[TD-2895] refactor

src/client/src/tscLocalMerge.c

@@ -86,7 +86,6 @@ static void tscInitSqlContext(SSqlCmd *pCmd, SLocalMerger *pReducer, tOrderDescr
     pCtx->outputBytes = pExpr->resBytes;
     pCtx->outputType = pExpr->resType;
 
-    pCtx->startOffset = 0;
     pCtx->size = 1;
     pCtx->hasNull = true;
     pCtx->currentStage = MERGE_STAGE;

src/client/src/tscSQLParser.c

@@ -4737,7 +4737,7 @@ static void setDefaultOrderInfo(SQueryInfo* pQueryInfo) {
 int32_t parseOrderbyClause(SSqlCmd* pCmd, SQueryInfo* pQueryInfo, SQuerySQL* pQuerySql, SSchema* pSchema) {
   const char* msg0 = "only support order by primary timestamp";
   const char* msg1 = "invalid column name";
-  const char* msg2 = "only support order by primary timestamp or first tag in groupby clause allowed";
+  const char* msg2 = "order by primary timestamp or first tag in groupby clause allowed";
   const char* msg3 = "invalid column in order by clause, only primary timestamp or first tag in groupby clause allowed";
 
   setDefaultOrderInfo(pQueryInfo);

src/client/tests/cliTest.cpp

@@ -57,7 +57,7 @@ void stmtInsertTest() {
   v.ts = start_ts + 20;
   v.k = 123;
 
-  char* str = "abc";
+  char str[] = "abc";
   uintptr_t len = strlen(str);
 
   v.a = str;
@@ -65,7 +65,7 @@ void stmtInsertTest() {
   params[2].buffer_length = len;
   params[2].buffer = str;
 
-  char* nstr = "999";
+  char nstr[] = "999";
   uintptr_t len1 = strlen(nstr);
 
   v.b = nstr;
@@ -84,18 +84,18 @@ void stmtInsertTest() {
   v.ts = start_ts + 30;
   v.k = 911;
 
-  str = "92";
-  len = strlen(str);
+  char str1[] = "92";
+  len = strlen(str1);
 
   params[2].length = &len;
   params[2].buffer_length = len;
-  params[2].buffer = str;
+  params[2].buffer = str1;
 
-  nstr = "1920";
-  len1 = strlen(nstr);
+  char nstr1[] = "1920";
+  len1 = strlen(nstr1);
 
   params[3].buffer_length = len1;
-  params[3].buffer = nstr;
+  params[3].buffer = nstr1;
   params[3].length = &len1;
 
   taos_stmt_bind_param(stmt, params);
@@ -103,7 +103,7 @@ void stmtInsertTest() {
 
   ret = taos_stmt_execute(stmt);
   if (ret != 0) {
-    printf("%p\n", ret);
+    printf("%d\n", ret);
     printf("\033[31mfailed to execute insert statement.\033[0m\n");
     return;
   }

src/cq/src/cqMain.c

@@ -256,6 +256,7 @@ void cqStop(void *handle) {
   if (tsEnableStream == 0) {
     return;
   }
+
   SCqContext *pContext = handle;
   cDebug("vgId:%d, stop all CQs", pContext->vgId);
   if (pContext->dbConn == NULL || pContext->master == 0) return;

src/query/inc/qAggMain.h

@@ -84,7 +84,7 @@ extern "C" {
 #define TSDB_FUNCSTATE_SO           0x1u    // single output
 #define TSDB_FUNCSTATE_MO           0x2u    // dynamic number of output, not multinumber of output e.g., TOP/BOTTOM
 #define TSDB_FUNCSTATE_STREAM       0x4u    // function avail for stream
-#define TSDB_FUNCSTATE_STABLE       0x8u    // function avail for metric
+#define TSDB_FUNCSTATE_STABLE       0x8u    // function avail for super table
 #define TSDB_FUNCSTATE_OF           0x10u   // outer forward
 #define TSDB_FUNCSTATE_NEED_TS      0x20u   // timestamp is required during query processing
 #define TSDB_FUNCSTATE_SELECTIVITY  0x40u   // selectivity functions, can exists along with tag columns
@@ -166,9 +166,8 @@ typedef struct SExtTagsInfo {
 
 // sql function runtime context
 typedef struct SQLFunctionCtx {
-  int32_t      startOffset;  // todo remove it
   int32_t      size;      // number of rows
-  void *       pInput;    //
+  void *       pInput;    // input data buffer
   uint32_t     order;     // asc|desc
   int16_t      inputType;
   int16_t      inputBytes;
@@ -184,7 +183,7 @@ typedef struct SQLFunctionCtx {
   uint8_t      currentStage;  // record current running step, default: 0
   int64_t      startTs;       // timestamp range of current query when function is executed on a specific data block
   int32_t      numOfParams;
-  tVariant     param[4];      // input parameter, e.g., top(k, 20), the number of results for top query is kept in param */
+  tVariant     param[4];      // input parameter, e.g., top(k, 20), the number of results for top query is kept in param
   int64_t     *ptsList;       // corresponding timestamp array list
   void        *ptsOutputBuf;  // corresponding output buffer for timestamp of each result, e.g., top/bottom*/
   SQLPreAggVal preAggVals;
@@ -228,7 +227,7 @@ int32_t getResultDataInfo(int32_t dataType, int32_t dataBytes, int32_t functionI
 #define IS_SINGLEOUTPUT(x)        (((x)&TSDB_FUNCSTATE_SO) != 0)
 #define IS_OUTER_FORWARD(x)       (((x)&TSDB_FUNCSTATE_OF) != 0)
 
-/* determine the real data need to calculated the result */
+// determine the real data need to calculated the result
 enum {
   BLK_DATA_NO_NEEDED     = 0x0,
   BLK_DATA_STATIS_NEEDED = 0x1,

src/query/inc/qExecutor.h

@@ -33,6 +33,36 @@ struct SColumnFilterElem;
 typedef bool (*__filter_func_t)(struct SColumnFilterElem* pFilter, const char* val1, const char* val2, int16_t type);
 typedef int32_t (*__block_search_fn_t)(char* data, int32_t num, int64_t key, int32_t order);
 
+#define IS_QUERY_KILLED(_q) ((_q)->code == TSDB_CODE_TSC_QUERY_CANCELLED)
+#define Q_STATUS_EQUAL(p, s)  (((p) & (s)) != 0u)
+#define QUERY_IS_ASC_QUERY(q) (GET_FORWARD_DIRECTION_FACTOR((q)->order.order) == QUERY_ASC_FORWARD_STEP)
+
+#define SET_STABLE_QUERY_OVER(_q) ((_q)->tableIndex = (int32_t)((_q)->tableqinfoGroupInfo.numOfTables))
+#define IS_STASBLE_QUERY_OVER(_q) ((_q)->tableIndex >= (int32_t)((_q)->tableqinfoGroupInfo.numOfTables))
+
+#define GET_TABLEGROUP(q, _index)   ((SArray*) taosArrayGetP((q)->tableqinfoGroupInfo.pGroupList, (_index)))
+
+enum {
+  // when query starts to execute, this status will set
+      QUERY_NOT_COMPLETED = 0x1u,
+
+  /* result output buffer is full, current query is paused.
+   * this status is only exist in group-by clause and diff/add/division/multiply/ query.
+   */
+      QUERY_RESBUF_FULL = 0x2u,
+
+  /* query is over
+   * 1. this status is used in one row result query process, e.g., count/sum/first/last/ avg...etc.
+   * 2. when all data within queried time window, it is also denoted as query_completed
+   */
+      QUERY_COMPLETED = 0x4u,
+
+  /* when the result is not completed return to client, this status will be
+   * usually used in case of interval query with interpolation option
+   */
+      QUERY_OVER = 0x8u,
+};
+
 typedef struct SResultRowPool {
   int32_t elemSize;
   int32_t blockSize;
@@ -66,7 +96,8 @@ typedef struct SResultRow {
 } SResultRow;
 
