diff --git a/source/libs/executor/inc/executorimpl.h b/source/libs/executor/inc/executorimpl.h
index 89c3b87b27..7d33f150ff 100644
--- a/source/libs/executor/inc/executorimpl.h
+++ b/source/libs/executor/inc/executorimpl.h
@@ -718,7 +718,8 @@ void doBuildResultDatablock(SOperatorInfo* pOperator, SOptrBasicInfo* pbInfo, SG
 
 bool hasLimitOffsetInfo(SLimitInfo* pLimitInfo);
 void initLimitInfo(const SNode* pLimit, const SNode* pSLimit, SLimitInfo* pLimitInfo);
-void applyLimitOffset(SLimitInfo* pLimitInfo, SSDataBlock* pBlock, SExecTaskInfo* pTaskInfo, SOperatorInfo* pOperator);
+void resetLimitInfoForNextGroup(SLimitInfo* pLimitInfo);
+bool applyLimitOffset(SLimitInfo* pLimitInfo, SSDataBlock* pBlock, SExecTaskInfo* pTaskInfo, SOperatorInfo* pOperator);
 
 void applyAggFunctionOnPartialTuples(SExecTaskInfo* taskInfo, SqlFunctionCtx* pCtx, SColumnInfoData* pTimeWindowData,
                                      int32_t offset, int32_t forwardStep, int32_t numOfTotal, int32_t numOfOutput);
diff --git a/source/libs/executor/src/exchangeoperator.c b/source/libs/executor/src/exchangeoperator.c
index 9873c52006..037b33dc9f 100644
--- a/source/libs/executor/src/exchangeoperator.c
+++ b/source/libs/executor/src/exchangeoperator.c
@@ -584,7 +584,13 @@ int32_t doExtractResultBlocks(SExchangeInfo* pExchangeInfo, SSourceDataInfo* pDa
   int32_t index = 0;
   int32_t code = 0;
   while (index++ < pRetrieveRsp->numOfBlocks) {
-    SSDataBlock* pb = createOneDataBlock(pExchangeInfo->pDummyBlock, false);
+    SSDataBlock* pb = NULL;
+    if (taosArrayGetSize(pExchangeInfo->pRecycledBlocks) > 0) {
+      pb = *(SSDataBlock**)taosArrayPop(pExchangeInfo->pRecycledBlocks);
+      blockDataCleanup(pb);
+    } else {
+      pb = createOneDataBlock(pExchangeInfo->pDummyBlock, false);
+    }
 
     code = extractDataBlockFromFetchRsp(pb, pStart, NULL, &pStart);
     if (code != 0) {
@@ -732,9 +738,7 @@ int32_t handleLimitOffset(SOperatorInfo* pOperator, SLimitInfo* pLimitInfo, SSDa
     }
 
     // reset the value for a new group data
-    pLimitInfo->numOfOutputRows = 0;
-    pLimitInfo->remainOffset = pLimitInfo->limit.offset;
-
+    resetLimitInfoForNextGroup(pLimitInfo);
     // existing rows that belongs to previous group.
     if (pBlock->info.rows > 0) {
       return PROJECT_RETRIEVE_DONE;
@@ -760,7 +764,12 @@ int32_t handleLimitOffset(SOperatorInfo* pOperator, SLimitInfo* pLimitInfo, SSDa
     int32_t keepRows = (int32_t)(pLimitInfo->limit.limit - pLimitInfo->numOfOutputRows);
     blockDataKeepFirstNRows(pBlock, keepRows);
     if (pLimitInfo->slimit.limit > 0 && pLimitInfo->slimit.limit <= pLimitInfo->numOfOutputGroups) {
-      pOperator->status = OP_EXEC_DONE;
+      setOperatorCompleted(pOperator);
+    } else {
+      // current group limitation is reached, and future blocks of this group need to be discarded.
+      if (pBlock->info.rows == 0) {
+        return PROJECT_RETRIEVE_CONTINUE;
+      }
     }
 
