_t initTableColSubmitData(STableDataCxt *pTableCxt);
diff --git a/source/libs/parser/src/parInsertSml.c b/source/libs/parser/src/parInsertSml.c
index c2ce77f02b..fcb5588717 100644
--- a/source/libs/parser/src/parInsertSml.c
+++ b/source/libs/parser/src/parInsertSml.c
@@ -113,7 +113,8 @@ static int32_t smlBuildTagRow(SArray* cols, SBoundColInfo* tags, SSchema* pSchem
     SSchema* pTagSchema = &pSchema[tags->pColIndex[i]];
     SSmlKv*  kv = taosArrayGet(cols, i);
 
-    if(kv->keyLen != strlen(pTagSchema->name) || memcmp(kv->key, pTagSchema->name, kv->keyLen) != 0 || kv->type != pTagSchema->type){
+    if (kv->keyLen != strlen(pTagSchema->name) || memcmp(kv->key, pTagSchema->name, kv->keyLen) != 0 ||
+        kv->type != pTagSchema->type) {
       code = TSDB_CODE_SML_INVALID_DATA;
       uError("SML smlBuildTagRow error col not same %s", pTagSchema->name);
       goto end;
@@ -200,7 +201,9 @@ int32_t smlBuildRow(STableDataCxt* pTableCxt) {
   if (TSDB_CODE_SUCCESS != ret) {
     return ret;
   }
-  insCheckTableDataOrder(pTableCxt, TD_ROW_KEY(*pRow));
+  SRowKey key;
+  tRowGetKey(*pRow, &key);
+  insCheckTableDataOrder(pTableCxt, &key);
   return TSDB_CODE_SUCCESS;
 }
 
@@ -209,15 +212,16 @@ int32_t smlBuildCol(STableDataCxt* pTableCxt, SSchema* schema, void* data, int32
   SSchema* pColSchema = schema + index;
   SColVal* pVal = taosArrayGet(pTableCxt->pValues, index);
   SSmlKv*  kv = (SSmlKv*)data;
-  if(kv->keyLen != strlen(pColSchema->name) || memcmp(kv->key, pColSchema->name, kv->keyLen) != 0 || kv->type != pColSchema->type){
+  if (kv->keyLen != strlen(pColSchema->name) || memcmp(kv->key, pColSchema->name, kv->keyLen) != 0 ||
+      kv->type != pColSchema->type) {
     ret = TSDB_CODE_SML_INVALID_DATA;
     char* tmp = taosMemoryCalloc(kv->keyLen + 1, 1);
-    if(tmp){
+    if (tmp) {
       memcpy(tmp, kv->key, kv->keyLen);
-      uInfo("SML data(name:%s type:%s) is not same like the db data(name:%s type:%s)",
-            tmp, tDataTypes[kv->type].name, pColSchema->name, tDataTypes[pColSchema->type].name);
+      uInfo("SML data(name:%s type:%s) is not same like the db data(name:%s type:%s)", tmp, tDataTypes[kv->type].name,
+            pColSchema->name, tDataTypes[pColSchema->type].name);
       taosMemoryFree(tmp);
-    }else{
+    } else {
       uError("SML smlBuildCol out of memory");
     }
     goto end;
@@ -225,11 +229,11 @@ int32_t smlBuildCol(STableDataCxt* pTableCxt, SSchema* schema, void* data, int32
   if (kv->type == TSDB_DATA_TYPE_NCHAR) {
     int32_t len = 0;
     int64_t size = pColSchema->bytes - VARSTR_HEADER_SIZE;
-    if(size <= 0){
+    if (size <= 0) {
       ret = TSDB_CODE_SML_INVALID_DATA;
       goto end;
     }
-    char*   pUcs4 = taosMemoryCalloc(1, size);
+    char* pUcs4 = taosMemoryCalloc(1, size);
     if (NULL == pUcs4) {
       ret = TSDB_CODE_OUT_OF_MEMORY;
       goto end;
@@ -351,7 +355,7 @@ int32_t smlBindData(SQuery* query, bool dataFormat, SArray* tags, SArray* colsSc
         continue;
       }
       SSmlKv* kv = *(SSmlKv**)p;
-      if(kv->type != pColSchema->type){
+      if (kv->type != pColSchema->type) {
         ret = buildInvalidOperationMsg(&pBuf, "kv type not equal to col type");
         goto end;
       }
@@ -367,7 +371,8 @@ int32_t smlBindData(SQuery* query, bool dataFormat, SArray* tags, SArray* colsSc
         }
         if (!taosMbsToUcs4(kv->value, kv->length, (TdUcs4*)pUcs4, pColSchema->bytes - VARSTR_HEADER_SIZE, &len)) {
           if (errno == E2BIG) {
-            uError("sml bind taosMbsToUcs4 error, kv length:%d, bytes:%d, kv->value:%s", (int)kv->length, pColSchema->bytes, kv->value);
+            uError("sml bind taosMbsToUcs4 error, kv length:%d, bytes:%d, kv->value:%s", (int)kv->length,
+                   pColSchema->bytes, kv->value);
             buildInvalidOperationMsg(&pBuf, "value too long");
             ret = TSDB_CODE_PAR_VALUE_TOO_LONG;
             goto end;
@@ -396,7 +401,9 @@ int32_t smlBindData(SQuery* query, bool dataFormat, SArray* tags, SArray* colsSc
       buildInvalidOperationMsg(&pBuf, "tRowBuild error");
       goto end;
     }
-    insCheckTableDataOrder(pTableCxt, TD_ROW_KEY(*pRow));
+    SRowKey key;
+    tRowGetKey(*pRow, &key);
+    insCheckTableDataOrder(pTableCxt, &key);
     clearColValArraySml(pTableCxt->pValues);
   }
 
diff --git a/source/libs/parser/src/parInsertSql.c b/source/libs/parser/src/parInsertSql.c
index 77c63c38d8..1a41ba5fb9 100644
--- a/source/libs/parser/src/parInsertSql.c
+++ b/source/libs/parser/src/parInsertSql.c
@@ -13,12 +13,12 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
+#include "geosWrapper.h"
 #include "parInsertUtil.h"
 #include "parToken.h"
 #include "scalar.h"
 #include "tglobal.h"
 #include "ttime.h"
-#include "geosWrapper.h"
 
 typedef struct SInsertParseContext {
   SParseContext* pComCxt;
@@ -29,7 +29,7 @@ typedef struct SInsertParseContext {
   bool           usingDuplicateTable;
   bool           forceUpdate;
   bool           needTableTagVal;
-  bool           needRequest;       // whether or not request server
+  bool           needRequest;  // whether or not request server
 } SInsertParseContext;
 
 typedef int32_t (*_row_append_fn_t)(SMsgBuf* pMsgBuf, const void* value, int32_t len, void* param);
@@ -154,19 +154,15 @@ static int32_t parseDuplicateUsingClause(SInsertParseContext* pCxt, SVnodeModify
   return code;
 }
 
-typedef enum {
-  BOUND_TAGS,
-  BOUND_COLUMNS,
-  BOUND_ALL_AND_TBNAME
-} EBoundColumnsType;
+typedef enum { BOUND_TAGS, BOUND_COLUMNS, BOUND_ALL_AND_TBNAME } EBoundColumnsType;
 
 static int32_t getTbnameSchemaIndex(STableMeta* pTableMeta) {
   return pTableMeta->tableInfo.numOfTags + pTableMeta->tableInfo.numOfColumns;
 }
 
 // pStmt->pSql -> field1_name, ...)
-static int32_t parseBoundColumns(SInsertParseContext* pCxt, const char** pSql, EBoundColumnsType boundColsType, STableMeta* pTableMeta,
-                                 SBoundColInfo* pBoundInfo) {
+static int32_t parseBoundColumns(SInsertParseContext* pCxt, const char** pSql, EBoundColumnsType boundColsType,
+                                 STableMeta* pTableMeta, SBoundColInfo* pBoundInfo) {
   SSchema* pSchema = NULL;
   if (boundColsType == BOUND_TAGS) {
     pSchema = getTableTagSchema(pTableMeta);
@@ -202,8 +198,7 @@ static int32_t parseBoundColumns(SInsertParseContext* pCxt, const char** pSql, E
     token.z = tmpTokenBuf;
     token.n = strdequote(token.z);
 
-    if (boundColsType == BOUND_ALL_AND_TBNAME &&
-        token.n == strlen("tbname") && (strcasecmp(token.z, "tbname") == 0)) {
+    if (boundColsType == BOUND_ALL_AND_TBNAME && token.n == strlen("tbname") && (strcasecmp(token.z, "tbname") == 0)) {
       pBoundInfo->pColIndex[pBoundInfo->numOfBound] = tbnameSchemaIndex;
       pUseCols[tbnameSchemaIndex] = true;
       ++pBoundInfo->numOfBound;
@@ -230,7 +225,7 @@ static int32_t parseBoundColumns(SInsertParseContext* pCxt, const char** pSql, E
   if (TSDB_CODE_SUCCESS == code && (BOUND_TAGS != boundColsType) && !pUseCols[0]) {
     code = buildInvalidOperationMsg(&pCxt->msg, "primary timestamp column can not be null");
   }
-  if (TSDB_CODE_SUCCESS == code && (BOUND_ALL_AND_TBNAME == boundColsType) &&!pUseCols[tbnameSchemaIndex]) {
+  if (TSDB_CODE_SUCCESS == code && (BOUND_ALL_AND_TBNAME == boundColsType) && !pUseCols[tbnameSchemaIndex]) {
     code = buildInvalidOperationMsg(&pCxt->msg, "tbname column can not be null");
   }
   taosMemoryFree(pUseCols);
@@ -285,10 +280,11 @@ static int parseTime(const char** end, SToken* pToken, int16_t timePrec, int64_t
   int32_t     index = 0, i = 0;
   int64_t     interval = 0, tempInterval = 0;
   int64_t     ts = 0, tempTs = 0;
-  bool firstIsTS = false, secondIsTs = false;
+  bool        firstIsTS = false, secondIsTs = false;
   const char* pTokenEnd = *end;
 
-  if (TSDB_CODE_SUCCESS != parseTimestampOrInterval(&pTokenEnd, pToken, timePrec, &ts, &interval, pMsgBuf, &firstIsTS)) {
+  if (TSDB_CODE_SUCCESS !=
+      parseTimestampOrInterval(&pTokenEnd, pToken, timePrec, &ts, &interval, pMsgBuf, &firstIsTS)) {
     return buildSyntaxErrMsg(pMsgBuf, "invalid timestamp format", pToken->z);
   }
 
@@ -330,8 +326,7 @@ static int parseTime(const char** end, SToken* pToken, int16_t timePrec, int64_t
     if (pTokenEnd[i] == ' ' || pTokenEnd[i] == '\t') {
       i++;
       continue;
-    }
-    else if (pTokenEnd[i] == ',' || pTokenEnd[i] == ')') {
+    } else if (pTokenEnd[i] == ',' || pTokenEnd[i] == ')') {
       *end = pTokenEnd + i;
       if (!firstIsTS) {
         return buildSyntaxErrMsg(pMsgBuf, "invalid timestamp format", pToken->z);
@@ -362,7 +357,8 @@ static int parseTime(const char** end, SToken* pToken, int16_t timePrec, int64_t
       valueToken.n = len;
     }
 
-    if (TSDB_CODE_SUCCESS != parseTimestampOrInterval(&pTokenEnd, &valueToken, timePrec, &tempTs, &tempInterval, pMsgBuf, &secondIsTs)) {
+    if (TSDB_CODE_SUCCESS !=
+        parseTimestampOrInterval(&pTokenEnd, &valueToken, timePrec, &tempTs, &tempInterval, pMsgBuf, &secondIsTs)) {
       return buildSyntaxErrMsg(pMsgBuf, "invalid timestamp format", pToken->z);
     }
 
@@ -376,7 +372,7 @@ static int parseTime(const char** end, SToken* pToken, int16_t timePrec, int64_t
         return buildSyntaxErrMsg(pMsgBuf, "invalid timestamp format", pToken->z);
       }
       ts = tempTs;
-    }else {
+    } else {
       // not support operator between tow interval, such as 2h + 3s
       if (!firstIsTS) {
         return buildSyntaxErrMsg(pMsgBuf, "invalid timestamp format", pToken->z);
@@ -413,7 +409,7 @@ static int parseTime(const char** end, SToken* pToken, int16_t timePrec, int64_t
 }
 
 // need to call geosFreeBuffer(*output) later
-static int parseGeometry(SToken *pToken, unsigned char **output, size_t *size) {
+static int parseGeometry(SToken* pToken, unsigned char** output, size_t* size) {
   int32_t code = TSDB_CODE_FAILED;
 
   //[ToDo] support to parse WKB as well as WKT
@@ -432,19 +428,19 @@ static int parseGeometry(SToken *pToken, unsigned char **output, size_t *size) {
   return code;
 }
 
-static int32_t parseVarbinary(SToken* pToken, uint8_t **pData, uint32_t *nData, int32_t bytes){
-  if(pToken->type != TK_NK_STRING){
+static int32_t parseVarbinary(SToken* pToken, uint8_t** pData, uint32_t* nData, int32_t bytes) {
+  if (pToken->type != TK_NK_STRING) {
     return TSDB_CODE_PAR_INVALID_VARBINARY;
   }
 
-  if(isHex(pToken->z + 1, pToken->n - 2)){
-    if(!isValidateHex(pToken->z + 1, pToken->n - 2)){
+  if (isHex(pToken->z + 1, pToken->n - 2)) {
+    if (!isValidateHex(pToken->z + 1, pToken->n - 2)) {
       return TSDB_CODE_PAR_INVALID_VARBINARY;
     }
 
-    void* data = NULL;
+    void*    data = NULL;
     uint32_t size = 0;
-    if(taosHex2Ascii(pToken->z + 1, pToken->n - 2, &data, &size) < 0){
+    if (taosHex2Ascii(pToken->z + 1, pToken->n - 2, &data, &size) < 0) {
       return TSDB_CODE_OUT_OF_MEMORY;
     }
 
@@ -454,7 +450,7 @@ static int32_t parseVarbinary(SToken* pToken, uint8_t **pData, uint32_t *nData,
     }
     *pData = data;
     *nData = size;
-  }else{
+  } else {
     *pData = taosMemoryCalloc(1, pToken->n);
     int32_t len = trimString(pToken->z, pToken->n, *pData, pToken->n);
     *nData = len;
@@ -633,7 +629,7 @@ static int32_t parseTagToken(const char** end, SToken* pToken, SSchema* pSchema,
     }
     case TSDB_DATA_TYPE_VARBINARY: {
       code = parseVarbinary(pToken, &val->pData, &val->nData, pSchema->bytes);
-      if(code != TSDB_CODE_SUCCESS){
+      if (code != TSDB_CODE_SUCCESS) {
         return generateSyntaxErrMsg(pMsgBuf, code, pSchema->name);
       }
       break;
@@ -759,8 +755,8 @@ static int32_t buildCreateTbReq(SVnodeModifyOpStmt* pStmt, STag* pTag, SArray* p
 int32_t checkAndTrimValue(SToken* pToken, char* tmpTokenBuf, SMsgBuf* pMsgBuf, int8_t type) {
   if ((pToken->type != TK_NOW && pToken->type != TK_TODAY && pToken->type != TK_NK_INTEGER &&
        pToken->type != TK_NK_STRING && pToken->type != TK_NK_FLOAT && pToken->type != TK_NK_BOOL &&
-       pToken->type != TK_NULL && pToken->type != TK_NK_HEX && pToken->type != TK_NK_OCT &&
-       pToken->type != TK_NK_BIN && pToken->type != TK_NK_VARIABLE) ||
+       pToken->type != TK_NULL && pToken->type != TK_NK_HEX && pToken->type != TK_NK_OCT && pToken->type != TK_NK_BIN &&
+       pToken->type != TK_NK_VARIABLE) ||
       (pToken->n == 0) || (pToken->type == TK_NK_RP)) {
     return buildSyntaxErrMsg(pMsgBuf, "invalid data or symbol", pToken->z);
   }
@@ -1114,8 +1110,8 @@ static int32_t checkAuth(SParseContext* pCxt, SName* pTbName, bool* pMissCache,
   return code;
 }
 