 typedef struct SGroupResInfo {
-  int32_t rowId;
+  int32_t totalGroup;
+  int32_t currentGroup;
   int32_t index;
   SArray* pRows;      // SArray<SResultRow*>
 } SGroupResInfo;
@@ -112,7 +143,7 @@ typedef struct STableQueryInfo {
   STimeWindow win;
   STSCursor   cur;
   void*       pTable;         // for retrieve the page id list
-  SResultRowInfo windowResInfo;
+  SResultRowInfo resInfo;
 } STableQueryInfo;
 
 typedef struct SQueryCostInfo {
@@ -193,7 +224,7 @@ typedef struct SQueryRuntimeEnv {
   uint16_t*            offset;
   uint16_t             scanFlag;         // denotes reversed scan of data or not
   SFillInfo*           pFillInfo;
-  SResultRowInfo       windowResInfo;
+  SResultRowInfo       resultRowInfo;
 
   SQueryCostInfo       summary;
   void*                pQueryHandle;
@@ -257,4 +288,51 @@ typedef struct SQInfo {
   char*            sql;         // query sql string
 } SQInfo;
 
+typedef struct SQueryParam {
+  char            *sql;
+  char            *tagCond;
+  char            *tbnameCond;
+  char            *prevResult;
+  SArray          *pTableIdList;
+  SSqlFuncMsg    **pExprMsg;
+  SSqlFuncMsg    **pSecExprMsg;
+  SExprInfo       *pExprs;
+  SExprInfo       *pSecExprs;
+
+  SColIndex       *pGroupColIndex;
+  SColumnInfo     *pTagColumnInfo;
+  SSqlGroupbyExpr *pGroupbyExpr;
+} SQueryParam;
+
+void freeParam(SQueryParam *param);
+int32_t convertQueryMsg(SQueryTableMsg *pQueryMsg, SQueryParam* param);
+int32_t createQueryFuncExprFromMsg(SQueryTableMsg *pQueryMsg, int32_t numOfOutput, SExprInfo **pExprInfo, SSqlFuncMsg **pExprMsg,
+                                   SColumnInfo* pTagCols);
+SSqlGroupbyExpr *createGroupbyExprFromMsg(SQueryTableMsg *pQueryMsg, SColIndex *pColIndex, int32_t *code);
+SQInfo *createQInfoImpl(SQueryTableMsg *pQueryMsg, SSqlGroupbyExpr *pGroupbyExpr, SExprInfo *pExprs,
+                        SExprInfo *pSecExprs, STableGroupInfo *pTableGroupInfo, SColumnInfo* pTagCols, bool stableQuery, char* sql);
+int32_t initQInfo(SQueryTableMsg *pQueryMsg, void *tsdb, int32_t vgId, SQInfo *pQInfo, SQueryParam* param, bool isSTable);
+void freeColumnFilterInfo(SColumnFilterInfo* pFilter, int32_t numOfFilters);
+
+bool isQueryKilled(SQInfo *pQInfo);
+int32_t checkForQueryBuf(size_t numOfTables);
+bool doBuildResCheck(SQInfo* pQInfo);
+void setQueryStatus(SQuery *pQuery, int8_t status);
+
+bool onlyQueryTags(SQuery* pQuery);
+void buildTagQueryResult(SQInfo *pQInfo);
+void stableQueryImpl(SQInfo *pQInfo);
+void buildTableBlockDistResult(SQInfo *pQInfo);
+void tableQueryImpl(SQInfo *pQInfo);
+bool isValidQInfo(void *param);
+
+int32_t doDumpQueryResult(SQInfo *pQInfo, char *data);
+
+size_t getResultSize(SQInfo *pQInfo, int64_t *numOfRows);
+void setQueryKilled(SQInfo *pQInfo);
+void queryCostStatis(SQInfo *pQInfo);
+void freeQInfo(SQInfo *pQInfo);
+
+int32_t getMaximumIdleDurationSec();
+
 #endif  // TDENGINE_QUERYEXECUTOR_H

src/query/inc/qUtil.h

@@ -85,4 +85,12 @@ void interResToBinary(SBufferWriter* bw, SArray* pRes, int32_t tagLen);
 SArray* interResFromBinary(const char* data, int32_t len);
 void freeInterResult(void* param);
 
+void    initGroupResInfo(SGroupResInfo* pGroupResInfo, SResultRowInfo* pResultInfo, int32_t offset);
+void    cleanupGroupResInfo(SGroupResInfo* pGroupResInfo);
+bool    hasRemainData(SGroupResInfo* pGroupResInfo);
+bool    incNextGroup(SGroupResInfo* pGroupResInfo);
+int32_t getNumOfTotalRes(SGroupResInfo* pGroupResInfo);
+
+int32_t mergeIntoGroupResult(SGroupResInfo* pGroupResInfo, SQInfo *pQInfo);
+
 #endif  // TDENGINE_QUERYUTIL_H

src/query/src/qAggMain.c

@@ -26,10 +26,12 @@
 #include "qTsbuf.h"
 #include "queryLog.h"
 
-#define GET_INPUT_DATA_LIST(x) (((char *)((x)->pInput)) + ((x)->startOffset) * ((x)->inputBytes))
+//#define GET_INPUT_DATA_LIST(x) (((char *)((x)->pInput)) + ((x)->startOffset) * ((x)->inputBytes))
+#define GET_INPUT_DATA_LIST(x) ((char *)((x)->pInput))
 #define GET_INPUT_DATA(x, y) (GET_INPUT_DATA_LIST(x) + (y) * (x)->inputBytes)
 
-#define GET_TS_LIST(x)    ((TSKEY*)&((x)->ptsList[(x)->startOffset]))
+//#define GET_TS_LIST(x)    ((TSKEY*)&((x)->ptsList[(x)->startOffset]))
+#define GET_TS_LIST(x)    ((TSKEY*)((x)->ptsList))
 #define GET_TS_DATA(x, y) (GET_TS_LIST(x)[(y)])
 
 #define GET_TRUE_DATA_TYPE()                          \
@@ -379,12 +381,8 @@ static bool function_setup(SQLFunctionCtx *pCtx) {
 static void function_finalizer(SQLFunctionCtx *pCtx) {
   SResultRowCellInfo *pResInfo = GET_RES_INFO(pCtx);
   if (pResInfo->hasResult != DATA_SET_FLAG) {
-    if (pCtx->outputType == TSDB_DATA_TYPE_BINARY || pCtx->outputType == TSDB_DATA_TYPE_NCHAR) {
-      setVardataNull(pCtx->pOutput, pCtx->outputType);
-    } else {
     setNull(pCtx->pOutput, pCtx->outputType, pCtx->outputBytes);
   }
-  }
   
   doFinalizer(pCtx);
 }
@@ -414,10 +412,7 @@ static void count_function(SQLFunctionCtx *pCtx) {
         numOfElem += 1;
       }
     } else {
-      /*
-       * when counting on the primary time stamp column and no statistics data is provided,
-       * simple use the size value
-       */
+      //when counting on the primary time stamp column and no statistics data is presented, use the size value directly.
       numOfElem = pCtx->size;
     }
   }
@@ -944,9 +939,9 @@ static void minMax_function(SQLFunctionCtx *pCtx, char *pOutput, int32_t isMin,
        *
        * The following codes of 3 lines will be removed later.
        */
-      if (index < 0 || index >= pCtx->size + pCtx->startOffset) {
-        index = 0;
-      }
+//      if (index < 0 || index >= pCtx->size + pCtx->startOffset) {
+//        index = 0;
+//      }
 