     return PROJECT_RETRIEVE_DONE;
diff --git a/source/libs/executor/src/executil.c b/source/libs/executor/src/executil.c
index cbb056b3b1..1d16a3418d 100644
--- a/source/libs/executor/src/executil.c
+++ b/source/libs/executor/src/executil.c
@@ -954,7 +954,7 @@ static int32_t optimizeTbnameInCondImpl(void* metaHandle, int64_t suid, SArray*
           return -1;
         }
       } else {
-        qWarn("failed to get tableIds from by table name: %s, reason: %s", name, tstrerror(terrno));
+//        qWarn("failed to get tableIds from by table name: %s, reason: %s", name, tstrerror(terrno));
         terrno = 0;
       }
     }
@@ -1771,6 +1771,11 @@ void initLimitInfo(const SNode* pLimit, const SNode* pSLimit, SLimitInfo* pLimit
   pLimitInfo->remainGroupOffset = slimit.offset;
 }
 
+void resetLimitInfoForNextGroup(SLimitInfo* pLimitInfo) {
+  pLimitInfo->numOfOutputRows = 0;
+  pLimitInfo->remainOffset = pLimitInfo->limit.offset;
+}
+
 uint64_t tableListGetSize(const STableListInfo* pTableList) {
   ASSERT(taosArrayGetSize(pTableList->pTableList) == taosHashGetSize(pTableList->map));
   return taosArrayGetSize(pTableList->pTableList);
diff --git a/source/libs/executor/src/projectoperator.c b/source/libs/executor/src/projectoperator.c
index 73c3ac4311..6c2bcf086d 100644
--- a/source/libs/executor/src/projectoperator.c
+++ b/source/libs/executor/src/projectoperator.c
@@ -175,8 +175,7 @@ static int32_t setInfoForNewGroup(SSDataBlock* pBlock, SLimitInfo* pLimitInfo, S
 
     // reset the value for a new group data
     // existing rows that belongs to previous group.
-    pLimitInfo->numOfOutputRows = 0;
-    pLimitInfo->remainOffset = pLimitInfo->limit.offset;
+    resetLimitInfoForNextGroup(pLimitInfo);
   }
 
   return PROJECT_RETRIEVE_DONE;
@@ -200,10 +199,18 @@ static int32_t doIngroupLimitOffset(SLimitInfo* pLimitInfo, uint64_t groupId, SS
   if (pLimitInfo->limit.limit >= 0 && pLimitInfo->numOfOutputRows + pBlock->info.rows >= pLimitInfo->limit.limit) {
     int32_t keepRows = (int32_t)(pLimitInfo->limit.limit - pLimitInfo->numOfOutputRows);
     blockDataKeepFirstNRows(pBlock, keepRows);
+
     // TODO: optimize it later when partition by + limit
+    // all retrieved requirement has been fulfilled, let's finish this
     if ((pLimitInfo->slimit.limit == -1 && pLimitInfo->currentGroupId == 0) ||
         (pLimitInfo->slimit.limit > 0 && pLimitInfo->slimit.limit <= pLimitInfo->numOfOutputGroups)) {
       setOperatorCompleted(pOperator);
+    } else {
+      // Even current group is done, there may be many vgroups remain existed, and we need to continue to retrieve data
+      // from next group. So let's continue this retrieve process
+      if (keepRows == 0) {
+        return PROJECT_RETRIEVE_CONTINUE;
+      }
     }
   }
 
@@ -358,7 +365,6 @@ SSDataBlock* doProjectOperation(SOperatorInfo* pOperator) {
     pOperator->cost.openCost = (taosGetTimestampUs() - st) / 1000.0;
   }
 
-  // printDataBlock1(p, "project");
   return (p->info.rows > 0) ? p : NULL;
 }
 
diff --git a/source/libs/executor/src/scanoperator.c b/source/libs/executor/src/scanoperator.c
index fdfd81f52f..37c33c44e2 100644
--- a/source/libs/executor/src/scanoperator.c
+++ b/source/libs/executor/src/scanoperator.c
@@ -257,7 +257,7 @@ static void doSetTagColumnData(STableScanBase* pTableScanInfo, SSDataBlock* pBlo
 }
 
 // todo handle the slimit info
-void applyLimitOffset(SLimitInfo* pLimitInfo, SSDataBlock* pBlock, SExecTaskInfo* pTaskInfo, SOperatorInfo* pOperator) {
+bool applyLimitOffset(SLimitInfo* pLimitInfo, SSDataBlock* pBlock, SExecTaskInfo* pTaskInfo, SOperatorInfo* pOperator) {
   SLimit*     pLimit = &pLimitInfo->limit;
   const char* id = GET_TASKID(pTaskInfo);
 