-static int32_t getTableMeta(SInsertParseContext* pCxt, SName* pTbName, STableMeta** pTableMeta,
-                            bool* pMissCache, bool bUsingTable) {
+static int32_t getTableMeta(SInsertParseContext* pCxt, SName* pTbName, STableMeta** pTableMeta, bool* pMissCache,
+                            bool bUsingTable) {
   SParseContext* pComCxt = pCxt->pComCxt;
   int32_t        code = TSDB_CODE_SUCCESS;
   if (pComCxt->async) {
@@ -1220,7 +1216,7 @@ static int32_t getTargetTableSchema(SInsertParseContext* pCxt, SVnodeModifyOpStm
     pCxt->missCache = true;
     return TSDB_CODE_SUCCESS;
   }
-  SNode* pTagCond = NULL;
+  SNode*  pTagCond = NULL;
   int32_t code = checkAuth(pCxt->pComCxt, &pStmt->targetTableName, &pCxt->missCache, &pTagCond);
   if (TSDB_CODE_SUCCESS == code && !pCxt->missCache) {
     code = getTargetTableMetaAndVgroup(pCxt, pStmt, &pCxt->missCache);
@@ -1365,13 +1361,11 @@ static int32_t parseBoundColumnsClause(SInsertParseContext* pCxt, SVnodeModifyOp
       return buildSyntaxErrMsg(&pCxt->msg, "keyword VALUES or FILE is expected", token.z);
     }
     // pStmt->pSql -> field1_name, ...)
-    return parseBoundColumns(pCxt, &pStmt->pSql, BOUND_COLUMNS, pStmt->pTableMeta,
-                             &pTableCxt->boundColsInfo);
+    return parseBoundColumns(pCxt, &pStmt->pSql, BOUND_COLUMNS, pStmt->pTableMeta, &pTableCxt->boundColsInfo);
   }
 
   if (NULL != pStmt->pBoundCols) {
-    return parseBoundColumns(pCxt, &pStmt->pBoundCols, BOUND_COLUMNS, pStmt->pTableMeta,
-                             &pTableCxt->boundColsInfo);
+    return parseBoundColumns(pCxt, &pStmt->pBoundCols, BOUND_COLUMNS, pStmt->pTableMeta, &pTableCxt->boundColsInfo);
   }
 
   return TSDB_CODE_SUCCESS;
@@ -1679,34 +1673,35 @@ static void clearColValArray(SArray* pCols) {
 typedef struct SStbRowsDataContext {
   SName stbName;
 
-  STableMeta* pStbMeta;
-  SNode* pTagCond;
+  STableMeta*   pStbMeta;
+  SNode*        pTagCond;
   SBoundColInfo boundColsInfo;
 
   // the following fields are for each stb row
-  SArray* aTagVals;
-  SArray* aColVals;
-  SArray* aTagNames;
-  SName ctbName;
-  STag* pTag;
-  STableMeta* pCtbMeta;
+  SArray*        aTagVals;
+  SArray*        aColVals;
+  SArray*        aTagNames;
+  SName          ctbName;
+  STag*          pTag;
+  STableMeta*    pCtbMeta;
   SVCreateTbReq* pCreateCtbReq;
   bool           hasTimestampTag;
   bool           isJsonTag;
 } SStbRowsDataContext;
 
-typedef union SRowsDataContext{
-  STableDataCxt* pTableDataCxt;
+typedef union SRowsDataContext {
+  STableDataCxt*       pTableDataCxt;
   SStbRowsDataContext* pStbRowsCxt;
 } SRowsDataContext;
 
-static int32_t parseTbnameToken(SInsertParseContext* pCxt, SStbRowsDataContext* pStbRowsCxt, SToken* pToken, bool* pFoundCtbName) {
+static int32_t parseTbnameToken(SInsertParseContext* pCxt, SStbRowsDataContext* pStbRowsCxt, SToken* pToken,
+                                bool* pFoundCtbName) {
   *pFoundCtbName = false;
   int32_t code = checkAndTrimValue(pToken, pCxt->tmpTokenBuf, &pCxt->msg, TSDB_DATA_TYPE_BINARY);
   if (TK_NK_VARIABLE == pToken->type) {
     code = buildInvalidOperationMsg(&pCxt->msg, "not expected tbname");
   }
-  if (code == TSDB_CODE_SUCCESS){
+  if (code == TSDB_CODE_SUCCESS) {
     if (isNullValue(TSDB_DATA_TYPE_BINARY, pToken)) {
       return buildInvalidOperationMsg(&pCxt->msg, "tbname can not be null value");
     }
@@ -1717,7 +1712,7 @@ static int32_t parseTbnameToken(SInsertParseContext* pCxt, SStbRowsDataContext*
           if (pToken->z[i] == '.') {
             return buildInvalidOperationMsg(&pCxt->msg, "tbname can not contain '.'");
           } else {
-            pStbRowsCxt->ctbName.tname[i] = pToken->z[i]; 
+            pStbRowsCxt->ctbName.tname[i] = pToken->z[i];
           }
         }
         pStbRowsCxt->ctbName.tname[pToken->n] = '\0';
@@ -1733,9 +1728,8 @@ static int32_t parseTbnameToken(SInsertParseContext* pCxt, SStbRowsDataContext*
 }
 
 static int32_t processCtbTagsAfterCtbName(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt,
-                                            SStbRowsDataContext* pStbRowsCxt, bool ctbFirst,
-                                          const SToken* tagTokens, SSchema* const* tagSchemas,
-                                            int numOfTagTokens) {
+                                          SStbRowsDataContext* pStbRowsCxt, bool ctbFirst, const SToken* tagTokens,
+                                          SSchema* const* tagSchemas, int numOfTagTokens) {
   int32_t code = TSDB_CODE_SUCCESS;
   uint8_t precision = pStmt->pTableMeta->tableInfo.precision;
 
@@ -1749,8 +1743,8 @@ static int32_t processCtbTagsAfterCtbName(SInsertParseContext* pCxt, SVnodeModif
       }
 
       if (code == TSDB_CODE_SUCCESS) {
-        code = parseTagValue(&pCxt->msg, NULL, precision, pTagSchema, pTagToken, pStbRowsCxt->aTagNames, pStbRowsCxt->aTagVals,
-                             &pStbRowsCxt->pTag);
+        code = parseTagValue(&pCxt->msg, NULL, precision, pTagSchema, pTagToken, pStbRowsCxt->aTagNames,
+                             pStbRowsCxt->aTagVals, &pStbRowsCxt->pTag);
       }
     }
     if (code == TSDB_CODE_SUCCESS && !pStbRowsCxt->isJsonTag) {
@@ -1765,9 +1759,9 @@ static int32_t processCtbTagsAfterCtbName(SInsertParseContext* pCxt, SVnodeModif
 }
 
 static int32_t doGetStbRowValues(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt, const char** ppSql,
-                                 SStbRowsDataContext* pStbRowsCxt, SToken* pToken,
-                                 const SBoundColInfo* pCols, const SSchema* pSchemas,
-                                 SToken* tagTokens, SSchema** tagSchemas, int* pNumOfTagTokens, bool* bFoundTbName) {
+                                 SStbRowsDataContext* pStbRowsCxt, SToken* pToken, const SBoundColInfo* pCols,
+                                 const SSchema* pSchemas, SToken* tagTokens, SSchema** tagSchemas, int* pNumOfTagTokens,
+                                 bool* bFoundTbName) {
   int32_t code = TSDB_CODE_SUCCESS;
   SArray* pTagNames = pStbRowsCxt->aTagNames;
   SArray* pTagVals = pStbRowsCxt->aTagVals;
@@ -1791,10 +1785,10 @@ static int32_t doGetStbRowValues(SInsertParseContext* pCxt, SVnodeModifyOpStmt*
     }
     if (pCols->pColIndex[i] < numOfCols) {
       const SSchema* pSchema = &pSchemas[pCols->pColIndex[i]];
-      SColVal* pVal = taosArrayGet(pStbRowsCxt->aColVals, pCols->pColIndex[i]);
+      SColVal*       pVal = taosArrayGet(pStbRowsCxt->aColVals, pCols->pColIndex[i]);
       code = parseValueToken(pCxt, ppSql, pToken, (SSchema*)pSchema, precision, pVal);
       if (TK_NK_VARIABLE == pToken->type) {
-          code = buildInvalidOperationMsg(&pCxt->msg, "not expected row value");
+        code = buildInvalidOperationMsg(&pCxt->msg, "not expected row value");
       }
     } else if (pCols->pColIndex[i] < tbnameIdx) {
       const SSchema* pTagSchema = &pSchemas[pCols->pColIndex[i]];
@@ -1808,11 +1802,11 @@ static int32_t doGetStbRowValues(SInsertParseContext* pCxt, SVnodeModifyOpStmt*
           code = buildInvalidOperationMsg(&pCxt->msg, "not expected row value");
         }
         if (code == TSDB_CODE_SUCCESS) {
-          code = parseTagValue(&pCxt->msg, ppSql, precision, (SSchema*)pTagSchema, pToken, pTagNames, pTagVals, &pStbRowsCxt->pTag);
+          code = parseTagValue(&pCxt->msg, ppSql, precision, (SSchema*)pTagSchema, pToken, pTagNames, pTagVals,
+                               &pStbRowsCxt->pTag);
         }
       }
-    }
-    else if (pCols->pColIndex[i] == tbnameIdx) {
+    } else if (pCols->pColIndex[i] == tbnameIdx) {
       code = parseTbnameToken(pCxt, pStbRowsCxt, pToken, bFoundTbName);
     }
 
@@ -1826,22 +1820,21 @@ static int32_t doGetStbRowValues(SInsertParseContext* pCxt, SVnodeModifyOpStmt*
   return code;
 }
 
-static int32_t getStbRowValues(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt,  const char** ppSql,
-                               SStbRowsDataContext* pStbRowsCxt, bool* pGotRow,
-                               SToken* pToken, bool *pCtbFirst) {
+static int32_t getStbRowValues(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt, const char** ppSql,
+                               SStbRowsDataContext* pStbRowsCxt, bool* pGotRow, SToken* pToken, bool* pCtbFirst) {
   SBoundColInfo* pCols = &pStbRowsCxt->boundColsInfo;
   SSchema*       pSchemas = getTableColumnSchema(pStbRowsCxt->pStbMeta);
 
-  bool bFoundTbName = false;
+  bool        bFoundTbName = false;
   const char* pOrigSql = *ppSql;
 
-  int32_t code = TSDB_CODE_SUCCESS;
-  SToken tagTokens[TSDB_MAX_TAGS] = {0};
+  int32_t  code = TSDB_CODE_SUCCESS;
+  SToken   tagTokens[TSDB_MAX_TAGS] = {0};
   SSchema* tagSchemas[TSDB_MAX_TAGS] = {0};
-  int numOfTagTokens = 0;
+  int      numOfTagTokens = 0;
 
-  code = doGetStbRowValues(pCxt, pStmt, ppSql, pStbRowsCxt, pToken, pCols, pSchemas, tagTokens,
-                    tagSchemas, &numOfTagTokens, &bFoundTbName);
+  code = doGetStbRowValues(pCxt, pStmt, ppSql, pStbRowsCxt, pToken, pCols, pSchemas, tagTokens, tagSchemas,
+                           &numOfTagTokens, &bFoundTbName);
 
   if (code == TSDB_CODE_SUCCESS && !bFoundTbName) {
     code = buildSyntaxErrMsg(&pCxt->msg, "tbname value expected", pOrigSql);
@@ -1870,7 +1863,8 @@ static int32_t getStbRowValues(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pS
   return code;
 }
 
-static int32_t processCtbAutoCreationAndCtbMeta(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt, SStbRowsDataContext* pStbRowsCxt) {
+static int32_t processCtbAutoCreationAndCtbMeta(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt,
+                                                SStbRowsDataContext* pStbRowsCxt) {
   int32_t code = TSDB_CODE_SUCCESS;
 
   pStbRowsCxt->pCreateCtbReq = taosMemoryCalloc(1, sizeof(SVCreateTbReq));
@@ -1878,16 +1872,16 @@ static int32_t processCtbAutoCreationAndCtbMeta(SInsertParseContext* pCxt, SVnod
     code = TSDB_CODE_OUT_OF_MEMORY;
   }
   if (code == TSDB_CODE_SUCCESS) {
-    insBuildCreateTbReq(pStbRowsCxt->pCreateCtbReq, pStbRowsCxt->ctbName.tname, pStbRowsCxt->pTag, pStbRowsCxt->pStbMeta->uid,
-                        pStbRowsCxt->stbName.tname, pStbRowsCxt->aTagNames, getNumOfTags(pStbRowsCxt->pStbMeta),
-                        TSDB_DEFAULT_TABLE_TTL);
+    insBuildCreateTbReq(pStbRowsCxt->pCreateCtbReq, pStbRowsCxt->ctbName.tname, pStbRowsCxt->pTag,
+                        pStbRowsCxt->pStbMeta->uid, pStbRowsCxt->stbName.tname, pStbRowsCxt->aTagNames,
+                        getNumOfTags(pStbRowsCxt->pStbMeta), TSDB_DEFAULT_TABLE_TTL);
     pStbRowsCxt->pTag = NULL;
   }
 
   if (code == TSDB_CODE_SUCCESS) {
     char ctbFName[TSDB_TABLE_FNAME_LEN];
     tNameExtractFullName(&pStbRowsCxt->ctbName, ctbFName);
-    SVgroupInfo vg;
+    SVgroupInfo      vg;
     SRequestConnInfo conn = {.pTrans = pCxt->pComCxt->pTransporter,
                              .requestId = pCxt->pComCxt->requestId,
                              .requestObjRefId = pCxt->pComCxt->requestRid,
@@ -1907,7 +1901,6 @@ static int32_t processCtbAutoCreationAndCtbMeta(SInsertParseContext* pCxt, SVnod
   return code;
 }
 
-
 static void clearStbRowsDataContext(SStbRowsDataContext* pStbRowsCxt) {
   if (pStbRowsCxt == NULL) return;
 
@@ -1938,7 +1931,7 @@ static int32_t parseOneStbRow(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pSt
   }
 
   if (code == TSDB_CODE_SUCCESS && bFirstTable) {
-      code = processCtbAutoCreationAndCtbMeta(pCxt, pStmt, pStbRowsCxt);
+    code = processCtbAutoCreationAndCtbMeta(pCxt, pStmt, pStbRowsCxt);
   }
 
   code = insGetTableDataCxt(pStmt->pTableBlockHashObj, &pStbRowsCxt->pCtbMeta->uid, sizeof(pStbRowsCxt->pCtbMeta->uid),
@@ -1948,7 +1941,9 @@ static int32_t parseOneStbRow(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pSt
     SRow** pRow = taosArrayReserve((*ppTableDataCxt)->pData->aRowP, 1);
     code = tRowBuild(pStbRowsCxt->aColVals, (*ppTableDataCxt)->pSchema, pRow);
     if (TSDB_CODE_SUCCESS == code) {
-      insCheckTableDataOrder(*ppTableDataCxt, TD_ROW_KEY(*pRow));
+      SRowKey key;
+      tRowGetKey(*pRow, &key);
+      insCheckTableDataOrder(*ppTableDataCxt, &key);
     }
   }
 
@@ -1957,11 +1952,12 @@ static int32_t parseOneStbRow(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pSt
   }
 
   clearStbRowsDataContext(pStbRowsCxt);
-  
+
   return TSDB_CODE_SUCCESS;
 }
 
-static int parseOneRow(SInsertParseContext* pCxt, const char** pSql, STableDataCxt* pTableCxt, bool* pGotRow, SToken* pToken) {
+static int parseOneRow(SInsertParseContext* pCxt, const char** pSql, STableDataCxt* pTableCxt, bool* pGotRow,
+                       SToken* pToken) {
   SBoundColInfo* pCols = &pTableCxt->boundColsInfo;
   bool           isParseBindParam = false;
   SSchema*       pSchemas = getTableColumnSchema(pTableCxt->pMeta);
@@ -2014,7 +2010,9 @@ static int parseOneRow(SInsertParseContext* pCxt, const char** pSql, STableDataC
     SRow** pRow = taosArrayReserve(pTableCxt->pData->aRowP, 1);
     code = tRowBuild(pTableCxt->pValues, pTableCxt->pSchema, pRow);
     if (TSDB_CODE_SUCCESS == code) {
-      insCheckTableDataOrder(pTableCxt, TD_ROW_KEY(*pRow));
+      SRowKey key;
+      tRowGetKey(*pRow, &key);
+      insCheckTableDataOrder(pTableCxt, &key);
     }
   }
 
@@ -2112,10 +2110,11 @@ static int32_t parseCsvFile(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt
         code = parseOneRow(pCxt, (const char**)&pRow, rowsDataCxt.pTableDataCxt, &gotRow, &token);
       } else {
         STableDataCxt* pTableDataCxt = NULL;
-        code = parseOneStbRow(pCxt, pStmt, (const char**)&pRow, rowsDataCxt.pStbRowsCxt, &gotRow, &token, &pTableDataCxt);
+        code =
+            parseOneStbRow(pCxt, pStmt, (const char**)&pRow, rowsDataCxt.pStbRowsCxt, &gotRow, &token, &pTableDataCxt);
         if (code == TSDB_CODE_SUCCESS) {
           SStbRowsDataContext* pStbRowsCxt = rowsDataCxt.pStbRowsCxt;
-          void* pData = pTableDataCxt;
+          void*                pData = pTableDataCxt;
           taosHashPut(pStmt->pTableCxtHashObj, &pStbRowsCxt->pCtbMeta->uid, sizeof(pStbRowsCxt->pCtbMeta->uid), &pData,
                       POINTER_BYTES);
         }
@@ -2149,11 +2148,11 @@ static int32_t parseCsvFile(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt
   return code;
 }
 