       // the index is the original position, not the relative position
       key = pCtx->ptsList[index];
@@ -3487,9 +3482,7 @@ static void arithmetic_function(SQLFunctionCtx *pCtx) {
   SArithmeticSupport *sas = (SArithmeticSupport *)pCtx->param[1].pz;
   
   arithmeticTreeTraverse(sas->pArithExpr->pExpr, pCtx->size, pCtx->pOutput, sas, pCtx->order, getArithColumnData);
-  
   pCtx->pOutput += pCtx->outputBytes * pCtx->size;
-  pCtx->param[1].pz = NULL;
 }
 
 static void arithmetic_function_f(SQLFunctionCtx *pCtx, int32_t index) {
@@ -3977,6 +3970,12 @@ static void interp_function_impl(SQLFunctionCtx *pCtx) {
       } else {
         assignVal(pCtx->pOutput, pCtx->start.ptr, pCtx->outputBytes, pCtx->inputType);
       }
+    } else if (type == TSDB_FILL_NEXT) {
+      if (IS_NUMERIC_TYPE(pCtx->inputType) || pCtx->inputType == TSDB_DATA_TYPE_BOOL) {
+        SET_TYPED_DATA(pCtx->pOutput, pCtx->inputType, pCtx->end.val);
+      } else {
+        assignVal(pCtx->pOutput, pCtx->end.ptr, pCtx->outputBytes, pCtx->inputType);
+      }
     } else if (type == TSDB_FILL_LINEAR) {
       SPoint point1 = {.key = pCtx->start.key, .val = &pCtx->start.val};
       SPoint point2 = {.key = pCtx->end.key, .val = &pCtx->end.val};

src/query/src/qExecutor.c

@@ -15,7 +15,6 @@
 #include "os.h"
 #include "qFill.h"
 #include "taosmsg.h"
-#include "tcache.h"
 #include "tglobal.h"
 
 #include "exception.h"
@@ -24,11 +23,9 @@
 #include "qExecutor.h"
 #include "qResultbuf.h"
 #include "qUtil.h"
-#include "query.h"
 #include "queryLog.h"
 #include "tlosertree.h"
 #include "ttype.h"
-#include "tcompare.h"
 
 #define MAX_ROWS_PER_RESBUF_PAGE  ((1u<<12) - 1)
 
@@ -36,8 +33,6 @@
  * check if the primary column is load by default, otherwise, the program will
  * forced to load primary column explicitly.
  */
-#define Q_STATUS_EQUAL(p, s)  (((p) & (s)) != 0u)
-#define QUERY_IS_ASC_QUERY(q) (GET_FORWARD_DIRECTION_FACTOR((q)->order.order) == QUERY_ASC_FORWARD_STEP)
 
 #define IS_MASTER_SCAN(runtime)        ((runtime)->scanFlag == MASTER_SCAN)
 #define IS_REVERSE_SCAN(runtime)       ((runtime)->scanFlag == REVERSE_SCAN)
@@ -56,27 +51,6 @@
    (_dst).ekey = (_src).ekey;\
 } while (0)
 
-enum {
-  // when query starts to execute, this status will set
-  QUERY_NOT_COMPLETED = 0x1u,
-
-  /* result output buffer is full, current query is paused.
-   * this status is only exist in group-by clause and diff/add/division/multiply/ query.
-   */
-  QUERY_RESBUF_FULL = 0x2u,
-
-  /* query is over
-   * 1. this status is used in one row result query process, e.g., count/sum/first/last/ avg...etc.
-   * 2. when all data within queried time window, it is also denoted as query_completed
-   */
-  QUERY_COMPLETED = 0x4u,
-
-  /* when the result is not completed return to client, this status will be
-   * usually used in case of interval query with interpolation option
-   */
-  QUERY_OVER = 0x8u,
-};
-
 enum {
   TS_JOIN_TS_EQUAL       = 0,
   TS_JOIN_TS_NOT_EQUALS  = 1,
@@ -134,13 +108,11 @@ static UNUSED_FUNC void* u_realloc(void* p, size_t __size) {
 
 #define CLEAR_QUERY_STATUS(q, st)   ((q)->status &= (~(st)))
 #define GET_NUM_OF_TABLEGROUP(q)    taosArrayGetSize((q)->tableqinfoGroupInfo.pGroupList)
-#define GET_TABLEGROUP(q, _index)   ((SArray*) taosArrayGetP((q)->tableqinfoGroupInfo.pGroupList, (_index)))
 #define QUERY_IS_INTERVAL_QUERY(_q) ((_q)->interval.interval > 0)
 
-static void setQueryStatus(SQuery *pQuery, int8_t status);
 static void finalizeQueryResult(SQueryRuntimeEnv *pRuntimeEnv);
 
-static int32_t getMaximumIdleDurationSec() {
+int32_t getMaximumIdleDurationSec() {
   return tsShellActivityTimer * 2;
 }
 
@@ -181,27 +153,19 @@ static void getNextTimeWindow(SQuery* pQuery, STimeWindow* tw) {
   tw->ekey -= 1;
 }
 
-#define SET_STABLE_QUERY_OVER(_q) ((_q)->tableIndex = (int32_t)((_q)->tableqinfoGroupInfo.numOfTables))
-#define IS_STASBLE_QUERY_OVER(_q) ((_q)->tableIndex >= (int32_t)((_q)->tableqinfoGroupInfo.numOfTables))
-
-// todo move to utility
-static int32_t mergeIntoGroupResultImpl(SGroupResInfo* pGroupResInfo, SArray *pTableList, SQInfo* pQInfo);
-
 static void setResultOutputBuf(SQueryRuntimeEnv *pRuntimeEnv, SResultRow *pResult);
 static void setResultRowOutputBufInitCtx(SQueryRuntimeEnv *pRuntimeEnv, SResultRow *pResult);
 static bool functionNeedToExecute(SQueryRuntimeEnv *pRuntimeEnv, SQLFunctionCtx *pCtx, int32_t functionId);
 
 static void setExecParams(SQuery *pQuery, SQLFunctionCtx *pCtx, void* inputData, TSKEY *tsCol, SDataBlockInfo* pBlockInfo,
-                          SDataStatis *pStatis, void *param, int32_t colIndex, int32_t vgId);
+                          SDataStatis *pStatis, SExprInfo* pExprInfo);
 
 static void initCtxOutputBuf(SQueryRuntimeEnv *pRuntimeEnv);
 static void destroyTableQueryInfoImpl(STableQueryInfo *pTableQueryInfo);
 static void resetDefaultResInfoOutputBuf(SQueryRuntimeEnv *pRuntimeEnv);
 static bool hasMainOutput(SQuery *pQuery);
-static void buildTagQueryResult(SQInfo *pQInfo);
 
 static int32_t setTimestampListJoinInfo(SQInfo *pQInfo, STableQueryInfo *pTableQueryInfo);
-static int32_t checkForQueryBuf(size_t numOfTables);
 static void releaseQueryBuf(size_t numOfTables);
 static int32_t binarySearchForKey(char *pValue, int num, TSKEY key, int order);
 static void doRowwiseTimeWindowInterpolation(SQueryRuntimeEnv* pRuntimeEnv, SArray* pDataBlock, TSKEY prevTs, int32_t prevRowIndex, TSKEY curTs, int32_t curRowIndex, TSKEY windowKey, int32_t type);
@@ -296,11 +260,6 @@ void updateNumOfResult(SQueryRuntimeEnv *pRuntimeEnv, int32_t numOfRes) {
   }
 }
 