@@ -266,6 +266,7 @@ void applyLimitOffset(SLimitInfo* pLimitInfo, SSDataBlock* pBlock, SExecTaskInfo
       pLimitInfo->remainOffset -= pBlock->info.rows;
       blockDataEmpty(pBlock);
       qDebug("current block ignore due to offset, current:%" PRId64 ", %s", pLimitInfo->remainOffset, id);
+      return false;
     } else {
       blockDataTrimFirstNRows(pBlock, pLimitInfo->remainOffset);
       pLimitInfo->remainOffset = 0;
@@ -274,13 +275,14 @@ void applyLimitOffset(SLimitInfo* pLimitInfo, SSDataBlock* pBlock, SExecTaskInfo
 
   if (pLimit->limit != -1 && pLimit->limit <= (pLimitInfo->numOfOutputRows + pBlock->info.rows)) {
     // limit the output rows
-    int32_t overflowRows = pLimitInfo->numOfOutputRows + pBlock->info.rows - pLimit->limit;
-    int32_t keep = pBlock->info.rows - overflowRows;
+    int32_t keep = (int32_t)(pLimit->limit - pLimitInfo->numOfOutputRows);
 
     blockDataKeepFirstNRows(pBlock, keep);
     qDebug("output limit %" PRId64 " has reached, %s", pLimit->limit, id);
-    pOperator->status = OP_EXEC_DONE;
+    return true;
   }
+
+  return false;
 }
 
 static int32_t loadDataBlock(SOperatorInfo* pOperator, STableScanBase* pTableScanInfo, SSDataBlock* pBlock,
@@ -395,7 +397,10 @@ static int32_t loadDataBlock(SOperatorInfo* pOperator, STableScanBase* pTableSca
     }
   }
 
-  applyLimitOffset(&pTableScanInfo->limitInfo, pBlock, pTaskInfo, pOperator);
+  bool limitReached = applyLimitOffset(&pTableScanInfo->limitInfo, pBlock, pTaskInfo, pOperator);
+  if (limitReached) { // set operator flag is done
+    setOperatorCompleted(pOperator);
+  }
 
   pCost->totalRows += pBlock->info.rows;
   pTableScanInfo->limitInfo.numOfOutputRows = pCost->totalRows;
@@ -772,8 +777,7 @@ static SSDataBlock* doTableScan(SOperatorInfo* pOperator) {
 
     // reset value for the next group data output
     pOperator->status = OP_OPENED;
-    pInfo->base.limitInfo.numOfOutputRows = 0;
-    pInfo->base.limitInfo.remainOffset = pInfo->base.limitInfo.limit.offset;
+    resetLimitInfoForNextGroup(&pInfo->base.limitInfo);
 
     int32_t        num = 0;
     STableKeyInfo* pList = NULL;
@@ -2685,9 +2689,12 @@ int32_t stopGroupTableMergeScan(SOperatorInfo* pOperator) {
   taosArrayDestroy(pInfo->queryConds);
   pInfo->queryConds = NULL;
 
+  resetLimitInfoForNextGroup(&pInfo->limitInfo);
   return TSDB_CODE_SUCCESS;
 }
 
+// all data produced by this function only belongs to one group
+// slimit/soffset does not need to be concerned here, since this function only deal with data within one group.
 SSDataBlock* getSortedTableMergeScanBlockData(SSortHandle* pHandle, SSDataBlock* pResBlock, int32_t capacity,
                                               SOperatorInfo* pOperator) {
   STableMergeScanInfo* pInfo = pOperator->info;
@@ -2707,10 +2714,12 @@ SSDataBlock* getSortedTableMergeScanBlockData(SSortHandle* pHandle, SSDataBlock*
     }
   }
 
-  qDebug("%s get sorted row blocks, rows:%d", GET_TASKID(pTaskInfo), pResBlock->info.rows);
   applyLimitOffset(&pInfo->limitInfo, pResBlock, pTaskInfo, pOperator);
   pInfo->limitInfo.numOfOutputRows += pResBlock->info.rows;
 