-static int32_t parseDataFromFileImpl(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt, SRowsDataContext rowsDataCxt) {
+static int32_t parseDataFromFileImpl(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt,
+                                     SRowsDataContext rowsDataCxt) {
   // init only for file
   if (NULL == pStmt->pTableCxtHashObj) {
-    pStmt->pTableCxtHashObj =
-        taosHashInit(128, taosGetDefaultHashFunction(TSDB_DATA_TYPE_BINARY), true, HASH_NO_LOCK);
+    pStmt->pTableCxtHashObj = taosHashInit(128, taosGetDefaultHashFunction(TSDB_DATA_TYPE_BINARY), true, HASH_NO_LOCK);
   }
   int32_t numOfRows = 0;
   int32_t code = parseCsvFile(pCxt, pStmt, rowsDataCxt, &numOfRows);
@@ -2267,7 +2266,7 @@ static int32_t constructStbRowsDataContext(SVnodeModifyOpStmt* pStmt, SStbRowsDa
   cloneTableMeta(pStbRowsCxt->pStbMeta, &pStbRowsCxt->pCtbMeta);
   pStbRowsCxt->pCtbMeta->tableType = TSDB_CHILD_TABLE;
   pStbRowsCxt->pCtbMeta->suid = pStbRowsCxt->pStbMeta->uid;
-  
+
   pStbRowsCxt->aTagNames = taosArrayInit(8, TSDB_COL_NAME_LEN);
   pStbRowsCxt->aTagVals = taosArrayInit(8, sizeof(STagVal));
 
@@ -2285,7 +2284,8 @@ static int32_t constructStbRowsDataContext(SVnodeModifyOpStmt* pStmt, SStbRowsDa
 static int32_t parseInsertStbClauseBottom(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt) {
   int32_t code = TSDB_CODE_SUCCESS;
   if (!pStmt->pBoundCols) {
-    return buildSyntaxErrMsg(&pCxt->msg, "(...tbname, ts...) bounded cols is expected for supertable insertion", pStmt->pSql);
+    return buildSyntaxErrMsg(&pCxt->msg, "(...tbname, ts...) bounded cols is expected for supertable insertion",
+                             pStmt->pSql);
   }
 
   SStbRowsDataContext* pStbRowsCxt = NULL;
@@ -2297,7 +2297,7 @@ static int32_t parseInsertStbClauseBottom(SInsertParseContext* pCxt, SVnodeModif
     pStbRowsCxt->hasTimestampTag = false;
     for (int32_t i = 0; i < pStbRowsCxt->boundColsInfo.numOfBound; ++i) {
       int16_t schemaIndex = pStbRowsCxt->boundColsInfo.pColIndex[i];
-      if (schemaIndex != getTbnameSchemaIndex(pStmt->pTableMeta) && schemaIndex >= getNumOfColumns(pStmt->pTableMeta) ) {
+      if (schemaIndex != getTbnameSchemaIndex(pStmt->pTableMeta) && schemaIndex >= getNumOfColumns(pStmt->pTableMeta)) {
         if (pStmt->pTableMeta->schema[schemaIndex].type == TSDB_DATA_TYPE_TIMESTAMP) {
           pStbRowsCxt->hasTimestampTag = true;
         }
@@ -2326,8 +2326,8 @@ static int32_t parseInsertTableClauseBottom(SInsertParseContext* pCxt, SVnodeMod
     return buildInvalidOperationMsg(&pCxt->msg, "insert into super table syntax is not supported for stmt");
   }
   if (!pStmt->stbSyntax) {
-    STableDataCxt* pTableCxt = NULL;
-    int32_t        code = parseSchemaClauseBottom(pCxt, pStmt, &pTableCxt);
+    STableDataCxt*   pTableCxt = NULL;
+    int32_t          code = parseSchemaClauseBottom(pCxt, pStmt, &pTableCxt);
     SRowsDataContext rowsDataCxt;
     rowsDataCxt.pTableDataCxt = pTableCxt;
     if (TSDB_CODE_SUCCESS == code) {
@@ -2442,7 +2442,7 @@ static int32_t parseInsertBodyBottom(SInsertParseContext* pCxt, SVnodeModifyOpSt
 
   bool fileOnly = (pStmt->insertType == TSDB_QUERY_TYPE_FILE_INSERT);
   if (fileOnly) {
-    // none data, skip merge & buildvgdata 
+    // none data, skip merge & buildvgdata
     if (0 == taosHashGetSize(pStmt->pTableCxtHashObj)) {
       pCxt->needRequest = false;
       return TSDB_CODE_SUCCESS;
@@ -2503,7 +2503,7 @@ static int32_t createVnodeModifOpStmt(SInsertParseContext* pCxt, bool reentry, S
   pStmt->freeHashFunc = insDestroyTableDataCxtHashMap;
   pStmt->freeArrayFunc = insDestroyVgroupDataCxtList;
   pStmt->freeStbRowsCxtFunc = destroyStbRowsDataContext;
-  
+
   if (!reentry) {
     pStmt->pVgroupsHashObj = taosHashInit(128, taosGetDefaultHashFunction(TSDB_DATA_TYPE_INT), true, HASH_NO_LOCK);
     pStmt->pTableBlockHashObj =
@@ -2625,7 +2625,7 @@ static int32_t checkSubtablePrivilegeForTable(const SArray* pTables, SVnodeModif
 }
 
 static int32_t processTableSchemaFromMetaData(SInsertParseContext* pCxt, const SMetaData* pMetaData,
-                                          SVnodeModifyOpStmt* pStmt, bool isStb) {
+                                              SVnodeModifyOpStmt* pStmt, bool isStb) {
   int32_t code = TSDB_CODE_SUCCESS;
   if (!isStb && TSDB_SUPER_TABLE == pStmt->pTableMeta->tableType) {
     code = buildInvalidOperationMsg(&pCxt->msg, "insert data into super table is not supported");
@@ -2767,11 +2767,11 @@ static int32_t parseInsertSqlFromStart(SInsertParseContext* pCxt, SVnodeModifyOp
 }
 
 static int32_t parseInsertSqlFromCsv(SInsertParseContext* pCxt, SVnodeModifyOpStmt* pStmt) {
-  int32_t code = TSDB_CODE_SUCCESS;
+  int32_t          code = TSDB_CODE_SUCCESS;
   SRowsDataContext rowsDataCxt;
 
   if (!pStmt->stbSyntax) {
-    STableDataCxt*   pTableCxt = NULL;
+    STableDataCxt* pTableCxt = NULL;
     code = getTableDataCxt(pCxt, pStmt, &pTableCxt);
     rowsDataCxt.pTableDataCxt = pTableCxt;
   } else {
diff --git a/source/libs/parser/src/parInsertUtil.c b/source/libs/parser/src/parInsertUtil.c
index 3f87f79301..7e99867b2a 100644
--- a/source/libs/parser/src/parInsertUtil.c
+++ b/source/libs/parser/src/parInsertUtil.c
@@ -170,9 +170,7 @@ static void initColValues(STableMeta* pTableMeta, SArray* pValues) {
   }
 }
 
-void insInitColValues(STableMeta* pTableMeta, SArray* aColValues) {
-    initColValues(pTableMeta, aColValues);
-}
+void insInitColValues(STableMeta* pTableMeta, SArray* aColValues) { initColValues(pTableMeta, aColValues); }
 
 int32_t insInitBoundColsInfo(int32_t numOfBound, SBoundColInfo* pInfo) {
   pInfo->numOfCols = numOfBound;
@@ -187,21 +185,22 @@ int32_t insInitBoundColsInfo(int32_t numOfBound, SBoundColInfo* pInfo) {
   return TSDB_CODE_SUCCESS;
 }
 
-void insCheckTableDataOrder(STableDataCxt* pTableCxt, TSKEY tsKey) {
+void insCheckTableDataOrder(STableDataCxt* pTableCxt, SRowKey* rowKey) {
   // once the data block is disordered, we do NOT keep last timestamp any more
   if (!pTableCxt->ordered) {
     return;
   }
 
-  if (tsKey < pTableCxt->lastTs) {
+  if (tRowKeyCompare(rowKey, &pTableCxt->lastKey) < 0) {
     pTableCxt->ordered = false;
   }
 
-  if (tsKey == pTableCxt->lastTs) {
+  if (tRowKeyCompare(rowKey, &pTableCxt->lastKey) == 0) {
     pTableCxt->duplicateTs = true;
   }
 
-  pTableCxt->lastTs = tsKey;
+  // TODO: for variable length data type, we need to copy it out
+  pTableCxt->lastKey = *rowKey;
   return;
 }
 
@@ -217,7 +216,7 @@ static int32_t createTableDataCxt(STableMeta* pTableMeta, SVCreateTbReq** pCreat
 
   int32_t code = TSDB_CODE_SUCCESS;
 
-  pTableCxt->lastTs = 0;
+  pTableCxt->lastKey = (SRowKey){0};
   pTableCxt->ordered = true;
   pTableCxt->duplicateTs = false;
 
@@ -254,7 +253,7 @@ static int32_t createTableDataCxt(STableMeta* pTableMeta, SVCreateTbReq** pCreat
       pTableCxt->pData->uid = pTableMeta->uid;
       pTableCxt->pData->sver = pTableMeta->sversion;
       pTableCxt->pData->pCreateTbReq = pCreateTbReq != NULL ? *pCreateTbReq : NULL;
-      if(pCreateTbReq != NULL) *pCreateTbReq = NULL;
+      if (pCreateTbReq != NULL) *pCreateTbReq = NULL;
       if (pTableCxt->pData->flags & SUBMIT_REQ_COLUMN_DATA_FORMAT) {
         pTableCxt->pData->aCol = taosArrayInit(128, sizeof(SColData));
         if (NULL == pTableCxt->pData->aCol) {
@@ -276,7 +275,7 @@ static int32_t createTableDataCxt(STableMeta* pTableMeta, SVCreateTbReq** pCreat
 }
 
 static int32_t rebuildTableData(SSubmitTbData* pSrc, SSubmitTbData** pDst) {
-  int32_t code = TSDB_CODE_SUCCESS;
+  int32_t        code = TSDB_CODE_SUCCESS;
   SSubmitTbData* pTmp = taosMemoryCalloc(1, sizeof(SSubmitTbData));
   if (NULL == pTmp) {
     code = TSDB_CODE_OUT_OF_MEMORY;
@@ -313,11 +312,10 @@ static int32_t rebuildTableData(SSubmitTbData* pSrc, SSubmitTbData** pDst) {
   if (TSDB_CODE_SUCCESS == code) {
     *pDst = pTmp;
   }
-  
+
   return code;
 }
 
-
 static void resetColValues(SArray* pValues) {
   int32_t num = taosArrayGetSize(pValues);
   for (int32_t i = 0; i < num; ++i) {
@@ -463,7 +461,7 @@ static int32_t createVgroupDataCxt(STableDataCxt* pTableCxt, SHashObj* pVgroupHa
   int32_t code = taosHashPut(pVgroupHash, &pVgCxt->vgId, sizeof(pVgCxt->vgId), &pVgCxt, POINTER_BYTES);
   if (TSDB_CODE_SUCCESS == code) {
     taosArrayPush(pVgroupList, &pVgCxt);
-//    uDebug("td23101 2vgId:%d, uid:%" PRIu64, pVgCxt->vgId, pTableCxt->pMeta->uid);
+    //    uDebug("td23101 2vgId:%d, uid:%" PRIu64, pVgCxt->vgId, pTableCxt->pMeta->uid);
     *pOutput = pVgCxt;
   } else {
     insDestroyVgroupDataCxt(pVgCxt);
@@ -613,7 +611,7 @@ int32_t insBuildVgDataBlocks(SHashObj* pVgroupsHashObj, SArray* pVgDataCxtList,
     if (TSDB_CODE_SUCCESS == code) {
       dst->numOfTables = taosArrayGetSize(src->pData->aSubmitTbData);
       code = taosHashGetDup(pVgroupsHashObj, (const char*)&src->vgId, sizeof(src->vgId), &dst->vg);
-//      uError("td23101 3vgId:%d, numEps:%d", src->vgId, dst->vg.epSet.numOfEps);
+      //      uError("td23101 3vgId:%d, numEps:%d", src->vgId, dst->vg.epSet.numOfEps);
     }
     if (TSDB_CODE_SUCCESS == code) {
       code = buildSubmitReq(src->vgId, src->pData, &dst->pData, &dst->size);
@@ -634,7 +632,7 @@ int32_t insBuildVgDataBlocks(SHashObj* pVgroupsHashObj, SArray* pVgDataCxtList,
 
 static bool findFileds(SSchema* pSchema, TAOS_FIELD* fields, int numFields) {
   for (int i = 0; i < numFields; i++) {
-    if(strcmp(pSchema->name, fields[i].name) == 0){
+    if (strcmp(pSchema->name, fields[i].name) == 0) {
       return true;
     }
   }
@@ -644,7 +642,8 @@ static bool findFileds(SSchema* pSchema, TAOS_FIELD* fields, int numFields) {
 
 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));
+  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, false);
@@ -654,7 +653,7 @@ int rawBlockBindData(SQuery* query, STableMeta* pTableMeta, void* data, SVCreate
   }
 
   pTableCxt->pData->flags |= TD_REQ_FROM_TAOX;
-  if(tmp == NULL){
+  if (tmp == NULL) {
     ret = initTableColSubmitData(pTableCxt);
     if (ret != TSDB_CODE_SUCCESS) {
       uError("initTableColSubmitData error");
@@ -663,8 +662,8 @@ int rawBlockBindData(SQuery* query, STableMeta* pTableMeta, void* data, SVCreate
   }
 
   char* p = (char*)data;
-  // | version | total length | total rows | blankFill | total columns | flag seg| block group id | column schema | each column
-  // length |
+  // | version | total length | total rows | blankFill | total columns | flag seg| block group id | column schema | each
+  // column length |
   int32_t version = *(int32_t*)data;
   p += sizeof(int32_t);
   p += sizeof(int32_t);
@@ -699,8 +698,8 @@ int rawBlockBindData(SQuery* query, STableMeta* pTableMeta, void* data, SVCreate
     ret = TSDB_CODE_INVALID_PARA;
     goto end;
   }
-  if(tFields == NULL){
-    for (int j = 0; j < boundInfo->numOfBound; j++){
+  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) {
@@ -717,7 +716,7 @@ int rawBlockBindData(SQuery* query, STableMeta* pTableMeta, void* data, SVCreate
       }
       char* pData = pStart;
       ret = tColDataAddValueByDataBlock(pCol, pColSchema->type, pColSchema->bytes, numOfRows, offset, pData);
-      if(ret != 0){
+      if (ret != 0) {
         goto end;
       }
       fields += sizeof(int8_t) + sizeof(int32_t);
@@ -727,11 +726,11 @@ int rawBlockBindData(SQuery* query, STableMeta* pTableMeta, void* data, SVCreate
         pStart += colLength[j];
       }
     }
-  }else{
+  } 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){
+      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;
@@ -748,7 +747,7 @@ int rawBlockBindData(SQuery* query, STableMeta* pTableMeta, void* data, SVCreate
 
           SColData* pCol = taosArrayGet(pTableCxt->pData->aCol, j);
           ret = tColDataAddValueByDataBlock(pCol, pColSchema->type, pColSchema->bytes, numOfRows, offset, pData);
-          if(ret != 0){
+          if (ret != 0) {
             goto end;
           }
           fields += sizeof(int8_t) + sizeof(int32_t);
@@ -761,17 +760,16 @@ int rawBlockBindData(SQuery* query, STableMeta* pTableMeta, void* data, SVCreate
           break;
         }
       }
-
     }
 
     for (int c = 0; c < boundInfo->numOfBound; ++c) {
-      if( boundInfo->pColIndex[c] != -1){
+      if (boundInfo->pColIndex[c] != -1) {
         SColData* pCol = taosArrayGet(pTableCxt->pData->aCol, c);
         ret = tColDataAddValueByDataBlock(pCol, 0, 0, numOfRows, NULL, NULL);
-        if(ret != 0){
+        if (ret != 0) {
           goto end;
         }
-      }else{
+      } else {
         boundInfo->pColIndex[c] = c;  // restore for next block
       }
     }

From 47410db46aab6548de286368b1477addf95a6407 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Tue, 26 Mar 2024 16:30:53 +0800
Subject: [PATCH 19/32] fix: set hasPk and pkBytes for partial func and merge
 func

---
 source/libs/function/src/builtinsimpl.c | 10 ++++++++++
 source/libs/function/src/functionMgt.c  |  8 ++++++--
 2 files changed, 16 insertions(+), 2 deletions(-)

diff --git a/source/libs/function/src/builtinsimpl.c b/source/libs/function/src/builtinsimpl.c
index 70fe2b19df..b63fa6a167 100644
--- a/source/libs/function/src/builtinsimpl.c
+++ b/source/libs/function/src/builtinsimpl.c
@@ -2343,6 +2343,11 @@ EFuncDataRequired firstDynDataReq(void* pRes, SDataBlockInfo* pBlockInfo) {
 
   SFirstLastRes* pResult = GET_ROWCELL_INTERBUF(pEntry);
   if (pResult->hasResult) {
+    if (pResult->pkBytes > 0) {
+      pResult->pkData = pResult->buf + pResult->bytes;
+    } else {
+      pResult->pkData = NULL;
+    }    
     if (pResult->ts < pBlockInfo->window.skey) {
       return FUNC_DATA_REQUIRED_NOT_LOAD;
     } else if (pResult->ts == pBlockInfo->window.skey && pResult->pkData) {
@@ -2366,6 +2371,11 @@ EFuncDataRequired lastDynDataReq(void* pRes, SDataBlockInfo* pBlockInfo) {
 