-static UNUSED_FUNC int32_t getMergeResultGroupId(int32_t groupIndex) {
-  int32_t base = 50000000;
-  return base + (groupIndex * 10000);
-}
-
 bool isGroupbyColumn(SSqlGroupbyExpr *pGroupbyExpr) {
   if (pGroupbyExpr == NULL || pGroupbyExpr->numOfGroupCols == 0) {
     return false;
@@ -345,7 +304,7 @@ int16_t getGroupbyColumnType(SQuery *pQuery, SSqlGroupbyExpr *pGroupbyExpr) {
   return type;
 }
 
-bool isSelectivityWithTagsQuery(SQuery *pQuery) {
+static bool isSelectivityWithTagsQuery(SQuery *pQuery) {
   bool    hasTags = false;
   int32_t numOfSelectivity = 0;
 
@@ -368,7 +327,7 @@ bool isSelectivityWithTagsQuery(SQuery *pQuery) {
   return false;
 }
 
-bool isProjQuery(SQuery *pQuery) {
+static bool isProjQuery(SQuery *pQuery) {
   for (int32_t i = 0; i < pQuery->numOfOutput; ++i) {
     int32_t functId = pQuery->pExpr1[i].base.functionId;
     if (functId != TSDB_FUNC_PRJ && functId != TSDB_FUNC_TAGPRJ) {
@@ -379,17 +338,14 @@ bool isProjQuery(SQuery *pQuery) {
   return true;
 }
 
-bool isTsCompQuery(SQuery *pQuery) { return pQuery->pExpr1[0].base.functionId == TSDB_FUNC_TS_COMP; }
-
-static bool limitOperator(SQueryRuntimeEnv* pRuntimeEnv) {
-  SQInfo* pQInfo = GET_QINFO_ADDR(pRuntimeEnv);
-  SQuery* pQuery = pRuntimeEnv->pQuery;
+static bool isTsCompQuery(SQuery *pQuery) { return pQuery->pExpr1[0].base.functionId == TSDB_FUNC_TS_COMP; }
 
+static bool limitOperator(SQuery* pQuery, void* qinfo) {
   if ((pQuery->limit.limit > 0) && (pQuery->rec.total + pQuery->rec.rows > pQuery->limit.limit)) {
     pQuery->rec.rows = pQuery->limit.limit - pQuery->rec.total;
 
     qDebug("QInfo:%p discard remain data due to result limitation, limit:%"PRId64", current return:%" PRId64 ", total:%"PRId64,
-        pQInfo, pQuery->limit.limit, pQuery->rec.rows, pQuery->rec.total + pQuery->rec.rows);
+           qinfo, pQuery->limit.limit, pQuery->rec.rows, pQuery->rec.total + pQuery->rec.rows);
     assert(pQuery->rec.rows >= 0);
     setQueryStatus(pQuery, QUERY_COMPLETED);
     return true;
@@ -643,7 +599,7 @@ static int32_t addNewWindowResultBuf(SResultRow *pWindowRes, SDiskbasedResultBuf
 }
 
 static int32_t setWindowOutputBufByKey(SQueryRuntimeEnv *pRuntimeEnv, SResultRowInfo *pResultRowInfo, STimeWindow *win,
-    bool masterscan, SResultRow** pResult, int64_t groupId) {
+                                       bool masterscan, SResultRow **pResult, int64_t groupId) {
   assert(win->skey <= win->ekey);
   SDiskbasedResultBuf *pResultBuf = pRuntimeEnv->pResultBuf;
 
@@ -826,9 +782,11 @@ static int32_t getNumOfRowsInTimeWindow(SQuery *pQuery, SDataBlockInfo *pDataBlo
   return num;
 }
 
-// TODO decouple the data block and the SQLFunctionCtx
-static void doBlockwiseApplyFunctions(SQueryRuntimeEnv *pRuntimeEnv, STimeWindow *pWin, int32_t offset, int32_t forwardStep, TSKEY *tsCol, int32_t numOfTotal) {
-  SQuery         *pQuery = pRuntimeEnv->pQuery;
+static char *getDataBlock(SQueryRuntimeEnv *pRuntimeEnv, SArithmeticSupport *sas, int32_t col, int32_t size, SArray *pDataBlock);
+
+static void doBlockwiseApplyFunctions(SQueryRuntimeEnv *pRuntimeEnv, STimeWindow *pWin, int32_t offset,
+                                      int32_t forwardStep, TSKEY *tsCol, int32_t numOfTotal, SArray *pDataBlock) {
+  SQuery *        pQuery = pRuntimeEnv->pQuery;
   SQLFunctionCtx *pCtx = pRuntimeEnv->pCtx;
 
   bool hasPrev = pCtx[0].preAggVals.isSet;
@@ -836,7 +794,17 @@ static void doBlockwiseApplyFunctions(SQueryRuntimeEnv *pRuntimeEnv, STimeWindow
   for (int32_t k = 0; k < pQuery->numOfOutput; ++k) {
     pCtx[k].size = forwardStep;
     pCtx[k].startTs = pWin->skey;
-    pCtx[k].startOffset = (QUERY_IS_ASC_QUERY(pQuery)) ? offset : offset - (forwardStep - 1);
+
+    char *dataBlock = getDataBlock(pRuntimeEnv, &pRuntimeEnv->sasArray[k], k, numOfTotal, pDataBlock);
+
+    int32_t pos = (QUERY_IS_ASC_QUERY(pQuery)) ? offset : offset - (forwardStep - 1);
+    if (dataBlock != NULL) {
+      pCtx[k].pInput = (char *)dataBlock + pos * pCtx[k].inputBytes;
+    }
+
+    if (tsCol != NULL) {
+      pCtx[k].ptsList = &tsCol[pos];
+    }
 
     int32_t functionId = pQuery->pExpr1[k].base.functionId;
 
@@ -976,6 +944,7 @@ static void* getDataBlockImpl(SArray* pDataBlock, int32_t colId) {
   return NULL;
 }
 
+// todo refactor
 static char *getDataBlock(SQueryRuntimeEnv *pRuntimeEnv, SArithmeticSupport *sas, int32_t col, int32_t size, SArray *pDataBlock) {
   if (pDataBlock == NULL) {
     return NULL;
@@ -1190,10 +1159,9 @@ static void blockwiseApplyFunctions(SQueryRuntimeEnv *pRuntimeEnv, SDataStatis *
     tsCols = (TSKEY *)(pColInfo->pData);
   }
 
-  SQInfo *pQInfo = GET_QINFO_ADDR(pRuntimeEnv);
   for (int32_t k = 0; k < pQuery->numOfOutput; ++k) {
     char *dataBlock = getDataBlock(pRuntimeEnv, &pRuntimeEnv->sasArray[k], k, pDataBlockInfo->rows, pDataBlock);
-    setExecParams(pQuery, &pCtx[k], dataBlock, tsCols, pDataBlockInfo, pStatis, &pRuntimeEnv->sasArray[k], k, pQInfo->vgId);
+    setExecParams(pQuery, &pCtx[k], dataBlock, tsCols, pDataBlockInfo, pStatis, &pQuery->pExpr1[k]);
   }
 
   int32_t step = GET_FORWARD_DIRECTION_FACTOR(pQuery->order.order);
@@ -1234,7 +1202,7 @@ static void blockwiseApplyFunctions(SQueryRuntimeEnv *pRuntimeEnv, SDataStatis *
         setResultRowInterpo(pResult, RESULT_ROW_END_INTERP);
         setNotInterpoWindowKey(pRuntimeEnv->pCtx, pQuery->numOfOutput, RESULT_ROW_START_INTERP);
 