+  qDebug("%s get sorted row block, rows:%d, limit:%"PRId64, GET_TASKID(pTaskInfo), pResBlock->info.rows,
+      pInfo->limitInfo.numOfOutputRows);
+
   return (pResBlock->info.rows > 0) ? pResBlock : NULL;
 }
 
@@ -2749,11 +2758,13 @@ SSDataBlock* doTableMergeScan(SOperatorInfo* pOperator) {
       pOperator->resultInfo.totalRows += pBlock->info.rows;
       return pBlock;
     } else {
+      // Data of this group are all dumped, let's try the next group
       stopGroupTableMergeScan(pOperator);
       if (pInfo->tableEndIndex >= tableListSize - 1) {
         setOperatorCompleted(pOperator);
         break;
       }
+
       pInfo->tableStartIndex = pInfo->tableEndIndex + 1;
       pInfo->groupId = tableListGetInfo(pTaskInfo->pTableInfoList, pInfo->tableStartIndex)->groupId;
       startGroupTableMergeScan(pOperator);
diff --git a/source/libs/executor/src/sortoperator.c b/source/libs/executor/src/sortoperator.c
index f5dc6cc623..97b4fd9dc4 100644
--- a/source/libs/executor/src/sortoperator.c
+++ b/source/libs/executor/src/sortoperator.c
@@ -680,11 +680,13 @@ SSDataBlock* getMultiwaySortedBlockData(SSortHandle* pHandle, SSDataBlock* pData
       break;
     }
 
+    bool limitReached = applyLimitOffset(&pInfo->limitInfo, p, pTaskInfo, pOperator);
+    if (limitReached) {
+      resetLimitInfoForNextGroup(&pInfo->limitInfo);
+    }
+
     if (p->info.rows > 0) {
-      applyLimitOffset(&pInfo->limitInfo, p, pTaskInfo, pOperator);
-      if (p->info.rows > 0) {
-        break;
-      }
+      break;
     }
   }
 
@@ -698,7 +700,6 @@ SSDataBlock* getMultiwaySortedBlockData(SSortHandle* pHandle, SSDataBlock* pData
       colDataAssign(pDst, pSrc, p->info.rows, &pDataBlock->info);
     }
 
-    pInfo->limitInfo.numOfOutputRows += p->info.rows;
     pDataBlock->info.rows = p->info.rows;
     pDataBlock->info.id.groupId = pInfo->groupId;
     pDataBlock->info.dataLoad = 1;
diff --git a/source/libs/planner/src/planOptimizer.c b/source/libs/planner/src/planOptimizer.c
index b47f0bc043..18aaabf448 100644
--- a/source/libs/planner/src/planOptimizer.c
+++ b/source/libs/planner/src/planOptimizer.c
@@ -1080,29 +1080,29 @@ static bool sortPriKeyOptMayBeOptimized(SLogicNode* pNode) {
     return false;
   }
   SSortLogicNode* pSort = (SSortLogicNode*)pNode;
-  if (pSort->groupSort || !sortPriKeyOptIsPriKeyOrderBy(pSort->pSortKeys) || 1 != LIST_LENGTH(pSort->node.pChildren)) {
+  if (!sortPriKeyOptIsPriKeyOrderBy(pSort->pSortKeys) || 1 != LIST_LENGTH(pSort->node.pChildren)) {
     return false;
   }
   return true;
 }
 