   SFirstLastRes* pResult = GET_ROWCELL_INTERBUF(pEntry);
   if (pResult->hasResult) {
+    if (pResult->pkBytes > 0) {
+      pResult->pkData = pResult->buf + pResult->bytes;
+    } else {
+      pResult->pkData = NULL;
+    }
     if (pResult->ts > pBlockInfo->window.ekey) {
       return FUNC_DATA_REQUIRED_NOT_LOAD;
     } else if (pResult->ts == pBlockInfo->window.ekey && pResult->pkData) {
diff --git a/source/libs/function/src/functionMgt.c b/source/libs/function/src/functionMgt.c
index 697efb63fd..03bc1b4e59 100644
--- a/source/libs/function/src/functionMgt.c
+++ b/source/libs/function/src/functionMgt.c
@@ -415,6 +415,8 @@ static int32_t createPartialFunction(const SFunctionNode* pSrcFunc, SFunctionNod
   int32_t len = snprintf(name, sizeof(name) - 1, "%s.%p", (*pPartialFunc)->functionName, pSrcFunc);
   taosCreateMD5Hash(name, len);
   strncpy((*pPartialFunc)->node.aliasName, name, TSDB_COL_NAME_LEN - 1);
+  (*pPartialFunc)->hasPk = pSrcFunc->hasPk;
+  (*pPartialFunc)->pkBytes = pSrcFunc->pkBytes;
   return TSDB_CODE_SUCCESS;
 }
 
@@ -453,7 +455,8 @@ static int32_t createMidFunction(const SFunctionNode* pSrcFunc, const SFunctionN
   } else {
     nodesDestroyList(pParameterList);
   }
-
+  (*pMidFunc)->hasPk = pPartialFunc->hasPk;
+  (*pMidFunc)->pkBytes = pPartialFunc->pkBytes;
   return code;
 }
 
@@ -482,7 +485,8 @@ static int32_t createMergeFunction(const SFunctionNode* pSrcFunc, const SFunctio
   } else {
     nodesDestroyList(pParameterList);
   }
-
+  (*pMergeFunc)->hasPk = pPartialFunc->hasPk;
+  (*pMergeFunc)->pkBytes = pPartialFunc->pkBytes;
   return code;
 }
 

From 595ae2c5b811bf3526ea8c975290c8a41dab582b Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Tue, 26 Mar 2024 17:05:40 +0800
Subject: [PATCH 20/32] fix: irate on super table merge phase and translate

---
 source/libs/function/src/builtins.c     |  2 +-
 source/libs/function/src/builtinsimpl.c | 21 +++++++++++++--------
 2 files changed, 14 insertions(+), 9 deletions(-)

diff --git a/source/libs/function/src/builtins.c b/source/libs/function/src/builtins.c
index 0f88572dbc..2476449383 100644
--- a/source/libs/function/src/builtins.c
+++ b/source/libs/function/src/builtins.c
@@ -1632,7 +1632,7 @@ static int32_t translateIrate(SFunctionNode* pFunc, char* pErrBuf, int32_t len)
 static int32_t translateIrateImpl(SFunctionNode* pFunc, char* pErrBuf, int32_t len, bool isPartial) {
   uint8_t colType = getSDataTypeFromNode(nodesListGetNode(pFunc->pParameterList, 0))->type;
   if (isPartial) {
-    if (3 != LIST_LENGTH(pFunc->pParameterList)) {
+    if (3 != LIST_LENGTH(pFunc->pParameterList) && 4 != LIST_LENGTH(pFunc->pParameterList)) {
       return invaildFuncParaNumErrMsg(pErrBuf, len, pFunc->functionName);
     }
     if (!IS_NUMERIC_TYPE(colType)) {
diff --git a/source/libs/function/src/builtinsimpl.c b/source/libs/function/src/builtinsimpl.c
index b63fa6a167..c029a809eb 100644
--- a/source/libs/function/src/builtinsimpl.c
+++ b/source/libs/function/src/builtinsimpl.c
@@ -6223,12 +6223,17 @@ static void doSaveRateInfo(SRateInfo* pRateInfo, bool isFirst, int64_t ts, char*
   }
 }
 
-static void initializeRateInfo(SqlFunctionCtx* pCtx, SRateInfo* pRateInfo) {
+static void initializeRateInfo(SqlFunctionCtx* pCtx, SRateInfo* pRateInfo, bool isMerge) {
   if (pCtx->hasPrimaryKey) {
-    pRateInfo->pkType = pCtx->input.pPrimaryKey->info.type;
-    pRateInfo->pkBytes = pCtx->input.pPrimaryKey->info.bytes;
-    pRateInfo->firstPk = pRateInfo->pkData;
-    pRateInfo->lastPk = pRateInfo->pkData + pRateInfo->pkBytes;
+    if (!isMerge) {
+      pRateInfo->pkType = pCtx->input.pPrimaryKey->info.type;
+      pRateInfo->pkBytes = pCtx->input.pPrimaryKey->info.bytes;
+      pRateInfo->firstPk = pRateInfo->pkData;
+      pRateInfo->lastPk = pRateInfo->pkData + pRateInfo->pkBytes;
+    } else {
+      pRateInfo->firstPk = pRateInfo->pkData;
+      pRateInfo->lastPk = pRateInfo->pkData + pRateInfo->pkBytes;
+    }
   } else {
     pRateInfo->firstPk = NULL;
     pRateInfo->lastPk = NULL;
@@ -6246,7 +6251,7 @@ int32_t irateFunction(SqlFunctionCtx* pCtx) {
 
   funcInputUpdate(pCtx);
   
-  initializeRateInfo(pCtx, pRateInfo);
+  initializeRateInfo(pCtx, pRateInfo, false);
 
   int32_t numOfElems = 0;
   int32_t type = pInputCol->info.type;
@@ -6368,13 +6373,13 @@ int32_t irateFunctionMerge(SqlFunctionCtx* pCtx) {
   }
 
   SRateInfo* pInfo = GET_ROWCELL_INTERBUF(GET_RES_INFO(pCtx));
-  initializeRateInfo(pCtx, pInfo);
+  initializeRateInfo(pCtx, pInfo, true);
 
   int32_t start = pInput->startRowIndex;
   for (int32_t i = start; i < start + pInput->numOfRows; ++i) {
     char*        data = colDataGetData(pCol, i);
     SRateInfo*   pInputInfo = (SRateInfo*)varDataVal(data);
-    initializeRateInfo(pCtx, pInfo);
+    initializeRateInfo(pCtx, pInfo, true);
     if (pInputInfo->hasResult) {
       int32_t code = irateTransferInfo(pInputInfo, pInfo);
       if (code != TSDB_CODE_SUCCESS) {

From f61596706d5345ff44bbf33096601f036ce2541c Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Tue, 26 Mar 2024 17:34:30 +0800
Subject: [PATCH 21/32] fix(tsdb): add attributes for cache scan optr.

---
 include/libs/executor/storageapi.h           |  2 +-
 source/dnode/vnode/inc/vnode.h               |  2 +-
 source/dnode/vnode/src/tsdb/tsdbCache.c      |  2 ++
 source/dnode/vnode/src/tsdb/tsdbCacheRead.c  | 16 +++++++++++++---
 source/dnode/vnode/src/tsdb/tsdbReadUtil.h   |  2 ++
 source/libs/executor/src/cachescanoperator.c | 17 +++++++++++++++--
 6 files changed, 34 insertions(+), 7 deletions(-)

diff --git a/include/libs/executor/storageapi.h b/include/libs/executor/storageapi.h
index b47a162a1a..cdc978ae08 100644
--- a/include/libs/executor/storageapi.h
+++ b/include/libs/executor/storageapi.h
@@ -191,7 +191,7 @@ typedef struct TsdReader {
 typedef struct SStoreCacheReader {
   int32_t  (*openReader)(void *pVnode, int32_t type, void *pTableIdList, int32_t numOfTables, int32_t numOfCols,
                          SArray *pCidList, int32_t *pSlotIds, uint64_t suid, void **pReader, const char *idstr,
-                         SArray *pFuncTypeList);
+                         SArray *pFuncTypeList, SColumnInfo* pPkCol, int32_t numOfPks);
   void    *(*closeReader)(void *pReader);
   int32_t  (*retrieveRows)(void *pReader, SSDataBlock *pResBlock, const int32_t *slotIds, const int32_t *dstSlotIds,
                            SArray *pTableUidList);
diff --git a/source/dnode/vnode/inc/vnode.h b/source/dnode/vnode/inc/vnode.h
index 8ae7e9706d..87a0db1e71 100644
--- a/source/dnode/vnode/inc/vnode.h
+++ b/source/dnode/vnode/inc/vnode.h
@@ -181,7 +181,7 @@ void         tsdbReaderSetNotifyCb(STsdbReader* pReader, TsdReaderNotifyCbFn not
 int32_t tsdbReuseCacherowsReader(void *pReader, void *pTableIdList, int32_t numOfTables);
 int32_t tsdbCacherowsReaderOpen(void *pVnode, int32_t type, void *pTableIdList, int32_t numOfTables, int32_t numOfCols,
                                 SArray *pCidList, int32_t *pSlotIds, uint64_t suid, void **pReader, const char *idstr,
-                                SArray* pFuncTypeList);
+                                SArray* pFuncTypeList, SColumnInfo* pkCol, int32_t numOfPks);
 int32_t tsdbRetrieveCacheRows(void *pReader, SSDataBlock *pResBlock, const int32_t *slotIds, const int32_t *dstSlotIds,
                               SArray *pTableUids);
 void   *tsdbCacherowsReaderClose(void *pReader);
diff --git a/source/dnode/vnode/src/tsdb/tsdbCache.c b/source/dnode/vnode/src/tsdb/tsdbCache.c
index 4c88242fee..c293b63f5d 100644
--- a/source/dnode/vnode/src/tsdb/tsdbCache.c
+++ b/source/dnode/vnode/src/tsdb/tsdbCache.c
@@ -2288,6 +2288,8 @@ static int32_t lastIterOpen(SFSLastIter *iter, STFileSet *pFileSet, STsdb *pTsdb
       .loadTombFn = loadSttTomb,
       .pReader = pr,
       .idstr = pr->idstr,
+      .comparFn = pr->pkComparFn,
+      .pCurRowKey = &pr->rowKey,
   };
 
   code = tMergeTreeOpen2(&iter->mergeTree, &conf, NULL);
diff --git a/source/dnode/vnode/src/tsdb/tsdbCacheRead.c b/source/dnode/vnode/src/tsdb/tsdbCacheRead.c
index b7fe92c2c5..195ca59e9a 100644
--- a/source/dnode/vnode/src/tsdb/tsdbCacheRead.c
+++ b/source/dnode/vnode/src/tsdb/tsdbCacheRead.c
@@ -209,7 +209,7 @@ int32_t tsdbReuseCacherowsReader(void* reader, void* pTableIdList, int32_t numOf
 
 int32_t tsdbCacherowsReaderOpen(void* pVnode, int32_t type, void* pTableIdList, int32_t numOfTables, int32_t numOfCols,
                                 SArray* pCidList, int32_t* pSlotIds, uint64_t suid, void** pReader, const char* idstr,
-                                SArray* pFuncTypeList) {
+                                SArray* pFuncTypeList, SColumnInfo* pPkCol, int32_t numOfPks) {
   *pReader = NULL;
   SCacheRowsReader* p = taosMemoryCalloc(1, sizeof(SCacheRowsReader));
   if (p == NULL) {
@@ -226,6 +226,15 @@ int32_t tsdbCacherowsReaderOpen(void* pVnode, int32_t type, void* pTableIdList,
   p->pSlotIds = pSlotIds;
   p->pFuncTypeList = pFuncTypeList;
 
+  p->rowKey.numOfPKs = numOfPks;
+  if (numOfPks > 0) {
+    p->pkComparFn = getComparFunc(pPkCol->type, 0);
+    p->rowKey.pks[0].type = pPkCol->type;
+    if (IS_VAR_DATA_TYPE(pPkCol->type)) {
+      p->rowKey.pks[0].pData = taosMemoryCalloc(1, pPkCol->bytes);
+    }
+  }
+
   if (numOfTables == 0) {
     *pReader = p;
     return TSDB_CODE_SUCCESS;
@@ -359,10 +368,11 @@ int32_t tsdbRetrieveCacheRows(void* pReader, SSDataBlock* pResBlock, const int32
 
   for (int32_t j = 0; j < pr->numOfCols; ++j) {
     int32_t bytes;
-    if (slotIds[j] == -1)
+    if (slotIds[j] == -1) {
       bytes = 1;
-    else
+    } else {
       bytes = pr->pSchema->columns[slotIds[j]].bytes;
+    }
 
     pRes[j] = taosMemoryCalloc(1, sizeof(SFirstLastRes) + bytes + VARSTR_HEADER_SIZE);
     SFirstLastRes* p = (SFirstLastRes*)varDataVal(pRes[j]);
diff --git a/source/dnode/vnode/src/tsdb/tsdbReadUtil.h b/source/dnode/vnode/src/tsdb/tsdbReadUtil.h
index 9cdb36e648..c48b7479dc 100644
--- a/source/dnode/vnode/src/tsdb/tsdbReadUtil.h
+++ b/source/dnode/vnode/src/tsdb/tsdbReadUtil.h
@@ -369,6 +369,8 @@ typedef struct SCacheRowsReader {
   char*                   idstr;
   int64_t                 lastTs;
   SArray*                 pFuncTypeList;
+  __compar_fn_t           pkComparFn;
+  SRowKey                 rowKey;
 } SCacheRowsReader;
 
 int32_t tsdbCacheGetBatch(STsdb* pTsdb, tb_uid_t uid, SArray* pLastArray, SCacheRowsReader* pr, int8_t ltype);
diff --git a/source/libs/executor/src/cachescanoperator.c b/source/libs/executor/src/cachescanoperator.c
index e4fa9f7580..56052434a4 100644
--- a/source/libs/executor/src/cachescanoperator.c
+++ b/source/libs/executor/src/cachescanoperator.c
@@ -46,6 +46,8 @@ typedef struct SCacheRowsScanInfo {
   int32_t         indexOfBufferedRes;
   STableListInfo* pTableList;
   SArray*         pFuncTypeList;
+  int32_t         numOfPks;
+  SColumnInfo     pkCol;
 } SCacheRowsScanInfo;
 
 static SSDataBlock* doScanCache(SOperatorInfo* pOperator);
@@ -106,6 +108,16 @@ SOperatorInfo* createCacherowsScanOperator(SLastRowScanPhysiNode* pScanNode, SRe
     goto _error;
   }
 
+  for(int32_t i = 0; i < taosArrayGetSize(pInfo->matchInfo.pList); ++i) {
+    SColMatchItem* pItem = taosArrayGet(pInfo->matchInfo.pList, i);
+    if (pItem->isPk) {
+      pInfo->numOfPks += 1;
+      pInfo->pkCol.type = pItem->dataType.type;    // only record one primary key
+      pInfo->pkCol.bytes = pItem->dataType.bytes;  // only record one primary key
+      pInfo->pkCol.pk = 1;
+    }
+  }
+
   SArray* pCidList = taosArrayInit(numOfCols, sizeof(int16_t));
   pInfo->pFuncTypeList = taosArrayInit(taosArrayGetSize(pScanNode->pFuncTypes), sizeof(int32_t));
   taosArrayAddAll(pInfo->pFuncTypeList, pScanNode->pFuncTypes);
@@ -140,7 +152,8 @@ SOperatorInfo* createCacherowsScanOperator(SLastRowScanPhysiNode* pScanNode, SRe
     uint64_t suid = tableListGetSuid(pTableListInfo);
     code = pInfo->readHandle.api.cacheFn.openReader(pInfo->readHandle.vnode, pInfo->retrieveType, pList, totalTables,
                                                     taosArrayGetSize(pInfo->matchInfo.pList), pCidList, pInfo->pSlotIds,
-                                                    suid, &pInfo->pLastrowReader, pTaskInfo->id.str, pScanNode->pFuncTypes);
+                                                    suid, &pInfo->pLastrowReader, pTaskInfo->id.str, pScanNode->pFuncTypes,
+                                                    &pInfo->pkCol, pInfo->numOfPks);
     if (code != TSDB_CODE_SUCCESS) {
       goto _error;
     }
@@ -282,7 +295,7 @@ SSDataBlock* doScanCache(SOperatorInfo* pOperator) {
       if (NULL == pInfo->pLastrowReader) {
         code = pInfo->readHandle.api.cacheFn.openReader(pInfo->readHandle.vnode, pInfo->retrieveType, pList, num,
                                        taosArrayGetSize(pInfo->matchInfo.pList), pInfo->pCidList, pInfo->pSlotIds, suid, &pInfo->pLastrowReader,
-                                       pTaskInfo->id.str, pInfo->pFuncTypeList);
+                                       pTaskInfo->id.str, pInfo->pFuncTypeList, &pInfo->pkCol, pInfo->numOfPks);
         if (code != TSDB_CODE_SUCCESS) {
           pInfo->currentGroupIndex += 1;
           taosArrayClear(pInfo->pUidList);