-        doBlockwiseApplyFunctions(pRuntimeEnv, &w, startPos, 0, tsCols, pDataBlockInfo->rows);
+        doBlockwiseApplyFunctions(pRuntimeEnv, &w, startPos, 0, tsCols, pDataBlockInfo->rows, pDataBlock);
       }
 
       // restore current time window
@@ -1244,7 +1212,7 @@ static void blockwiseApplyFunctions(SQueryRuntimeEnv *pRuntimeEnv, SDataStatis *
 
     // window start key interpolation
     doWindowBorderInterpolation(pRuntimeEnv, pDataBlockInfo, pDataBlock, pResult, &win, pQuery->pos, forwardStep);
-    doBlockwiseApplyFunctions(pRuntimeEnv, &win, startPos, forwardStep, tsCols, pDataBlockInfo->rows);
+    doBlockwiseApplyFunctions(pRuntimeEnv, &win, startPos, forwardStep, tsCols, pDataBlockInfo->rows, pDataBlock);
 
     STimeWindow nextWin = win;
     while (1) {
@@ -1265,7 +1233,7 @@ static void blockwiseApplyFunctions(SQueryRuntimeEnv *pRuntimeEnv, SDataStatis *
 
       // window start(end) key interpolation
       doWindowBorderInterpolation(pRuntimeEnv, pDataBlockInfo, pDataBlock, pResult, &nextWin, startPos, forwardStep);
-      doBlockwiseApplyFunctions(pRuntimeEnv, &nextWin, startPos, forwardStep, tsCols, pDataBlockInfo->rows);
+      doBlockwiseApplyFunctions(pRuntimeEnv, &nextWin, startPos, forwardStep, tsCols, pDataBlockInfo->rows, pDataBlock);
     }
 
   } else {
@@ -1305,7 +1273,7 @@ static int32_t setGroupResultOutputBuf(SQueryRuntimeEnv *pRuntimeEnv, char *pDat
     return -1;
   }
 
-  SResultRow *pResultRow = doPrepareResultRowFromKey(pRuntimeEnv, &pRuntimeEnv->windowResInfo, d, len, true, groupIndex);
+  SResultRow *pResultRow = doPrepareResultRowFromKey(pRuntimeEnv, &pRuntimeEnv->resultRowInfo, d, len, true, groupIndex);
   assert (pResultRow != NULL);
 
   int64_t v = -1;
@@ -1556,7 +1524,7 @@ static void rowwiseApplyFunctions(SQueryRuntimeEnv *pRuntimeEnv, SDataStatis *pS
   SQInfo* pQInfo = GET_QINFO_ADDR(pRuntimeEnv);
   for (int32_t k = 0; k < pQuery->numOfOutput; ++k) {
     char *dataBlock = getDataBlock(pRuntimeEnv, &pRuntimeEnv->sasArray[k], k, pDataBlockInfo->rows, pDataBlock);
-    setExecParams(pQuery, &pCtx[k], dataBlock, tsCols, pDataBlockInfo, pStatis, &pRuntimeEnv->sasArray[k], k, pQInfo->vgId);
+    setExecParams(pQuery, &pCtx[k], dataBlock, tsCols, pDataBlockInfo, pStatis, &pQuery->pExpr1[k]);
     pCtx[k].size = 1;
   }
 
@@ -1723,7 +1691,7 @@ static int32_t tableApplyFunctionsOnBlock(SQueryRuntimeEnv *pRuntimeEnv, SDataBl
   SQuery *pQuery = pRuntimeEnv->pQuery;
 
   STableQueryInfo* pTableQueryInfo = pQuery->current;
-  SResultRowInfo*  pResultRowInfo = &pRuntimeEnv->windowResInfo;
+  SResultRowInfo*  pResultRowInfo = &pRuntimeEnv->resultRowInfo;
 
   if (pQuery->numOfFilterCols > 0 || pRuntimeEnv->pTsBuf != NULL || pRuntimeEnv->groupbyColumn) {
     rowwiseApplyFunctions(pRuntimeEnv, pStatis, pDataBlockInfo, pResultRowInfo, pDataBlock);
@@ -1767,14 +1735,13 @@ static int32_t tableApplyFunctionsOnBlock(SQueryRuntimeEnv *pRuntimeEnv, SDataBl
 }
 
 void setExecParams(SQuery *pQuery, SQLFunctionCtx *pCtx, void* inputData, TSKEY *tsCol, SDataBlockInfo* pBlockInfo,
-                   SDataStatis *pStatis, void *param, int32_t colIndex, int32_t vgId) {
+                   SDataStatis *pStatis, SExprInfo* pExprInfo) {
 
-  int32_t functionId = pQuery->pExpr1[colIndex].base.functionId;
-  int32_t colId = pQuery->pExpr1[colIndex].base.colInfo.colId;
+  int32_t functionId = pExprInfo->base.functionId;
+  int32_t colId = pExprInfo->base.colInfo.colId;
 
   SDataStatis *tpField = NULL;
-  pCtx->hasNull = hasNullValue(&pQuery->pExpr1[colIndex].base.colInfo, pStatis, &tpField);
-  pCtx->pInput = inputData;
+  pCtx->hasNull = hasNullValue(&pExprInfo->base.colInfo, pStatis, &tpField);
 
   if (tpField != NULL) {
     pCtx->preAggVals.isSet  = true;
@@ -1789,73 +1756,24 @@ void setExecParams(SQuery *pQuery, SQLFunctionCtx *pCtx, void* inputData, TSKEY
   // limit/offset query will affect this value
   pCtx->size = QUERY_IS_ASC_QUERY(pQuery) ? pBlockInfo->rows - pQuery->pos : pQuery->pos + 1;
 
-  // minimum value no matter ascending/descending order query
-  pCtx->startOffset = QUERY_IS_ASC_QUERY(pQuery) ? pQuery->pos: (pQuery->pos - pCtx->size + 1);
-  assert(pCtx->startOffset >= 0);
+  // set the start position in current block
+  int32_t offset = QUERY_IS_ASC_QUERY(pQuery) ? pQuery->pos: (pQuery->pos - pCtx->size + 1);
+  if (inputData != NULL) {
+    pCtx->pInput = (char*)inputData + offset * pCtx->inputBytes;
+  }
 
   uint32_t status = aAggs[functionId].status;
   if (((status & (TSDB_FUNCSTATE_SELECTIVITY | TSDB_FUNCSTATE_NEED_TS)) != 0) && (tsCol != NULL)) {
-    pCtx->ptsList = tsCol;
-  }
-
-  if (functionId >= TSDB_FUNC_FIRST_DST && functionId <= TSDB_FUNC_LAST_DST) {
-    // last_dist or first_dist function
-    // store the first&last timestamp into the intermediate buffer [1], the true
-    // value may be null but timestamp will never be null
-  } else if (functionId == TSDB_FUNC_TOP || functionId == TSDB_FUNC_BOTTOM || functionId == TSDB_FUNC_TWA ||
-             functionId == TSDB_FUNC_DIFF || (functionId >= TSDB_FUNC_RATE && functionId <= TSDB_FUNC_AVG_IRATE)) {
-    /*
-     * least squares function needs two columns of input, currently, the x value of linear equation is set to
-     * timestamp column, and the y-value is the column specified in pQuery->pExpr1[i].colIdxInBuffer
-     *
-     * top/bottom function needs timestamp to indicate when the
-     * top/bottom values emerge, so does diff function
-     */
-    if (functionId == TSDB_FUNC_TWA) {
-       pCtx->param[1].i64 = pQuery->window.skey;
-       pCtx->param[1].nType = TSDB_DATA_TYPE_BIGINT;
-       pCtx->param[2].i64 = pQuery->window.ekey;
-       pCtx->param[2].nType = TSDB_DATA_TYPE_BIGINT;
+    pCtx->ptsList = tsCol + offset;
   }
 