-static int32_t sortPriKeyOptGetSequencingNodesImpl(SLogicNode* pNode, bool* pNotOptimize,
+static int32_t sortPriKeyOptGetSequencingNodesImpl(SLogicNode* pNode, bool groupSort, bool* pNotOptimize,
                                                    SNodeList** pSequencingNodes) {
   switch (nodeType(pNode)) {
     case QUERY_NODE_LOGIC_PLAN_SCAN: {
       SScanLogicNode* pScan = (SScanLogicNode*)pNode;
-      if (NULL != pScan->pGroupTags || TSDB_SYSTEM_TABLE == pScan->tableType) {
+      if ((!groupSort && NULL != pScan->pGroupTags) || TSDB_SYSTEM_TABLE == pScan->tableType) {
         *pNotOptimize = true;
         return TSDB_CODE_SUCCESS;
       }
       return nodesListMakeAppend(pSequencingNodes, (SNode*)pNode);
     }
     case QUERY_NODE_LOGIC_PLAN_JOIN: {
-      int32_t code = sortPriKeyOptGetSequencingNodesImpl((SLogicNode*)nodesListGetNode(pNode->pChildren, 0),
+      int32_t code = sortPriKeyOptGetSequencingNodesImpl((SLogicNode*)nodesListGetNode(pNode->pChildren, 0), groupSort,
                                                          pNotOptimize, pSequencingNodes);
       if (TSDB_CODE_SUCCESS == code) {
-        code = sortPriKeyOptGetSequencingNodesImpl((SLogicNode*)nodesListGetNode(pNode->pChildren, 1), pNotOptimize,
-                                                   pSequencingNodes);
+        code = sortPriKeyOptGetSequencingNodesImpl((SLogicNode*)nodesListGetNode(pNode->pChildren, 1), groupSort,
+                                                   pNotOptimize, pSequencingNodes);
       }
       return code;
     }
@@ -1121,13 +1121,13 @@ static int32_t sortPriKeyOptGetSequencingNodesImpl(SLogicNode* pNode, bool* pNot
     return TSDB_CODE_SUCCESS;
   }
 
-  return sortPriKeyOptGetSequencingNodesImpl((SLogicNode*)nodesListGetNode(pNode->pChildren, 0), pNotOptimize,
-                                             pSequencingNodes);
+  return sortPriKeyOptGetSequencingNodesImpl((SLogicNode*)nodesListGetNode(pNode->pChildren, 0), groupSort,
+                                             pNotOptimize, pSequencingNodes);
 }
 
-static int32_t sortPriKeyOptGetSequencingNodes(SLogicNode* pNode, SNodeList** pSequencingNodes) {
+static int32_t sortPriKeyOptGetSequencingNodes(SLogicNode* pNode, bool groupSort, SNodeList** pSequencingNodes) {
   bool    notOptimize = false;
-  int32_t code = sortPriKeyOptGetSequencingNodesImpl(pNode, &notOptimize, pSequencingNodes);
+  int32_t code = sortPriKeyOptGetSequencingNodesImpl(pNode, groupSort, &notOptimize, pSequencingNodes);
   if (TSDB_CODE_SUCCESS != code || notOptimize) {
     NODES_CLEAR_LIST(*pSequencingNodes);
   }
@@ -1175,8 +1175,8 @@ static int32_t sortPriKeyOptApply(SOptimizeContext* pCxt, SLogicSubplan* pLogicS
 
 static int32_t sortPrimaryKeyOptimizeImpl(SOptimizeContext* pCxt, SLogicSubplan* pLogicSubplan, SSortLogicNode* pSort) {
   SNodeList* pSequencingNodes = NULL;
-  int32_t    code =
-      sortPriKeyOptGetSequencingNodes((SLogicNode*)nodesListGetNode(pSort->node.pChildren, 0), &pSequencingNodes);
+  int32_t    code = sortPriKeyOptGetSequencingNodes((SLogicNode*)nodesListGetNode(pSort->node.pChildren, 0),
+                                                    pSort->groupSort, &pSequencingNodes);
   if (TSDB_CODE_SUCCESS == code && NULL != pSequencingNodes) {
     code = sortPriKeyOptApply(pCxt, pLogicSubplan, pSort, pSequencingNodes);
   }
diff --git a/tests/parallel_test/cases.task b/tests/parallel_test/cases.task
index 13cc3bce28..c28d0745c4 100644
--- a/tests/parallel_test/cases.task
+++ b/tests/parallel_test/cases.task
@@ -146,6 +146,7 @@
 ,,y,script,./test.sh -f tsim/parser/precision_ns.sim
 ,,y,script,./test.sh -f tsim/parser/projection_limit_offset.sim
 ,,y,script,./test.sh -f tsim/parser/regex.sim
+,,y,script,./test.sh -f tsim/parser/regressiontest.sim
 ,,y,script,./test.sh -f tsim/parser/select_across_vnodes.sim
 ,,y,script,./test.sh -f tsim/parser/select_distinct_tag.sim
 ,,y,script,./test.sh -f tsim/parser/select_from_cache_disk.sim