From 6e8bc410054b977f6029f8059db590e8975c3087 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Wed, 27 Mar 2024 09:43:41 +0800
Subject: [PATCH 22/32] feat: add test case pk func

---
 tests/parallel_test/cases.task       |   1 +
 tests/system-test/2-query/pk_func.py | 652 +++++++++++++++++++++++++++
 2 files changed, 653 insertions(+)
 create mode 100644 tests/system-test/2-query/pk_func.py

diff --git a/tests/parallel_test/cases.task b/tests/parallel_test/cases.task
index 2402f6a94a..b33f0050d1 100644
--- a/tests/parallel_test/cases.task
+++ b/tests/parallel_test/cases.task
@@ -45,6 +45,7 @@
 #,,n,system-test,python3 ./test.py -f 8-stream/snode_restart.py -N 4
 ,,n,system-test,python3 ./test.py -f 8-stream/snode_restart_with_checkpoint.py -N 4
 
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/pk_func.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/partition_expr.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/project_group.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/tbname_vgroup.py
diff --git a/tests/system-test/2-query/pk_func.py b/tests/system-test/2-query/pk_func.py
new file mode 100644
index 0000000000..30b6671f98
--- /dev/null
+++ b/tests/system-test/2-query/pk_func.py
@@ -0,0 +1,652 @@
+import sys 
+from util.log import *
+from util.cases import *
+from util.sql import *
+from util.dnodes import tdDnodes
+from math import inf
+
+class TDTestCase:
+    def caseDescription(self):
+        '''
+        case1<shenglian zhou>: [TD-] 
+        ''' 
+        return
+    
+    def init(self, conn, logSql, replicaVer=1):
+        tdLog.debug("start to execute %s" % __file__)
+        tdSql.init(conn.cursor(), True)
+        self.conn = conn
+        
+    def restartTaosd(self, index=1, dbname="db"):
+        tdDnodes.stop(index)
+        tdDnodes.startWithoutSleep(index)
+        tdSql.execute(f"use pk_func")
+
+    def run(self):
+        print("running {}".format(__file__))
+        tdSql.execute("drop database if exists pk_func")
+        tdSql.execute("create database if not exists pk_func")
+        tdSql.execute('use pk_func')
+        tdSql.execute('drop database IF EXISTS d1;')
+
+        tdSql.execute('drop database IF EXISTS d2;')
+
+        tdSql.execute('create database d1 vgroups 1')
+
+        tdSql.execute('use d1;')
+
+        tdSql.execute('create table st(ts timestamp, pk int primary key, f int) tags(t int);')
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:00', 1, 1);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:00', 2, 2);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:00', 3, 3);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:00', 4, 4);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:01', 1, 1);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:01', 4, 4);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:01', 3, 3);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:01', 2, 2);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:02', 6, 6);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:02', 5, 5);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:02', 8, 8);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:02', 7, 7);")
+
+        tdSql.query('select * from d1.ct1 order by ts,pk;')
+        tdSql.checkRows(6)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, 1)
+        tdSql.checkData(2, 2, 1)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(3, 1, 4)
+        tdSql.checkData(3, 2, 4)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 5)
+        tdSql.checkData(4, 2, 5)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(5, 1, 6)
+        tdSql.checkData(5, 2, 6)
+
+        tdSql.query('select * from d1.ct2 order by ts,pk;')
+        tdSql.checkRows(6)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 3)
+        tdSql.checkData(0, 2, 3)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 4)
+        tdSql.checkData(1, 2, 4)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, 2)
+        tdSql.checkData(2, 2, 2)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(3, 1, 3)
+        tdSql.checkData(3, 2, 3)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 7)
+        tdSql.checkData(4, 2, 7)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(5, 1, 8)
+        tdSql.checkData(5, 2, 8)
+
+        tdSql.query('select * from d1.st order by ts,pk;')
+        tdSql.checkRows(12)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(0, 3, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(1, 3, 1)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(2, 1, 3)
+        tdSql.checkData(2, 2, 3)
+        tdSql.checkData(2, 3, 2)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(3, 1, 4)
+        tdSql.checkData(3, 2, 4)
+        tdSql.checkData(3, 3, 2)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(4, 1, 1)
+        tdSql.checkData(4, 2, 1)
+        tdSql.checkData(4, 3, 1)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(5, 1, 2)
+        tdSql.checkData(5, 2, 2)
+        tdSql.checkData(5, 3, 2)
+        tdSql.checkData(6, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(6, 1, 3)
+        tdSql.checkData(6, 2, 3)
+        tdSql.checkData(6, 3, 2)
+        tdSql.checkData(7, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(7, 1, 4)
+        tdSql.checkData(7, 2, 4)
+        tdSql.checkData(7, 3, 1)
+        tdSql.checkData(8, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(8, 1, 5)
+        tdSql.checkData(8, 2, 5)
+        tdSql.checkData(8, 3, 1)
+        tdSql.checkData(9, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(9, 1, 6)
+        tdSql.checkData(9, 2, 6)
+        tdSql.checkData(9, 3, 1)
+        tdSql.checkData(10, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(10, 1, 7)
+        tdSql.checkData(10, 2, 7)
+        tdSql.checkData(10, 3, 2)
+        tdSql.checkData(11, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(11, 1, 8)
+        tdSql.checkData(11, 2, 8)
+        tdSql.checkData(11, 3, 2)
+
+        tdSql.query('select first(*) from d1.ct1;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+
+        tdSql.query('select first(*) from d1.ct2;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 3)
+        tdSql.checkData(0, 2, 3)
+
+        tdSql.query('select first(*) from d1.st;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+
+        tdSql.query('select last(*) from d1.ct1;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 6)
+        tdSql.checkData(0, 2, 6)
+
+        tdSql.query('select last(*) from d1.ct2;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 8)
+        tdSql.checkData(0, 2, 8)
+
+        tdSql.query('select last(*) from d1.st;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 8)
+        tdSql.checkData(0, 2, 8)
+
+        tdSql.query('select last_row(*) from d1.ct1;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 6)
+        tdSql.checkData(0, 2, 6)
+
+        tdSql.query('select last_row(*) from d1.ct2;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 8)
+        tdSql.checkData(0, 2, 8)
+
+        tdSql.query('select last_row(*) from d1.st;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 8)
+        tdSql.checkData(0, 2, 8)
+
+        tdSql.query('select ts,diff(f) from d1.ct1;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4)
+
+        tdSql.query('select ts,diff(f) from d1.ct2;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, -1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 5)
+
+        tdSql.query('select ts,diff(f) from d1.st;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4)
+
+        tdSql.query('select irate(f) from d1.ct1;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 4.0)
+
+        tdSql.query('select irate(f) from d1.ct2;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 5.0)
+
+        tdSql.query('select irate(f) from d1.st;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 4.0)
+
+        tdSql.query('select ts,derivative(f, 1s, 0) from d1.ct1;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0.0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4.0)
+
+        tdSql.query('select ts,derivative(f, 1s, 0) from d1.ct2;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, -1.0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 5.0)
+
+        tdSql.query('select ts,derivative(f, 1s, 0) from d1.st;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0.0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4.0)
+
+        tdSql.query('select twa(f) from d1.ct1;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 2.0)
+
+        tdSql.query('select twa(f) from d1.ct2;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 3.5)
+
+        tdSql.query('select twa(f) from d1.st;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 2.0)
+
+        tdSql.query('select ts,pk,unique(f) from d1.ct1 order by ts,pk;')
+        tdSql.checkRows(5)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, 4)
+        tdSql.checkData(2, 2, 4)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 5)
+        tdSql.checkData(3, 2, 5)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 6)
+        tdSql.checkData(4, 2, 6)
+
+        tdSql.query('select ts,pk,unique(f) from d1.ct2 order by ts,pk;')
+        tdSql.checkRows(5)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 3)
+        tdSql.checkData(0, 2, 3)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 4)
+        tdSql.checkData(1, 2, 4)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, 2)
+        tdSql.checkData(2, 2, 2)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 7)
+        tdSql.checkData(3, 2, 7)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 8)
+        tdSql.checkData(4, 2, 8)
+
+        tdSql.query('select ts,pk,unique(f) from d1.st order by ts,pk;')
+        tdSql.checkRows(8)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(2, 1, 3)
+        tdSql.checkData(2, 2, 3)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(3, 1, 4)
+        tdSql.checkData(3, 2, 4)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 5)
+        tdSql.checkData(4, 2, 5)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(5, 1, 6)
+        tdSql.checkData(5, 2, 6)
+        tdSql.checkData(6, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(6, 1, 7)
+        tdSql.checkData(6, 2, 7)
+        tdSql.checkData(7, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(7, 1, 8)
+        tdSql.checkData(7, 2, 8)
+
+        tdSql.execute('create database d2 vgroups 2')
+
+        tdSql.execute('use d2;')
+
+        tdSql.execute('create table st(ts timestamp, pk int primary key, f int) tags(t int);')
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:00', 1, 1);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:00', 2, 2);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:00', 3, 3);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:00', 4, 4);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:01', 1, 1);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:01', 4, 4);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:01', 3, 3);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:01', 2, 2);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:02', 6, 6);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:02', 5, 5);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:02', 8, 8);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:02', 7, 7);")
+
+        tdSql.query('select * from d2.ct1 order by ts,pk;')
+        tdSql.checkRows(6)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, 1)
+        tdSql.checkData(2, 2, 1)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(3, 1, 4)
+        tdSql.checkData(3, 2, 4)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 5)
+        tdSql.checkData(4, 2, 5)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(5, 1, 6)
+        tdSql.checkData(5, 2, 6)
+
+        tdSql.query('select * from d2.ct2 order by ts,pk;')
+        tdSql.checkRows(6)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 3)
+        tdSql.checkData(0, 2, 3)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 4)
+        tdSql.checkData(1, 2, 4)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, 2)
+        tdSql.checkData(2, 2, 2)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(3, 1, 3)
+        tdSql.checkData(3, 2, 3)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 7)
+        tdSql.checkData(4, 2, 7)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(5, 1, 8)
+        tdSql.checkData(5, 2, 8)
+
+        tdSql.query('select * from d2.st order by ts,pk;')
+        tdSql.checkRows(12)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(0, 3, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(1, 3, 1)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(2, 1, 3)
+        tdSql.checkData(2, 2, 3)
+        tdSql.checkData(2, 3, 2)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(3, 1, 4)
+        tdSql.checkData(3, 2, 4)
+        tdSql.checkData(3, 3, 2)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(4, 1, 1)
+        tdSql.checkData(4, 2, 1)
+        tdSql.checkData(4, 3, 1)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(5, 1, 2)
+        tdSql.checkData(5, 2, 2)
+        tdSql.checkData(5, 3, 2)
+        tdSql.checkData(6, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(6, 1, 3)
+        tdSql.checkData(6, 2, 3)
+        tdSql.checkData(6, 3, 2)
+        tdSql.checkData(7, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(7, 1, 4)
+        tdSql.checkData(7, 2, 4)
+        tdSql.checkData(7, 3, 1)
+        tdSql.checkData(8, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(8, 1, 5)
+        tdSql.checkData(8, 2, 5)
+        tdSql.checkData(8, 3, 1)
+        tdSql.checkData(9, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(9, 1, 6)
+        tdSql.checkData(9, 2, 6)
+        tdSql.checkData(9, 3, 1)
+        tdSql.checkData(10, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(10, 1, 7)
+        tdSql.checkData(10, 2, 7)
+        tdSql.checkData(10, 3, 2)
+        tdSql.checkData(11, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(11, 1, 8)
+        tdSql.checkData(11, 2, 8)
+        tdSql.checkData(11, 3, 2)
+
+        tdSql.query('select first(*) from d2.ct1;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+
+        tdSql.query('select first(*) from d2.ct2;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 3)
+        tdSql.checkData(0, 2, 3)
+
+        tdSql.query('select first(*) from d2.st;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+
+        tdSql.query('select last(*) from d2.ct1;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 6)
+        tdSql.checkData(0, 2, 6)
+
+        tdSql.query('select last(*) from d2.ct2;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 8)
+        tdSql.checkData(0, 2, 8)
+
+        tdSql.query('select last(*) from d2.st;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 8)
+        tdSql.checkData(0, 2, 8)
+
+        tdSql.query('select last_row(*) from d2.ct1;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 6)
+        tdSql.checkData(0, 2, 6)
+
+        tdSql.query('select last_row(*) from d2.ct2;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 8)
+        tdSql.checkData(0, 2, 8)
+
+        tdSql.query('select last_row(*) from d2.st;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 8)
+        tdSql.checkData(0, 2, 8)
+
+        tdSql.query('select ts,diff(f) from d2.ct1;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4)
+
+        tdSql.query('select ts,diff(f) from d2.ct2;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, -1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 5)
+
+        tdSql.query('select ts,diff(f) from d2.st;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4)
+
+        tdSql.query('select irate(f) from d2.ct1;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 4.0)
+
+        tdSql.query('select irate(f) from d2.ct2;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 5.0)
+
+        tdSql.query('select irate(f) from d2.st;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 4.0)
+
+        tdSql.query('select ts,derivative(f, 1s, 0) from d2.ct1;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0.0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4.0)
+
+        tdSql.query('select ts,derivative(f, 1s, 0) from d2.ct2;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, -1.0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 5.0)
+
+        tdSql.query('select ts,derivative(f, 1s, 0) from d2.st;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0.0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4.0)
+
+        tdSql.query('select twa(f) from d2.ct1;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 2.0)
+
+        tdSql.query('select twa(f) from d2.ct2;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 3.5)
+
+        tdSql.query('select twa(f) from d2.st;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 2.0)
+
+        tdSql.query('select ts,pk,unique(f) from d2.ct1 order by ts,pk;')
+        tdSql.checkRows(5)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, 4)
+        tdSql.checkData(2, 2, 4)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 5)
+        tdSql.checkData(3, 2, 5)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 6)
+        tdSql.checkData(4, 2, 6)
+
+        tdSql.query('select ts,pk,unique(f) from d2.ct2 order by ts,pk;')
+        tdSql.checkRows(5)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 3)
+        tdSql.checkData(0, 2, 3)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 4)
+        tdSql.checkData(1, 2, 4)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, 2)
+        tdSql.checkData(2, 2, 2)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 7)
+        tdSql.checkData(3, 2, 7)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 8)
+        tdSql.checkData(4, 2, 8)
+
+        tdSql.query('select ts,pk,unique(f) from d2.st order by ts,pk;')
+        tdSql.checkRows(8)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(2, 1, 3)
+        tdSql.checkData(2, 2, 3)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(3, 1, 4)
+        tdSql.checkData(3, 2, 4)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 5)
+        tdSql.checkData(4, 2, 5)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(5, 1, 6)
+        tdSql.checkData(5, 2, 6)
+        tdSql.checkData(6, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(6, 1, 7)
+        tdSql.checkData(6, 2, 7)
+        tdSql.checkData(7, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(7, 1, 8)
+        tdSql.checkData(7, 2, 8)
+
+        tdSql.execute('drop database pk_func')
+    def stop(self):
+        tdSql.close()
+        tdLog.success("%s successfully executed" % __file__)
+
+tdCases.addWindows(__file__, TDTestCase())
+tdCases.addLinux(__file__, TDTestCase())

From a2112625f9b62b18943e5a56ba79b3173d4df4fc Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Wed, 27 Mar 2024 15:06:51 +0800
Subject: [PATCH 23/32] fix(tsdb): set correct uid list.