-  } else if (functionId == TSDB_FUNC_ARITHM) {
-    pCtx->param[1].pz = param;
-  } else if (functionId == TSDB_FUNC_SPREAD) {  // set the statistics data for primary time stamp column
+  if (functionId == TSDB_FUNC_SPREAD) {  // set the statistics data for primary time stamp column
     if (colId == PRIMARYKEY_TIMESTAMP_COL_INDEX) {
       pCtx->preAggVals.isSet  = true;
       pCtx->preAggVals.statis.min = pBlockInfo->window.skey;
       pCtx->preAggVals.statis.max = pBlockInfo->window.ekey;
     }
-  } else if (functionId == TSDB_FUNC_INTERP) {
-    pCtx->param[2].i64 = (int8_t) pQuery->fillType;
-    if (pQuery->fillVal != NULL) {
-      if (isNull((const char*) &pQuery->fillVal[colIndex], pCtx->inputType)) {
-        pCtx->param[1].nType = TSDB_DATA_TYPE_NULL;
-      } else { // todo refactor, tVariantCreateFromBinary should handle the NULL value
-        if (pCtx->inputType != TSDB_DATA_TYPE_BINARY && pCtx->inputType != TSDB_DATA_TYPE_NCHAR) {
-          tVariantCreateFromBinary(&pCtx->param[1], (char*) &pQuery->fillVal[colIndex], pCtx->inputBytes, pCtx->inputType);
-        }
-      }
-    }
-  } else if (functionId == TSDB_FUNC_TS_COMP) {
-    pCtx->param[0].i64 = vgId;
-    pCtx->param[0].nType = TSDB_DATA_TYPE_BIGINT;
   }
-
-#if defined(_DEBUG_VIEW)
-  //  int64_t *tsList = (int64_t *)primaryColumnData;
-//  int64_t  s = tsList[0];
-//  int64_t  e = tsList[size - 1];
-
-//    if (IS_DATA_BLOCK_LOADED(blockStatus)) {
-//        qDebug("QInfo:%p query ts:%lld-%lld, offset:%d, rows:%d, bstatus:%d,
-//        functId:%d", GET_QINFO_ADDR(pQuery),
-//               s, e, startOffset, size, blockStatus, functionId);
-//    } else {
-//        qDebug("QInfo:%p block not loaded, bstatus:%d",
-//        GET_QINFO_ADDR(pQuery), blockStatus);
-//    }
-#endif
 }
 
 // set the output buffer for the selectivity + tag query
@@ -1900,7 +1818,7 @@ static int32_t setCtxTagColumnInfo(SQueryRuntimeEnv *pRuntimeEnv, SQLFunctionCtx
   return TSDB_CODE_SUCCESS;
 }
 
-static int32_t setupQueryRuntimeEnv(SQueryRuntimeEnv *pRuntimeEnv, int32_t numOfTables, int16_t order) {
+static int32_t setupQueryRuntimeEnv(SQueryRuntimeEnv *pRuntimeEnv, int32_t numOfTables, int16_t order, int32_t vgId) {
   qDebug("QInfo:%p setup runtime env", GET_QINFO_ADDR(pRuntimeEnv));
   SQuery *pQuery = pRuntimeEnv->pQuery;
 
@@ -1912,30 +1830,29 @@ static int32_t setupQueryRuntimeEnv(SQueryRuntimeEnv *pRuntimeEnv, int32_t numOf
   pRuntimeEnv->pool    = initResultRowPool(getResultRowSize(pRuntimeEnv));
   pRuntimeEnv->prevRow = malloc(POINTER_BYTES * pQuery->numOfCols + pQuery->srcRowSize);
   pRuntimeEnv->tagVal  = malloc(pQuery->tagLen);
-
-  char* start = POINTER_BYTES * pQuery->numOfCols + (char*) pRuntimeEnv->prevRow;
-  pRuntimeEnv->prevRow[0] = start;
-
-  for(int32_t i = 1; i < pQuery->numOfCols; ++i) {
-    pRuntimeEnv->prevRow[i] = pRuntimeEnv->prevRow[i - 1] + pQuery->colList[i-1].bytes;
-  }
-
   pRuntimeEnv->pCtx    = (SQLFunctionCtx *)calloc(pQuery->numOfOutput, sizeof(SQLFunctionCtx));
   pRuntimeEnv->offset  = calloc(pQuery->numOfOutput, sizeof(int16_t));
   pRuntimeEnv->rowCellInfoOffset = calloc(pQuery->numOfOutput, sizeof(int32_t));
   pRuntimeEnv->sasArray = calloc(pQuery->numOfOutput, sizeof(SArithmeticSupport));
 
-  // TODO check malloc failure
-  if (pRuntimeEnv->offset == NULL || pRuntimeEnv->pCtx == NULL || pRuntimeEnv->rowCellInfoOffset == NULL || pRuntimeEnv->sasArray == NULL) {
+  if (pRuntimeEnv->offset == NULL || pRuntimeEnv->pCtx == NULL || pRuntimeEnv->rowCellInfoOffset == NULL ||
+      pRuntimeEnv->sasArray == NULL || pRuntimeEnv->pResultRowHashTable == NULL || pRuntimeEnv->keyBuf == NULL ||
+      pRuntimeEnv->prevRow == NULL || pRuntimeEnv->tagVal == NULL) {
     goto _clean;
   }
 
+  char* start = POINTER_BYTES * pQuery->numOfCols + (char*) pRuntimeEnv->prevRow;
+  pRuntimeEnv->prevRow[0] = start;
+  for(int32_t i = 1; i < pQuery->numOfCols; ++i) {
+    pRuntimeEnv->prevRow[i] = pRuntimeEnv->prevRow[i - 1] + pQuery->colList[i-1].bytes;
+  }
+
   pRuntimeEnv->offset[0] = 0;
   for (int32_t i = 0; i < pQuery->numOfOutput; ++i) {
     SSqlFuncMsg *pSqlFuncMsg = &pQuery->pExpr1[i].base;
 