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

diff --git a/source/dnode/vnode/src/tsdb/tsdbMergeTree.c b/source/dnode/vnode/src/tsdb/tsdbMergeTree.c
index 97194acb00..b3dd1dec0e 100644
--- a/source/dnode/vnode/src/tsdb/tsdbMergeTree.c
+++ b/source/dnode/vnode/src/tsdb/tsdbMergeTree.c
@@ -365,7 +365,7 @@ static int32_t loadSttStatisticsBlockData(SSttFileReader *pSttFileReader, SSttBl
       }
 
       if (pStatisBlkArray->data[k].maxTbid.suid == suid) {
-        taosArrayAddBatch(pBlockLoadInfo->info.pUid, tBufferGetDataAt(&block.suids, i * sizeof(int64_t)), rows - i);
+        taosArrayAddBatch(pBlockLoadInfo->info.pUid, tBufferGetDataAt(&block.uids, i * sizeof(int64_t)), rows - i);
         taosArrayAddBatch(pBlockLoadInfo->info.pFirstKey,
                           tBufferGetDataAt(&block.firstKeyTimestamps, i * sizeof(int64_t)), rows - i);
         taosArrayAddBatch(pBlockLoadInfo->info.pLastKey,

From 43cefb9d55bd615b0fbcecf40f79dafc1b60bd65 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Wed, 27 Mar 2024 16:26:33 +0800
Subject: [PATCH 24/32] feat: find pk col from interp func node for interp
 operator

---
 source/libs/executor/src/timesliceoperator.c | 22 ++++++++++++++++++++
 1 file changed, 22 insertions(+)

diff --git a/source/libs/executor/src/timesliceoperator.c b/source/libs/executor/src/timesliceoperator.c
index abaebb1543..4b86f85d53 100644
--- a/source/libs/executor/src/timesliceoperator.c
+++ b/source/libs/executor/src/timesliceoperator.c
@@ -46,6 +46,8 @@ typedef struct STimeSliceOperatorInfo {
   SSDataBlock*         pNextGroupRes;
   SSDataBlock*         pRemainRes;    // save block unfinished processing
   int32_t              remainIndex;     // the remaining index in the block to be processed
+  bool                 hasPk;
+  SColumn              pkCol;
 } STimeSliceOperatorInfo;
 
 static void destroyTimeSliceOperatorInfo(void* param);
@@ -953,6 +955,25 @@ _finished:
   return pResBlock->info.rows == 0 ? NULL : pResBlock;
 }
 
+static int32_t extractPkColumnFromFuncs(SNodeList* pFuncs, bool* pHasPk, SColumn* pPkColumn) {
+  SNode* pNode;
+  FOREACH(pNode, pFuncs) {
+    if ((nodeType(pNode) == QUERY_NODE_TARGET) && 
+        (nodeType(((STargetNode*)pNode)->pExpr) == QUERY_NODE_FUNCTION)) {
+        SFunctionNode* pFunc = (SFunctionNode*)((STargetNode*)pNode)->pExpr;
+        if (fmIsInterpFunc(pFunc->funcId) && pFunc->hasPk) {
+          SNode* pNode2 = (pFunc->pParameterList->pTail->pNode);
+          if ((nodeType(pNode2) == QUERY_NODE_COLUMN) && ((SColumnNode*)pNode2)->isPk) {
+            *pHasPk = true;
+            *pPkColumn = extractColumnFromColumnNode((SColumnNode*)pNode2);
+            break;
+          }
+        }
+    }
+  }
+  return TSDB_CODE_SUCCESS;
+}
+
 SOperatorInfo* createTimeSliceOperatorInfo(SOperatorInfo* downstream, SPhysiNode* pPhyNode, SExecTaskInfo* pTaskInfo) {
   STimeSliceOperatorInfo* pInfo = taosMemoryCalloc(1, sizeof(STimeSliceOperatorInfo));
   SOperatorInfo*          pOperator = taosMemoryCalloc(1, sizeof(SOperatorInfo));
@@ -985,6 +1006,7 @@ SOperatorInfo* createTimeSliceOperatorInfo(SOperatorInfo* downstream, SPhysiNode
   }
 
   pInfo->tsCol = extractColumnFromColumnNode((SColumnNode*)pInterpPhyNode->pTimeSeries);
+  extractPkColumnFromFuncs(pInterpPhyNode->pFuncs, &pInfo->hasPk, &pInfo->pkCol);
   pInfo->fillType = convertFillType(pInterpPhyNode->fillMode);
   initResultSizeInfo(&pOperator->resultInfo, 4096);
 

From 6202e1f501b444e7626115e7c86623e435737b28 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Fri, 29 Mar 2024 10:26:09 +0800
Subject: [PATCH 25/32] feat: drop pk column and update pk column is not
 supported for stable

---
 include/util/taoserror.h               |  1 +
 source/libs/parser/src/parTranslater.c | 10 ++++++++--
 source/libs/parser/src/parUtil.c       |  4 +++-
 source/util/src/terror.c               |  1 +
 4 files changed, 13 insertions(+), 3 deletions(-)

diff --git a/include/util/taoserror.h b/include/util/taoserror.h
index ba66cfb06a..affa1f0345 100644
--- a/include/util/taoserror.h
+++ b/include/util/taoserror.h
@@ -765,6 +765,7 @@ int32_t* taosGetErrno();
 #define TSDB_CODE_PAR_TAG_IS_PRIMARY_KEY        TAOS_DEF_ERROR_CODE(0, 0x2671)
 #define TSDB_CODE_PAR_SECOND_COL_PK             TAOS_DEF_ERROR_CODE(0, 0x2672)
 #define TSDB_CODE_PAR_COL_PK_TYPE               TAOS_DEF_ERROR_CODE(0, 0x2673)
+#define TSDB_CODE_PAR_INVALID_PK_OP             TAOS_DEF_ERROR_CODE(0, 0x2674)
 #define TSDB_CODE_PAR_INTERNAL_ERROR            TAOS_DEF_ERROR_CODE(0, 0x26FF)
 
 //planner
diff --git a/source/libs/parser/src/parTranslater.c b/source/libs/parser/src/parTranslater.c
index 3d59146d13..9fc9f28a6a 100644
--- a/source/libs/parser/src/parTranslater.c
+++ b/source/libs/parser/src/parTranslater.c
@@ -927,7 +927,7 @@ static bool isPrimaryKey(STempTableNode* pTable, SNode* pExpr) {
   return isPrimaryKeyImpl(pExpr);
 }
 
-static bool hasPkInTable(STableMeta* pTableMeta) {
+static bool hasPkInTable(const STableMeta* pTableMeta) {
   return pTableMeta->tableInfo.numOfColumns>=2 && pTableMeta->schema[1].flags & COL_IS_KEY;
 }
 
@@ -6752,7 +6752,13 @@ static int32_t checkAlterSuperTableBySchema(STranslateContext* pCxt, SAlterTable
               : TSDB_CODE_PAR_INVALID_TAG_NAME,
           pStmt->colName);
     }
-
+    if (hasPkInTable(pTableMeta) && (pSchema->flags & COL_IS_KEY)) {
+      if (TSDB_ALTER_TABLE_DROP_COLUMN == pStmt->alterType ||
+          TSDB_ALTER_TABLE_UPDATE_COLUMN_BYTES == pStmt->alterType ||
+          TSDB_ALTER_TABLE_UPDATE_COLUMN_NAME == pStmt->alterType) {
+            return generateSyntaxErrMsg(&pCxt->msgBuf, TSDB_CODE_PAR_INVALID_PK_OP, pStmt->colName);
+          }
+    } 
     if ((TSDB_ALTER_TABLE_UPDATE_COLUMN_BYTES == pStmt->alterType ||
          TSDB_ALTER_TABLE_UPDATE_TAG_BYTES == pStmt->alterType) &&
         (!IS_VAR_DATA_TYPE(pSchema->type) || pSchema->type != pStmt->dataType.type ||
diff --git a/source/libs/parser/src/parUtil.c b/source/libs/parser/src/parUtil.c
index 39f50da9fd..7c82555e63 100644
--- a/source/libs/parser/src/parUtil.c
+++ b/source/libs/parser/src/parUtil.c
@@ -199,7 +199,9 @@ static char* getSyntaxErrFormat(int32_t errCode) {
     case TSDB_CODE_PAR_SECOND_COL_PK:
       return "primary key column must be second column";
     case TSDB_CODE_PAR_COL_PK_TYPE:
-      return "primary key column must be of type int, uint, bigint, ubigint, and varchar";  
+      return "primary key column must be of type int, uint, bigint, ubigint, and varchar";
+    case TSDB_CODE_PAR_INVALID_PK_OP:
+      return "primary key column can not be added, modified, and dropped";    
     default:
       return "Unknown error";
   }
diff --git a/source/util/src/terror.c b/source/util/src/terror.c
index 6df9bf0e16..a64c8642db 100644
--- a/source/util/src/terror.c
+++ b/source/util/src/terror.c
@@ -627,6 +627,7 @@ TAOS_DEFINE_ERROR(TSDB_CODE_PAR_NOT_SUPPORT_MULTI_RESULT,   "Operator not suppor
 TAOS_DEFINE_ERROR(TSDB_CODE_PAR_TAG_IS_PRIMARY_KEY,         "tag can not be primary key")
 TAOS_DEFINE_ERROR(TSDB_CODE_PAR_SECOND_COL_PK,              "primary key must be second column")
 TAOS_DEFINE_ERROR(TSDB_CODE_PAR_COL_PK_TYPE,                "primary key column must be of type int, uint, bigint, ubigint, and varchar")
+TAOS_DEFINE_ERROR(TSDB_CODE_PAR_INVALID_PK_OP,              "primary key column can not be added, modified, and dropped")
 TAOS_DEFINE_ERROR(TSDB_CODE_PAR_INTERNAL_ERROR,             "Parser internal error")
 
 //planner

From 1e5ef8271a3a78dd30a88d1c0dd2b7a5184e0306 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Fri, 29 Mar 2024 10:52:28 +0800
Subject: [PATCH 26/32] feat: update column and drop column is banned for
 normal table pk column

---
 source/libs/parser/src/parTranslater.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/source/libs/parser/src/parTranslater.c b/source/libs/parser/src/parTranslater.c
index 9fc9f28a6a..b08552fc3d 100644
--- a/source/libs/parser/src/parTranslater.c
+++ b/source/libs/parser/src/parTranslater.c
@@ -10832,6 +10832,14 @@ static int32_t rewriteAlterTableImpl(STranslateContext* pCxt, SAlterTableStmt* p
     return generateSyntaxErrMsg(&pCxt->msgBuf, TSDB_CODE_PAR_INVALID_ALTER_TABLE);
   }
 
+  const SSchema* pSchema = getNormalColSchema(pTableMeta, pStmt->colName);
+  if (hasPkInTable(pTableMeta) && pSchema && (pSchema->flags & COL_IS_KEY) && 
+      (TSDB_ALTER_TABLE_DROP_COLUMN == pStmt->alterType ||
+       TSDB_ALTER_TABLE_UPDATE_COLUMN_BYTES == pStmt->alterType ||
+       TSDB_ALTER_TABLE_UPDATE_COLUMN_NAME == pStmt->alterType)) {
+    return generateSyntaxErrMsg(&pCxt->msgBuf, TSDB_CODE_PAR_INVALID_PK_OP);
+  }
+
   SVAlterTbReq req = {0};
   int32_t      code = buildAlterTbReq(pCxt, pStmt, pTableMeta, &req);
 

From fe1a7e4b9c138781185d091d9e21d899a21bd4c8 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Fri, 29 Mar 2024 11:05:45 +0800
Subject: [PATCH 27/32] feat: add pk col to scan cols of scan logic node of
 delete operation

---
 source/libs/planner/src/planLogicCreater.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/source/libs/planner/src/planLogicCreater.c b/source/libs/planner/src/planLogicCreater.c
index c79fbd01f0..b5236fee9e 100644
--- a/source/libs/planner/src/planLogicCreater.c
+++ b/source/libs/planner/src/planLogicCreater.c
@@ -1696,6 +1696,11 @@ static int32_t createDeleteScanLogicNode(SLogicPlanContext* pCxt, SDeleteStmt* p
     }
   }
 
+  STableMeta* pMeta = ((SRealTableNode*)pDelete->pFromTable)->pMeta;
+  if (TSDB_CODE_SUCCESS == code && hasPkInTable(pMeta)) {
+    code = addPkCol(pMeta->uid, pMeta->schema + 1, &pScan->pScanCols, pMeta);
+  }
+
   if (TSDB_CODE_SUCCESS == code && NULL != pDelete->pTagCond) {
     pScan->pTagCond = nodesCloneNode(pDelete->pTagCond);
     if (NULL == pScan->pTagCond) {

From b50462783061cd60df5146b83d13e1de9e540eb7 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Fri, 29 Mar 2024 17:01:56 +0800
Subject: [PATCH 28/32] fix: new group block to reset func input iter hasPrev

---
 include/libs/function/function.h        | 7 ++-----
 source/libs/executor/src/tsort.c        | 5 +++--
 source/libs/function/src/builtinsimpl.c | 7 ++++++-
 3 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/include/libs/function/function.h b/include/libs/function/function.h
index 78bd1d807b..6c60e1c4a8 100644
--- a/include/libs/function/function.h
+++ b/include/libs/function/function.h
@@ -212,11 +212,8 @@ typedef struct SFuncInputRowIter {
   char* pPrevPk;
   SSDataBlock* pPrevRowBlock; // pre one row block
 
-  //TODO:
-  // int32_t prevStartOffset; // for diff, derivative.
-  // SPoint1 prevStartPoint; // for twa.   
-  // int32_t startOffset; // for diff, derivative.
-  // SPoint1 startPoint; // for twa. 
+  uint64_t groupId;
+  bool hasGroupId;
 
   bool finalRow;
 } SFuncInputRowIter;
diff --git a/source/libs/executor/src/tsort.c b/source/libs/executor/src/tsort.c
index 271d3e05a5..44404c345e 100644
--- a/source/libs/executor/src/tsort.c
+++ b/source/libs/executor/src/tsort.c
@@ -1306,6 +1306,7 @@ static void appendToRowIndexDataBlock(SSortHandle* pHandle, SSDataBlock* pSource
 static void initRowIdSort(SSortHandle* pHandle) {
   SBlockOrderInfo* pkOrder = (pHandle->bSortPk) ? taosArrayGet(pHandle->aExtRowsOrders, 1) : NULL;
   SColumnInfoData* extPkCol = (pHandle->bSortPk) ? taosArrayGet(pHandle->pDataBlock->pDataBlock, pkOrder->slotId) : NULL;
+  SColumnInfoData pkCol = {0};
 
   SSDataBlock* pSortInput = createDataBlock();
   SColumnInfoData tsCol = createColumnInfoData(TSDB_DATA_TYPE_TIMESTAMP, 8, 1);
@@ -1317,7 +1318,7 @@ static void initRowIdSort(SSortHandle* pHandle) {
   SColumnInfoData  lengthCol = createColumnInfoData(TSDB_DATA_TYPE_INT, 4, 4);
   blockDataAppendColInfo(pSortInput, &lengthCol);
   if (pHandle->bSortPk) {
-    SColumnInfoData pkCol = createColumnInfoData(extPkCol->info.type, extPkCol->info.bytes, 5);
+    pkCol = createColumnInfoData(extPkCol->info.type, extPkCol->info.bytes, 5);
     blockDataAppendColInfo(pSortInput, &pkCol);
   }
   blockDataDestroy(pHandle->pDataBlock);
@@ -1343,7 +1344,7 @@ static void initRowIdSort(SSortHandle* pHandle) {
     biPk.order = pkOrder->order;
     biPk.slotId = 4;
     biPk.nullFirst = (biPk.order == TSDB_ORDER_ASC);
-    biPk.compFn = getKeyComparFunc(extPkCol->info.type, biPk.order);
+    biPk.compFn = getKeyComparFunc(pkCol.info.type, biPk.order);
     taosArrayPush(aOrder, &biPk);
   }
   taosArrayDestroy(pHandle->pSortInfo);
diff --git a/source/libs/function/src/builtinsimpl.c b/source/libs/function/src/builtinsimpl.c
index c029a809eb..fafc313afc 100644
--- a/source/libs/function/src/builtinsimpl.c
+++ b/source/libs/function/src/builtinsimpl.c
@@ -434,7 +434,12 @@ int32_t funcInputUpdate(SqlFunctionCtx* pCtx) {
     pIter->pPkCol = pIter->pInput->pPrimaryKey;
     pIter->rowIndex = pIter->pInput->startRowIndex;
     pIter->inputEndIndex = pIter->rowIndex + pIter->pInput->numOfRows - 1;
-    pIter->pSrcBlock = pCtx->pSrcBlock; 
+    pIter->pSrcBlock = pCtx->pSrcBlock;
+    if (!pIter->hasGroupId || pIter->groupId != pIter->pSrcBlock->info.id.groupId) {
+      pIter->hasGroupId = true;
+      pIter->groupId = pIter->pSrcBlock->info.id.groupId;
+      pIter->hasPrev = false;
+    }
   } else {
     pIter->finalRow = true;
   }

From 110ac5bd161642a3250f60c234fbcc39aca1da25 Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Fri, 29 Mar 2024 17:20:28 +0800
Subject: [PATCH 29/32] feat: add test case for pk func partition by tbname

---
 tests/parallel_test/cases.task             |   1 +
 tests/system-test/2-query/pk_func_group.py | 284 +++++++++++++++++++++
 2 files changed, 285 insertions(+)
 create mode 100644 tests/system-test/2-query/pk_func_group.py

diff --git a/tests/parallel_test/cases.task b/tests/parallel_test/cases.task
index b33f0050d1..d15b12cfe6 100644
--- a/tests/parallel_test/cases.task
+++ b/tests/parallel_test/cases.task
@@ -46,6 +46,7 @@
 ,,n,system-test,python3 ./test.py -f 8-stream/snode_restart_with_checkpoint.py -N 4
 
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/pk_func.py
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/pk_func_group.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/partition_expr.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/project_group.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/tbname_vgroup.py
diff --git a/tests/system-test/2-query/pk_func_group.py b/tests/system-test/2-query/pk_func_group.py
new file mode 100644
index 0000000000..c6b2e0f846
--- /dev/null
+++ b/tests/system-test/2-query/pk_func_group.py
@@ -0,0 +1,284 @@
+import sys 
+from util.log import *
+from util.cases import *
+from util.sql import *
+from util.dnodes import tdDnodes
+from math import inf
+
+class TDTestCase:
+    def caseDescription(self):
+        '''
+        case1<shenglian zhou>: [TD-] 
+        ''' 
+        return
+    
+    def init(self, conn, logSql, replicaVer=1):
+        tdLog.debug("start to execute %s" % __file__)
+        tdSql.init(conn.cursor(), True)
+        self.conn = conn
+        
+    def restartTaosd(self, index=1, dbname="db"):
+        tdDnodes.stop(index)
+        tdDnodes.startWithoutSleep(index)
+        tdSql.execute(f"use pk_func_group")
+
+    def run(self):
+        print("running {}".format(__file__))
+        tdSql.execute("drop database if exists pk_func_group")
+        tdSql.execute("create database if not exists pk_func_group")
+        tdSql.execute('use pk_func_group')
+        tdSql.execute('drop database IF EXISTS d1;')
+
+        tdSql.execute('drop database IF EXISTS d2;')
+
+        tdSql.execute('create database d1 vgroups 1')
+
+        tdSql.execute('use d1;')
+
+        tdSql.execute('create table st(ts timestamp, pk int primary key, f int) tags(t int);')
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:00', 1, 1);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:00', 2, 2);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:00', 3, 3);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:00', 4, 4);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:01', 1, 1);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:01', 4, 4);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:01', 3, 3);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:01', 2, 2);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:02', 6, 6);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:02', 5, 5);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:02', 8, 8);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:02', 7, 7);")
+
+        tdSql.query('select first(*) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 3)
+        tdSql.checkData(1, 2, 3)
+
+        tdSql.query('select last(*) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 6)
+        tdSql.checkData(0, 2, 6)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 8)
+        tdSql.checkData(1, 2, 8)
+
+        tdSql.query('select last_row(*) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 6)
+        tdSql.checkData(0, 2, 6)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 8)
+        tdSql.checkData(1, 2, 8)
+
+        tdSql.query('select ts,diff(f) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(4)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, -1)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 5)
+
+        tdSql.query('select irate(f) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, 4.0)
+        tdSql.checkData(1, 0, 5.0)
+
+        tdSql.query('select ts,derivative(f, 1s, 0) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(4)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0.0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4.0)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, -1.0)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 5.0)
+
+        tdSql.query('select twa(f) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, 2.0)
+        tdSql.checkData(1, 0, 3.5)
+
+        tdSql.query('select ts,pk,unique(f) from d1.st partition by tbname order by tbname,ts,pk;')
+        tdSql.checkRows(10)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, 4)
+        tdSql.checkData(2, 2, 4)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 5)
+        tdSql.checkData(3, 2, 5)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 6)
+        tdSql.checkData(4, 2, 6)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(5, 1, 3)
+        tdSql.checkData(5, 2, 3)
+        tdSql.checkData(6, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(6, 1, 4)
+        tdSql.checkData(6, 2, 4)
+        tdSql.checkData(7, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(7, 1, 2)
+        tdSql.checkData(7, 2, 2)
+        tdSql.checkData(8, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(8, 1, 7)
+        tdSql.checkData(8, 2, 7)
+        tdSql.checkData(9, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(9, 1, 8)
+        tdSql.checkData(9, 2, 8)
+
+        tdSql.execute('create database d2 vgroups 2')
+
+        tdSql.execute('use d2;')
+
+        tdSql.execute('create table st(ts timestamp, pk int primary key, f int) tags(t int);')
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:00', 1, 1);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:00', 2, 2);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:00', 3, 3);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:00', 4, 4);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:01', 1, 1);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:01', 4, 4);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:01', 3, 3);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:01', 2, 2);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:02', 6, 6);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:02', 5, 5);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:02', 8, 8);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:02', 7, 7);")
+
+        tdSql.query('select first(*) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 3)
+        tdSql.checkData(1, 2, 3)
+
+        tdSql.query('select last(*) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 6)
+        tdSql.checkData(0, 2, 6)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 8)
+        tdSql.checkData(1, 2, 8)
+
+        tdSql.query('select last_row(*) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 6)
+        tdSql.checkData(0, 2, 6)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 8)
+        tdSql.checkData(1, 2, 8)
+
+        tdSql.query('select ts,diff(f) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(4)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, -1)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 5)
+
+        tdSql.query('select irate(f) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, 4.0)
+        tdSql.checkData(1, 0, 5.0)
+
+        tdSql.query('select ts,derivative(f, 1s, 0) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(4)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0.0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4.0)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, -1.0)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 5.0)
+
+        tdSql.query('select twa(f) from d1.st partition by tbname order by tbname;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, 2.0)
+        tdSql.checkData(1, 0, 3.5)
+
+        tdSql.query('select ts,pk,unique(f) from d1.st partition by tbname order by tbname,ts,pk;')
+        tdSql.checkRows(10)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, 4)
+        tdSql.checkData(2, 2, 4)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 5)
+        tdSql.checkData(3, 2, 5)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 6)
+        tdSql.checkData(4, 2, 6)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(5, 1, 3)
+        tdSql.checkData(5, 2, 3)
+        tdSql.checkData(6, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(6, 1, 4)
+        tdSql.checkData(6, 2, 4)
+        tdSql.checkData(7, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(7, 1, 2)
+        tdSql.checkData(7, 2, 2)
+        tdSql.checkData(8, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(8, 1, 7)
+        tdSql.checkData(8, 2, 7)
+        tdSql.checkData(9, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(9, 1, 8)
+        tdSql.checkData(9, 2, 8)
+
+        tdSql.execute('drop database pk_func_group')
+    def stop(self):
+        tdSql.close()
+        tdLog.success("%s successfully executed" % __file__)
+
+tdCases.addWindows(__file__, TDTestCase())
+tdCases.addLinux(__file__, TDTestCase())

From 55bb6ab341b524d99ac71d450331e47d3f28a052 Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Fri, 29 Mar 2024 18:15:03 +0800
Subject: [PATCH 30/32] fix(tsdb): set correct initial value for compare

---
 include/common/tdataformat.h                  |  1 +
 source/common/src/tdataformat.c               | 21 ++++++
 source/dnode/vnode/src/tq/tq.c                |  1 -
 source/dnode/vnode/src/tsdb/tsdbReadUtil.c    | 30 +++++++--
 source/dnode/vnode/src/tsdb/tsdbUtil.c        | 21 ------
 source/libs/executor/src/timesliceoperator.c  | 65 +++++++++++++++----
 source/libs/executor/src/timewindowoperator.c |  5 +-
 source/util/test/tbaseCodecTest.cpp           |  5 --
 8 files changed, 101 insertions(+), 48 deletions(-)

diff --git a/include/common/tdataformat.h b/include/common/tdataformat.h
index 73e6837475..6c8a39caf6 100644
--- a/include/common/tdataformat.h
+++ b/include/common/tdataformat.h
@@ -126,6 +126,7 @@ int32_t tRowMerge(SArray *aRowP, STSchema *pTSchema, int8_t flag);
 int32_t tRowUpsertColData(SRow *pRow, STSchema *pTSchema, SColData *aColData, int32_t nColData, int32_t flag);
 void    tRowGetKey(SRow *pRow, SRowKey *key);
 int32_t tRowKeyCompare(const void *p1, const void *p2);
+int32_t tRowKeyAssign(SRowKey* pDst, SRowKey* pSrc);
 
 // SRowIter ================================
 int32_t  tRowIterOpen(SRow *pRow, STSchema *pTSchema, SRowIter **ppIter);
diff --git a/source/common/src/tdataformat.c b/source/common/src/tdataformat.c
index cf8e28bf9a..ceb572edc3 100644
--- a/source/common/src/tdataformat.c
+++ b/source/common/src/tdataformat.c
@@ -1286,6 +1286,27 @@ int32_t tRowKeyCompare(const void *p1, const void *p2) {
   return 0;
 }
 
+int32_t tRowKeyAssign(SRowKey *pDst, SRowKey* pSrc) {
+  pDst->ts = pSrc->ts;
+  pDst->numOfPKs = pSrc->numOfPKs;
+
+  if (pSrc->numOfPKs > 0) {
+    for (int32_t i = 0; i < pSrc->numOfPKs; ++i) {
+      SValue *pVal = &pDst->pks[i];
+      pVal->type = pSrc->pks[i].type;
+
+      if (IS_NUMERIC_TYPE(pVal->type)) {
+        pVal->val = pSrc->pks[i].val;
+      } else {
+        memcpy(pVal->pData, pVal->pData, pVal->nData);
+        pVal->nData = pSrc->pks[i].nData;
+      }
+    }
+  }
+
+  return TSDB_CODE_SUCCESS;
+}
+
 // STag ========================================
 static int tTagValCmprFn(const void *p1, const void *p2) {
   if (((STagVal *)p1)->cid < ((STagVal *)p2)->cid) {
diff --git a/source/dnode/vnode/src/tq/tq.c b/source/dnode/vnode/src/tq/tq.c
index ffebd783ac..141fe88339 100644
--- a/source/dnode/vnode/src/tq/tq.c
+++ b/source/dnode/vnode/src/tq/tq.c
@@ -1183,7 +1183,6 @@ int32_t tqProcessTaskCheckPointSourceReq(STQ* pTq, SRpcMsg* pMsg, SRpcMsg* pRsp)
       return TSDB_CODE_SUCCESS;
     }
   } else {
-//    ASSERT(status == TASK_STATUS__HALT);
     if (status != TASK_STATUS__HALT) {
       tqError("s-task:%s should in halt status, let's halt it directly", pTask->id.idStr);
 //      streamTaskHandleEvent(pTask->status.pSM, TASK_EVENT_HALT);
diff --git a/source/dnode/vnode/src/tsdb/tsdbReadUtil.c b/source/dnode/vnode/src/tsdb/tsdbReadUtil.c
index 99f0e42261..86c6b70c92 100644
--- a/source/dnode/vnode/src/tsdb/tsdbReadUtil.c
+++ b/source/dnode/vnode/src/tsdb/tsdbReadUtil.c
@@ -130,14 +130,29 @@ STableBlockScanInfo* getTableBlockScanInfo(SSHashObj* pTableMap, uint64_t uid, c
   return *p;
 }
 
-static int32_t initSRowKey(SRowKey* pKey, int64_t ts, int32_t numOfPks, int32_t type, int32_t len) {
+static int32_t initSRowKey(SRowKey* pKey, int64_t ts, int32_t numOfPks, int32_t type, int32_t len, bool asc) {
   pKey->numOfPKs = numOfPks;
   pKey->ts = ts;
 
   if (numOfPks > 0) {
     pKey->pks[0].type = type;
     if (IS_NUMERIC_TYPE(pKey->pks[0].type)) {
-      pKey->pks[0].val = INT64_MIN;
+      char* p = (char*)&pKey->pks[0].val;
+      if (asc) {
+        switch(pKey->pks[0].type) {
+          case TSDB_DATA_TYPE_BIGINT:*(int64_t*)p = INT64_MIN;break;
+          case TSDB_DATA_TYPE_INT:*(int32_t*)p = INT32_MIN;break;
+          case TSDB_DATA_TYPE_SMALLINT:*(int16_t*)p = INT16_MIN;break;
+          case TSDB_DATA_TYPE_TINYINT:*(int8_t*)p = INT8_MIN;break;
+        }
+      } else {
+        switch(pKey->pks[0].type) {
+          case TSDB_DATA_TYPE_BIGINT:*(int64_t*)p = INT64_MAX;break;
+          case TSDB_DATA_TYPE_INT:*(int32_t*)p = INT32_MAX;break;
+          case TSDB_DATA_TYPE_SMALLINT:*(int16_t*)p = INT16_MAX;break;
+          case TSDB_DATA_TYPE_TINYINT:*(int8_t*)p = INT8_MAX;break;
+        }
+      }
     } else {
       pKey->pks[0].pData = taosMemoryCalloc(1, len);
       pKey->pks[0].nData = 0;
@@ -154,22 +169,23 @@ static int32_t initSRowKey(SRowKey* pKey, int64_t ts, int32_t numOfPks, int32_t
 
 static void initLastProcKey(STableBlockScanInfo *pScanInfo, STsdbReader* pReader) {
   int32_t numOfPks = pReader->suppInfo.numOfPks;
+  bool asc = ASCENDING_TRAVERSE(pReader->info.order);
 
   SRowKey* pRowKey = &pScanInfo->lastProcKey;
-  if (ASCENDING_TRAVERSE(pReader->info.order)) {
+  if (asc) {
     int64_t skey = pReader->info.window.skey;
     int64_t ts = (skey > INT64_MIN) ? (skey - 1) : skey;
 
-    initSRowKey(pRowKey, ts, numOfPks, pReader->suppInfo.pk.type, pReader->suppInfo.pk.bytes);
+    initSRowKey(pRowKey, ts, numOfPks, pReader->suppInfo.pk.type, pReader->suppInfo.pk.bytes, asc);
     initSRowKey(&pScanInfo->sttKeyInfo.nextProcKey, skey, numOfPks, pReader->suppInfo.pk.type,
-                pReader->suppInfo.pk.bytes);
+                pReader->suppInfo.pk.bytes, asc);
   } else {
     int64_t ekey = pReader->info.window.ekey;
     int64_t ts = (ekey < INT64_MAX) ? (ekey + 1) : ekey;
 
-    initSRowKey(pRowKey, ts, numOfPks, pReader->suppInfo.pk.type, pReader->suppInfo.pk.bytes);
+    initSRowKey(pRowKey, ts, numOfPks, pReader->suppInfo.pk.type, pReader->suppInfo.pk.bytes, asc);
     initSRowKey(&pScanInfo->sttKeyInfo.nextProcKey, ekey, numOfPks, pReader->suppInfo.pk.type,
-                pReader->suppInfo.pk.bytes);
+                pReader->suppInfo.pk.bytes, asc);
   }
 }
 
diff --git a/source/dnode/vnode/src/tsdb/tsdbUtil.c b/source/dnode/vnode/src/tsdb/tsdbUtil.c
index 59452ebb9d..2b8de9aff3 100644
--- a/source/dnode/vnode/src/tsdb/tsdbUtil.c
+++ b/source/dnode/vnode/src/tsdb/tsdbUtil.c
@@ -628,27 +628,6 @@ void tColRowGetKey(SBlockData* pBlock, int32_t irow, SRowKey* key) {
   }
 }
 
-int32_t tRowKeyAssign(SRowKey *pDst, SRowKey* pSrc) {
-  pDst->ts = pSrc->ts;
-  pDst->numOfPKs = pSrc->numOfPKs;
-
-  if (pSrc->numOfPKs > 0) {
-    for (int32_t i = 0; i < pSrc->numOfPKs; ++i) {
-      SValue *pVal = &pDst->pks[i];
-      pVal->type = pSrc->pks[i].type;
-
-      if (IS_NUMERIC_TYPE(pVal->type)) {
-        pVal->val = pSrc->pks[i].val;
-      } else {
-        memcpy(pVal->pData, pVal->pData, pVal->nData);
-        pVal->nData = pSrc->pks[i].nData;
-      }
-    }
-  }
-
-  return TSDB_CODE_SUCCESS;
-}
-
 int32_t tsdbRowKeyCmpr(const STsdbRowKey *key1, const STsdbRowKey *key2) {
   int32_t c = tRowKeyCompare(&key1->key, &key2->key);
 
diff --git a/source/libs/executor/src/timesliceoperator.c b/source/libs/executor/src/timesliceoperator.c
index 4b86f85d53..9b28a203b8 100644
--- a/source/libs/executor/src/timesliceoperator.c
+++ b/source/libs/executor/src/timesliceoperator.c
@@ -39,7 +39,7 @@ typedef struct STimeSliceOperatorInfo {
   SColumn              tsCol;         // primary timestamp column
   SExprSupp            scalarSup;     // scalar calculation
   struct SFillColInfo* pFillColInfo;  // fill column info
-  int64_t              prevTs;
+  SRowKey              prevKey;
   bool                 prevTsSet;
   uint64_t             groupId;
   SGroupKeys*          pPrevGroupKey;
@@ -178,22 +178,49 @@ static bool isIsfilledPseudoColumn(SExprInfo* pExprInfo) {
   return (IS_BOOLEAN_TYPE(pExprInfo->base.resSchema.type) && strcasecmp(name, "_isfilled") == 0);
 }
 
+static void tRowGetKeyFromColData(int64_t ts, SColumnInfoData* pPkCol, int32_t rowIndex, SRowKey* pKey) {
+  pKey->ts = ts;
+  pKey->numOfPKs = 1;
+
+  int8_t t = pPkCol->info.type;
+
+  pKey->pks[0].type = t;
+  if (IS_NUMERIC_TYPE(t)) {
+    GET_TYPED_DATA(pKey->pks[0].val, int64_t, t, colDataGetNumData(pPkCol, rowIndex));
+  } else {
+    char* p = colDataGetVarData(pPkCol, rowIndex);
+    pKey->pks[0].pData = (uint8_t*)varDataVal(p);
+    pKey->pks[0].nData = varDataLen(p);
+  }
+}
+
 static bool checkDuplicateTimestamps(STimeSliceOperatorInfo* pSliceInfo, SColumnInfoData* pTsCol,
-                                     int32_t curIndex, int32_t rows) {
-
-
+                                     SColumnInfoData* pPkCol, int32_t curIndex, int32_t rows) {
   int64_t currentTs = *(int64_t*)colDataGetData(pTsCol, curIndex);
   if (currentTs > pSliceInfo->win.ekey) {
     return false;
   }
 