     SQLFunctionCtx *pCtx = &pRuntimeEnv->pCtx[i];
-    SColIndex* pIndex = &pSqlFuncMsg->colInfo;
+    SColIndex *     pIndex = &pSqlFuncMsg->colInfo;
 
     if (TSDB_COL_REQ_NULL(pIndex->flag)) {
       pCtx->requireNull = true;
@@ -1947,7 +1864,7 @@ static int32_t setupQueryRuntimeEnv(SQueryRuntimeEnv *pRuntimeEnv, int32_t numOf
     int32_t index = pSqlFuncMsg->colInfo.colIndex;
     if (TSDB_COL_IS_TAG(pIndex->flag)) {
       if (pIndex->colId == TSDB_TBNAME_COLUMN_INDEX) {  // todo refactor
-        SSchema* s = tGetTbnameColumnSchema();
+        SSchema *s = tGetTbnameColumnSchema();
 
         pCtx->inputBytes = s->bytes;
         pCtx->inputType = s->type;
@@ -2008,18 +1925,38 @@ static int32_t setupQueryRuntimeEnv(SQueryRuntimeEnv *pRuntimeEnv, int32_t numOf
       pCtx->param[3].nType = TSDB_DATA_TYPE_BIGINT;
 
       pCtx->param[1].i64 = pQuery->order.orderColId;
+    } else if (functionId == TSDB_FUNC_INTERP) {
+      pCtx->param[2].i64 = (int8_t)pQuery->fillType;
+      if (pQuery->fillVal != NULL) {
+        if (isNull((const char *)&pQuery->fillVal[i], pCtx->inputType)) {
+          pCtx->param[1].nType = TSDB_DATA_TYPE_NULL;
+        } else {  // todo refactor, tVariantCreateFromBinary should handle the NULL value
+          if (pCtx->inputType != TSDB_DATA_TYPE_BINARY && pCtx->inputType != TSDB_DATA_TYPE_NCHAR) {
+            tVariantCreateFromBinary(&pCtx->param[1], (char *)&pQuery->fillVal[i], pCtx->inputBytes, pCtx->inputType);
           }
-
-    if (functionId == TSDB_FUNC_ARITHM) {
+        }
+      }
+    } else if (functionId == TSDB_FUNC_TS_COMP) {
+      pCtx->param[0].i64 = vgId;
+      pCtx->param[0].nType = TSDB_DATA_TYPE_BIGINT;
+    } else if (functionId == TSDB_FUNC_TWA) {
+      pCtx->param[1].i64 = pQuery->window.skey;
+      pCtx->param[1].nType = TSDB_DATA_TYPE_BIGINT;
+      pCtx->param[2].i64 = pQuery->window.ekey;
+      pCtx->param[2].nType = TSDB_DATA_TYPE_BIGINT;
+    } else if (functionId == TSDB_FUNC_ARITHM) {
       pRuntimeEnv->sasArray[i].data = calloc(pQuery->numOfCols, POINTER_BYTES);
       if (pRuntimeEnv->sasArray[i].data == NULL) {
         goto _clean;
       }
+
+      pCtx->param[1].pz = (char*) &pRuntimeEnv->sasArray[i];
     }
 
     if (i > 0) {
       pRuntimeEnv->offset[i] = pRuntimeEnv->offset[i - 1] + pRuntimeEnv->pCtx[i - 1].outputBytes;
-      pRuntimeEnv->rowCellInfoOffset[i] = pRuntimeEnv->rowCellInfoOffset[i - 1] + sizeof(SResultRowCellInfo) + pQuery->pExpr1[i - 1].interBytes;
+      pRuntimeEnv->rowCellInfoOffset[i] =
+          pRuntimeEnv->rowCellInfoOffset[i - 1] + sizeof(SResultRowCellInfo) + pQuery->pExpr1[i - 1].interBytes;
     }
   }
 
@@ -2043,6 +1980,10 @@ _clean:
   tfree(pRuntimeEnv->offset);
   tfree(pRuntimeEnv->rowCellInfoOffset);
   tfree(pRuntimeEnv->sasArray);
+  tfree(pRuntimeEnv->pResultRowHashTable);
+  tfree(pRuntimeEnv->keyBuf);
+  tfree(pRuntimeEnv->prevRow);
+  tfree(pRuntimeEnv->tagVal);
 
   return TSDB_CODE_QRY_OUT_OF_MEMORY;
 }
@@ -2070,7 +2011,7 @@ static void teardownQueryRuntimeEnv(SQueryRuntimeEnv *pRuntimeEnv) {
   SQInfo* pQInfo = (SQInfo*) GET_QINFO_ADDR(pRuntimeEnv);
 
   qDebug("QInfo:%p teardown runtime env", pQInfo);
-  cleanupResultRowInfo(&pRuntimeEnv->windowResInfo);
+  cleanupResultRowInfo(&pRuntimeEnv->resultRowInfo);
 
   if (isTsCompQuery(pQuery)) {
     FILE *f = *(FILE **)pQuery->sdata[0]->data;
@@ -2118,7 +2059,6 @@ static void teardownQueryRuntimeEnv(SQueryRuntimeEnv *pRuntimeEnv) {
   tfree(pRuntimeEnv->prevRow);
   tfree(pRuntimeEnv->tagVal);
 
-
   taosHashCleanup(pRuntimeEnv->pResultRowHashTable);
   pRuntimeEnv->pResultRowHashTable = NULL;
 
@@ -2131,9 +2071,7 @@ static bool needBuildResAfterQueryComplete(SQInfo* pQInfo) {
   return pQInfo->rspContext != NULL;
 }
 
-#define IS_QUERY_KILLED(_q) ((_q)->code == TSDB_CODE_TSC_QUERY_CANCELLED)
-
-static bool isQueryKilled(SQInfo *pQInfo) {
+bool isQueryKilled(SQInfo *pQInfo) {
   if (IS_QUERY_KILLED(pQInfo)) {
     return true;
   }
@@ -2152,7 +2090,7 @@ static bool isQueryKilled(SQInfo *pQInfo) {
   return false;
 }
 
-static void setQueryKilled(SQInfo *pQInfo) { pQInfo->code = TSDB_CODE_TSC_QUERY_CANCELLED;}
+void setQueryKilled(SQInfo *pQInfo) { pQInfo->code = TSDB_CODE_TSC_QUERY_CANCELLED;}
 
 static bool isFixedOutputQuery(SQueryRuntimeEnv* pRuntimeEnv) {
   SQuery* pQuery = pRuntimeEnv->pQuery;
@@ -2253,7 +2191,7 @@ static bool needReverseScan(SQuery *pQuery) {
  * The following 4 kinds of query are treated as the tags query
  * tagprj, tid_tag query, count(tbname), 'abc' (user defined constant value column) query
  */
-static bool onlyQueryTags(SQuery* pQuery) {
+bool onlyQueryTags(SQuery* pQuery) {
   for(int32_t i = 0; i < pQuery->numOfOutput; ++i) {
     SExprInfo* pExprInfo = &pQuery->pExpr1[i];
 
@@ -2351,7 +2289,6 @@ static bool onlyFirstQuery(SQuery *pQuery) { return onlyOneQueryType(pQuery, TSD
 
 static bool onlyLastQuery(SQuery *pQuery) { return onlyOneQueryType(pQuery, TSDB_FUNC_LAST, TSDB_FUNC_LAST_DST); }
 
-// todo refactor, add iterator
 static void doExchangeTimeWindow(SQInfo* pQInfo, STimeWindow* win) {
   size_t t = taosArrayGetSize(pQInfo->tableGroupInfo.pGroupList);
   for(int32_t i = 0; i < t; ++i) {
@@ -2769,56 +2706,42 @@ int32_t binarySearchForKey(char *pValue, int num, TSKEY key, int order) {
   return midPos;
 }
 
-static void ensureOutputBufferSimple(SQueryRuntimeEnv* pRuntimeEnv, int32_t capacity) {
+static void expandBuffer(SQueryRuntimeEnv* pRuntimeEnv, int32_t newSize, void* qinfo) {
   SQuery* pQuery = pRuntimeEnv->pQuery;
+  SResultRec *pRec = &pQuery->rec;
 
-  if (capacity < pQuery->rec.capacity) {
-    return;
-  }
+  assert(newSize > 0);
 
   for (int32_t i = 0; i < pQuery->numOfOutput; ++i) {
     int32_t bytes = pQuery->pExpr1[i].bytes;
-    assert(bytes > 0 && capacity > 0);
 