-  if ((pSliceInfo->prevTsSet == true) && (currentTs == pSliceInfo->prevTs)) {
-    return true;
+  SRowKey cur = {.ts = currentTs, .numOfPKs = (pPkCol != NULL)? 1:0};
+  if (pPkCol != NULL) {
+    cur.pks[0].type = pPkCol->info.type;
+  }
+
+  if ((pSliceInfo->prevTsSet == true) && (currentTs == pSliceInfo->prevKey.ts)) {
+//    if (pPkCol == NULL) {
+      return true;
+    /* } else {
+       tRowGetKeyFromColData(currentTs, pPkCol, curIndex, &cur);
+       if (tRowKeyCompare(&cur, &pSliceInfo->prevKey) == 0) {
+         return true;
+       }
+     }*/
   }
 
   pSliceInfo->prevTsSet = true;
-  pSliceInfo->prevTs = currentTs;
+  tRowKeyAssign(&pSliceInfo->prevKey, &cur);
 
+  // todo handle next
   if (currentTs == pSliceInfo->win.ekey && curIndex < rows - 1) {
     int64_t nextTs = *(int64_t*)colDataGetData(pTsCol, curIndex + 1);
     if (currentTs == nextTs) {
@@ -695,14 +722,20 @@ static void doTimesliceImpl(SOperatorInfo* pOperator, STimeSliceOperatorInfo* pS
   SInterval*   pInterval = &pSliceInfo->interval;
 
   SColumnInfoData* pTsCol = taosArrayGet(pBlock->pDataBlock, pSliceInfo->tsCol.slotId);
+  SColumnInfoData* pPkCol = NULL;
+
+  if (pSliceInfo->hasPk) {
+    pPkCol = taosArrayGet(pBlock->pDataBlock, pSliceInfo->pkCol.slotId);
+  }
 
   int32_t i = (pSliceInfo->pRemainRes == NULL) ? 0 : pSliceInfo->remainIndex;
   for (; i < pBlock->info.rows; ++i) {
     int64_t ts = *(int64_t*)colDataGetData(pTsCol, i);
 
     // check for duplicate timestamps
-    if (checkDuplicateTimestamps(pSliceInfo, pTsCol, i, pBlock->info.rows)) {
-      T_LONG_JMP(pTaskInfo->env, TSDB_CODE_FUNC_DUP_TIMESTAMP);
+    if (checkDuplicateTimestamps(pSliceInfo, pTsCol, pPkCol, i, pBlock->info.rows)) {
+      continue;
+//      T_LONG_JMP(pTaskInfo->env, TSDB_CODE_FUNC_DUP_TIMESTAMP);
     }
 
     if (checkNullRow(&pOperator->exprSupp, pBlock, i, ignoreNull)) {
@@ -875,11 +908,9 @@ static SSDataBlock* doTimeslice(SOperatorInfo* pOperator) {
     return NULL;
   }
 
-  SExecTaskInfo*          pTaskInfo = pOperator->pTaskInfo;
   STimeSliceOperatorInfo* pSliceInfo = pOperator->info;
   SSDataBlock*            pResBlock = pSliceInfo->pRes;
 
-  SOperatorInfo* downstream = pOperator->pDownstream[0];
   blockDataCleanup(pResBlock);
 
   while (1) {
@@ -1017,13 +1048,23 @@ SOperatorInfo* createTimeSliceOperatorInfo(SOperatorInfo* downstream, SPhysiNode
   pInfo->interval.interval = pInterpPhyNode->interval;
   pInfo->current = pInfo->win.skey;
   pInfo->prevTsSet = false;
-  pInfo->prevTs = 0;
+  pInfo->prevKey.ts = INT64_MIN;
   pInfo->groupId = 0;
   pInfo->pPrevGroupKey = NULL;
   pInfo->pNextGroupRes = NULL;
   pInfo->pRemainRes = NULL;
   pInfo->remainIndex = 0;
 
+  if (pInfo->hasPk) {
+    pInfo->prevKey.numOfPKs = 1;
+    pInfo->prevKey.ts = INT64_MIN;
+    pInfo->prevKey.pks[0].type = pInfo->pkCol.type;
+
+    if (IS_VAR_DATA_TYPE(pInfo->pkCol.type)) {
+      pInfo->prevKey.pks[0].pData = taosMemoryCalloc(1, pInfo->pkCol.bytes);
+    }
+  }
+
   if (downstream->operatorType == QUERY_NODE_PHYSICAL_PLAN_TABLE_SCAN) {
     STableScanInfo* pScanInfo = (STableScanInfo*)downstream->info;
     pScanInfo->base.cond.twindows = pInfo->win;
diff --git a/source/libs/executor/src/timewindowoperator.c b/source/libs/executor/src/timewindowoperator.c
index 57c038e75a..57cc0e82d2 100644
--- a/source/libs/executor/src/timewindowoperator.c
+++ b/source/libs/executor/src/timewindowoperator.c
@@ -658,11 +658,12 @@ static bool isCalculatedWin(SIntervalAggOperatorInfo* pInfo, const STimeWindow*
  */
 static bool filterWindowWithLimit(SIntervalAggOperatorInfo* pOperatorInfo, STimeWindow* win, uint64_t groupId) {
   if (!pOperatorInfo->limited  // if no limit info, no filter will be applied
-      || pOperatorInfo->binfo.inputTsOrder !=
-             pOperatorInfo->binfo.outputTsOrder  // if input/output ts order mismatch, no filter
+      || pOperatorInfo->binfo.inputTsOrder != pOperatorInfo->binfo.outputTsOrder
+      // if input/output ts order mismatch, no filter
   ) {
     return false;
   }
+
   if (pOperatorInfo->limit == 0) return true;
 
   if (pOperatorInfo->pBQ == NULL) {
diff --git a/source/util/test/tbaseCodecTest.cpp b/source/util/test/tbaseCodecTest.cpp
index 4c56979885..63bbfcaa68 100644
--- a/source/util/test/tbaseCodecTest.cpp
+++ b/source/util/test/tbaseCodecTest.cpp
@@ -17,11 +17,6 @@ using namespace std;
 #pragma GCC diagnostic ignored "-Wunused-variable"
 #pragma GCC diagnostic ignored "-Wsign-compare"
 
-int main(int argc, char **argv) {
-  testing::InitGoogleTest(&argc, argv);
-  return RUN_ALL_TESTS();
-}
-
 static void checkBase58Codec(uint8_t *pRaw, int32_t rawLen, int32_t index) {
   int64_t start = taosGetTimestampUs();
   char   *pEnc = base58_encode((const uint8_t *)pRaw, rawLen);

From dd2f6287ec96214176a506df136f821bad0e483b Mon Sep 17 00:00:00 2001
From: Haojun Liao <hjliao@taosdata.com>
Date: Fri, 29 Mar 2024 18:59:29 +0800
Subject: [PATCH 31/32] fix(tsdb): fix rowkey assign error.

---
 source/common/src/tdataformat.c         | 2 +-
 source/dnode/vnode/src/tsdb/tsdbRead2.c | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/source/common/src/tdataformat.c b/source/common/src/tdataformat.c
index ceb572edc3..991c17c5cc 100644
--- a/source/common/src/tdataformat.c
+++ b/source/common/src/tdataformat.c
@@ -1298,7 +1298,7 @@ int32_t tRowKeyAssign(SRowKey *pDst, SRowKey* pSrc) {
       if (IS_NUMERIC_TYPE(pVal->type)) {
         pVal->val = pSrc->pks[i].val;
       } else {
-        memcpy(pVal->pData, pVal->pData, pVal->nData);
+        memcpy(pVal->pData, pSrc->pks[i].pData, pVal->nData);
         pVal->nData = pSrc->pks[i].nData;
       }
     }
diff --git a/source/dnode/vnode/src/tsdb/tsdbRead2.c b/source/dnode/vnode/src/tsdb/tsdbRead2.c
index c7164d43de..5469eae1cd 100644
--- a/source/dnode/vnode/src/tsdb/tsdbRead2.c
+++ b/source/dnode/vnode/src/tsdb/tsdbRead2.c
@@ -2068,8 +2068,8 @@ static int32_t initMemDataIterator(STableBlockScanInfo* pBlockScanInfo, STsdbRea
     return TSDB_CODE_SUCCESS;
   }
 
-  STsdbRowKey startKey = {0};
-  tRowKeyAssign(&startKey.key, &pBlockScanInfo->lastProcKey);
+  STsdbRowKey startKey;
+  startKey.key = pBlockScanInfo->lastProcKey;
   startKey.version = asc ? pReader->info.verRange.minVer : pReader->info.verRange.maxVer;
   if ((asc && (startKey.key.ts < pWindow->skey)) || ((!asc) && startKey.key.ts > pWindow->ekey)) {
     startKey.key.ts = asc? pWindow->skey:pWindow->ekey;

From 719b8439f073d538888d2e6eab5f440f054ec00c Mon Sep 17 00:00:00 2001
From: slzhou <shenglian_zhou@163.com>
Date: Mon, 1 Apr 2024 08:58:43 +0800
Subject: [PATCH 32/32] feat: add test for pk modification

---
 tests/parallel_test/cases.task        |   1 +
 tests/system-test/2-query/pk_error.py | 267 ++++++++++++++++++++++++++
 2 files changed, 268 insertions(+)
 create mode 100644 tests/system-test/2-query/pk_error.py

diff --git a/tests/parallel_test/cases.task b/tests/parallel_test/cases.task
index d15b12cfe6..0cdac26a3a 100644
--- a/tests/parallel_test/cases.task
+++ b/tests/parallel_test/cases.task
@@ -45,6 +45,7 @@
 #,,n,system-test,python3 ./test.py -f 8-stream/snode_restart.py -N 4
 ,,n,system-test,python3 ./test.py -f 8-stream/snode_restart_with_checkpoint.py -N 4
 
+,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/pk_error.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/pk_func.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/pk_func_group.py
 ,,y,system-test,./pytest.sh python3 ./test.py -f 2-query/partition_expr.py
diff --git a/tests/system-test/2-query/pk_error.py b/tests/system-test/2-query/pk_error.py
new file mode 100644
index 0000000000..7aa1d81942
--- /dev/null
+++ b/tests/system-test/2-query/pk_error.py
@@ -0,0 +1,267 @@
+import sys 
+from util.log import *
+from util.cases import *
+from util.sql import *
+from util.dnodes import tdDnodes
+from math import inf
+
+class TDTestCase:
+    def caseDescription(self):
+        '''
+        case1<shenglian zhou>: [TD-] 
+        ''' 
+        return
+    
+    def init(self, conn, logSql, replicaVer=1):
+        tdLog.debug("start to execute %s" % __file__)
+        tdSql.init(conn.cursor(), True)
+        self.conn = conn
+        
+    def restartTaosd(self, index=1, dbname="db"):
+        tdDnodes.stop(index)
+        tdDnodes.startWithoutSleep(index)
+        tdSql.execute(f"use pk_error")
+
+    def run(self):
+        print("running {}".format(__file__))
+        tdSql.execute("drop database if exists pk_error")
+        tdSql.execute("create database if not exists pk_error")
+        tdSql.execute('use pk_error')
+        tdSql.execute('drop database IF EXISTS d1;')
+
+        tdSql.execute('drop database IF EXISTS d2;')
+
+        tdSql.execute('create database d1 vgroups 1')
+
+        tdSql.execute('use d1;')
+
+        tdSql.execute('create table st(ts timestamp, pk int primary key, f int) tags(t int);')
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:00', 1, 1);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:00', 2, 2);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:00', 3, 3);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:00', 4, 4);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:01', 1, 1);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:01', 4, 4);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:01', 3, 3);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:01', 2, 2);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:02', 6, 6);")
+
+        tdSql.execute("insert into ct1 using st tags(1) values('2021-04-19 00:00:02', 5, 5);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:02', 8, 8);")
+
+        tdSql.execute("insert into ct2 using st tags(2) values('2021-04-19 00:00:02', 7, 7);")
+
+        tdSql.query('select first(*) from d1.st partition by t order by t;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 3)
+        tdSql.checkData(1, 2, 3)
+
+        tdSql.query('select last(*) from d1.st partition by t order by t;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 6)
+        tdSql.checkData(0, 2, 6)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 8)
+        tdSql.checkData(1, 2, 8)
+
+        tdSql.query('select last_row(*) from d1.st partition by t order by t;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 6)
+        tdSql.checkData(0, 2, 6)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 8)
+        tdSql.checkData(1, 2, 8)
+
+        tdSql.query('select ts,diff(f) from d1.st partition by t order by t;')
+        tdSql.checkRows(4)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, -1)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 5)
+
+        tdSql.query('select irate(f) from d1.st partition by t order by t;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, 4.0)
+        tdSql.checkData(1, 0, 5.0)
+
+        tdSql.query('select ts,derivative(f, 1s, 0) from d1.st partition by t order by t;')
+        tdSql.checkRows(4)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0.0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4.0)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, -1.0)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 5.0)
+
+        tdSql.query('select twa(f) from d1.st partition by t order by t;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, 2.0)
+        tdSql.checkData(1, 0, 3.5)
+
+        tdSql.query('select ts,pk,unique(f) from d1.st partition by t order by t,ts,pk;')
+        tdSql.checkRows(10)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(2, 1, 4)
+        tdSql.checkData(2, 2, 4)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(3, 1, 5)
+        tdSql.checkData(3, 2, 5)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 6)
+        tdSql.checkData(4, 2, 6)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(5, 1, 3)
+        tdSql.checkData(5, 2, 3)
+        tdSql.checkData(6, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(6, 1, 4)
+        tdSql.checkData(6, 2, 4)
+        tdSql.checkData(7, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(7, 1, 2)
+        tdSql.checkData(7, 2, 2)
+        tdSql.checkData(8, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(8, 1, 7)
+        tdSql.checkData(8, 2, 7)
+        tdSql.checkData(9, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(9, 1, 8)
+        tdSql.checkData(9, 2, 8)
+
+        tdSql.execute('create table nt(ts timestamp, pk int primary key, f int);')
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:00', 1, 1);")
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:00', 2, 2);")
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:00', 3, 3);")
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:00', 4, 4);")
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:01', 1, 1);")
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:01', 4, 4);")
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:01', 3, 3);")
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:01', 2, 2);")
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:02', 6, 6);")
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:02', 5, 5);")
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:02', 8, 8);")
+
+        tdSql.execute("insert into nt values('2021-04-19 00:00:02', 7, 7);")
+
+        tdSql.query('select first(*) from d1.nt;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+
+        tdSql.query('select last(*) from d1.nt;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 8)
+        tdSql.checkData(0, 2, 8)
+
+        tdSql.query('select last_row(*) from d1.nt;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(0, 1, 8)
+        tdSql.checkData(0, 2, 8)
+
+        tdSql.query('select ts,diff(f) from d1.nt;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4)
+
+        tdSql.query('select irate(f) from d1.nt;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 4.0)
+
+        tdSql.query('select ts,derivative(f, 1s, 0) from d1.nt;')
+        tdSql.checkRows(2)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0, 1))
+        tdSql.checkData(0, 1, 0.0)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(1, 1, 4.0)
+
+        tdSql.query('select twa(f) from d1.nt;')
+        tdSql.checkRows(1)
+        tdSql.checkData(0, 0, 2.0)
+
+        tdSql.query('select ts,pk,unique(f) from d1.nt order by ts,pk;')
+        tdSql.checkRows(8)
+        tdSql.checkData(0, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(0, 1, 1)
+        tdSql.checkData(0, 2, 1)
+        tdSql.checkData(1, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(1, 1, 2)
+        tdSql.checkData(1, 2, 2)
+        tdSql.checkData(2, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(2, 1, 3)
+        tdSql.checkData(2, 2, 3)
+        tdSql.checkData(3, 0, datetime.datetime(2021, 4, 19, 0, 0))
+        tdSql.checkData(3, 1, 4)
+        tdSql.checkData(3, 2, 4)
+        tdSql.checkData(4, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(4, 1, 5)
+        tdSql.checkData(4, 2, 5)
+        tdSql.checkData(5, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(5, 1, 6)
+        tdSql.checkData(5, 2, 6)
+        tdSql.checkData(6, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(6, 1, 7)
+        tdSql.checkData(6, 2, 7)
+        tdSql.checkData(7, 0, datetime.datetime(2021, 4, 19, 0, 0, 2))
+        tdSql.checkData(7, 1, 8)
+        tdSql.checkData(7, 2, 8)
+
+        tdSql.execute('create table st2(ts timestamp, vcpk varchar(64) primary key, f int);')
+
+        tdSql.error('alter table st drop column pk;')
+
+        tdSql.error('alter table nt drop column pk;')
+
+        tdSql.error('alter table st modify column pk bigint;')
+
+        tdSql.error('alter table nt modify column pk bigint;')
+
+        tdSql.error('alter table st2 modify column vcpk varchar(32)')
+
+        tdSql.execute('drop database pk_error')
+    def stop(self):
+        tdSql.close()
+        tdLog.success("%s successfully executed" % __file__)
+
+tdCases.addWindows(__file__, TDTestCase())
+tdCases.addLinux(__file__, TDTestCase())