-    char *tmp = realloc(pQuery->sdata[i], bytes * capacity + sizeof(tFilePage));
+    char *tmp = realloc(pQuery->sdata[i], bytes * newSize + sizeof(tFilePage));
     if (tmp == NULL) {
       longjmp(pRuntimeEnv->env, TSDB_CODE_QRY_OUT_OF_MEMORY);
     } else {
+      memset(tmp + sizeof(tFilePage) + bytes * pRec->rows, 0, (size_t)((newSize - pRec->rows) * bytes));
       pQuery->sdata[i] = (tFilePage *)tmp;
     }
-
-    // set the pCtx output buffer position
-    pRuntimeEnv->pCtx[i].pOutput = pQuery->sdata[i]->data;
   }
 
-  qDebug("QInfo:%p realloc output buffer to inc output buffer from: %" PRId64 " rows to:%d rows", GET_QINFO_ADDR(pRuntimeEnv),
-         pQuery->rec.capacity, capacity);
-
-  pQuery->rec.capacity = capacity;
+  pRec->capacity = newSize;
+  qDebug("QInfo:%p realloc output buffer, new size: %d rows, old:%" PRId64 ", remain:%" PRId64, qinfo, newSize,
+      pRec->capacity, newSize - pRec->rows);
 }
 
-// TODO merge with enuserOutputBufferSimple
-static void ensureOutputBuffer(SQueryRuntimeEnv* pRuntimeEnv, SDataBlockInfo* pBlockInfo) {
+static void ensureOutputBuffer(SQueryRuntimeEnv* pRuntimeEnv, int32_t numOfRows) {
   // in case of prj/diff query, ensure the output buffer is sufficient to accommodate the results of current block
   SQuery* pQuery = pRuntimeEnv->pQuery;
   if (!QUERY_IS_INTERVAL_QUERY(pQuery) && !pRuntimeEnv->groupbyColumn && !isFixedOutputQuery(pRuntimeEnv) && !isTsCompQuery(pQuery)) {
     SResultRec *pRec = &pQuery->rec;
 
-    if (pQuery->rec.capacity - pQuery->rec.rows < pBlockInfo->rows) {
     int32_t remain  = (int32_t)(pRec->capacity - pRec->rows);
-      int32_t newSize = (int32_t)(pRec->capacity + (pBlockInfo->rows - remain));
+    if (remain < numOfRows) {
+      int32_t newSize = (int32_t)(pRec->capacity + (numOfRows - remain));
+      expandBuffer(pRuntimeEnv, newSize, GET_QINFO_ADDR(pRuntimeEnv));
 
       for (int32_t i = 0; i < pQuery->numOfOutput; ++i) {
         int32_t bytes = pQuery->pExpr1[i].bytes;
-        assert(bytes > 0 && newSize > 0);
-
-        char *tmp = realloc(pQuery->sdata[i], bytes * newSize + sizeof(tFilePage));
-        if (tmp == NULL) {
-          longjmp(pRuntimeEnv->env, TSDB_CODE_QRY_OUT_OF_MEMORY);
-        } else {
-          memset(tmp + sizeof(tFilePage) + bytes * pRec->rows, 0, (size_t)((newSize - pRec->rows) * bytes));
-          pQuery->sdata[i] = (tFilePage *)tmp;
-        }
 
         // set the pCtx output buffer position
         pRuntimeEnv->pCtx[i].pOutput = pQuery->sdata[i]->data + pRec->rows * bytes;
@@ -2828,11 +2751,6 @@ static void ensureOutputBuffer(SQueryRuntimeEnv* pRuntimeEnv, SDataBlockInfo* pB
           pRuntimeEnv->pCtx[i].ptsOutputBuf = pRuntimeEnv->pCtx[0].pOutput;
         }
       }
-
-      qDebug("QInfo:%p realloc output buffer, new size: %d rows, old:%" PRId64 ", remain:%" PRId64, GET_QINFO_ADDR(pRuntimeEnv),
-             newSize, pRec->capacity, newSize - pRec->rows);
-
-      pRec->capacity = newSize;
     }
   }
 }
@@ -2840,9 +2758,9 @@ static void ensureOutputBuffer(SQueryRuntimeEnv* pRuntimeEnv, SDataBlockInfo* pB
 static void doSetInitialTimewindow(SQueryRuntimeEnv* pRuntimeEnv, SDataBlockInfo* pBlockInfo) {
   SQuery* pQuery = pRuntimeEnv->pQuery;
 
-  if (QUERY_IS_INTERVAL_QUERY(pQuery) && pRuntimeEnv->windowResInfo.prevSKey == TSKEY_INITIAL_VAL) {
+  if (QUERY_IS_INTERVAL_QUERY(pQuery) && pRuntimeEnv->resultRowInfo.prevSKey == TSKEY_INITIAL_VAL) {
     STimeWindow w = TSWINDOW_INITIALIZER;
-    SResultRowInfo *pWindowResInfo = &pRuntimeEnv->windowResInfo;
+    SResultRowInfo *pWindowResInfo = &pRuntimeEnv->resultRowInfo;
 
     if (QUERY_IS_ASC_QUERY(pQuery)) {
       getAlignQueryTimeWindow(pQuery, pBlockInfo->window.skey, pBlockInfo->window.skey, pQuery->window.ekey, &w);
@@ -2882,13 +2800,13 @@ static int64_t doScanAllDataBlocks(SQueryRuntimeEnv *pRuntimeEnv) {
     doSetInitialTimewindow(pRuntimeEnv, &blockInfo);
 
     // in case of prj/diff query, ensure the output buffer is sufficient to accommodate the results of current block
-    ensureOutputBuffer(pRuntimeEnv, &blockInfo);
+    ensureOutputBuffer(pRuntimeEnv, blockInfo.rows);
 
     SDataStatis *pStatis = NULL;
     SArray *     pDataBlock = NULL;
     uint32_t     status = 0;
 
-    int32_t ret = loadDataBlockOnDemand(pRuntimeEnv, &pRuntimeEnv->windowResInfo, pQueryHandle, &blockInfo, &pStatis, &pDataBlock, &status);
+    int32_t ret = loadDataBlockOnDemand(pRuntimeEnv, &pRuntimeEnv->resultRowInfo, pQueryHandle, &blockInfo, &pStatis, &pDataBlock, &status);
     if (ret != TSDB_CODE_SUCCESS) {
       break;
     }
@@ -2923,8 +2841,8 @@ static int64_t doScanAllDataBlocks(SQueryRuntimeEnv *pRuntimeEnv) {
   }
 
   if (QUERY_IS_INTERVAL_QUERY(pQuery)) {
-    closeAllResultRows(&pRuntimeEnv->windowResInfo);
-    pRuntimeEnv->windowResInfo.curIndex = pRuntimeEnv->windowResInfo.size - 1;  // point to the last time window
+    closeAllResultRows(&pRuntimeEnv->resultRowInfo);
+    pRuntimeEnv->resultRowInfo.curIndex = pRuntimeEnv->resultRowInfo.size - 1;  // point to the last time window
   }
 
   return 0;
@@ -3142,253 +3060,35 @@ void UNUSED_FUNC displayInterResult(tFilePage **pdata, SQueryRuntimeEnv* pRuntim
   }
 }
 
-typedef struct SCompSupporter {
-  STableQueryInfo **pTableQueryInfo;
-  int32_t          *rowIndex;
-  int32_t           order;
-} SCompSupporter;
-
-int32_t tableResultComparFn(const void *pLeft, const void *pRight, void *param) {
-  int32_t left  = *(int32_t *)pLeft;
-  int32_t right = *(int32_t *)pRight;
-
-  SCompSupporter *  supporter = (SCompSupporter *)param;
-
-  int32_t leftPos  = supporter->rowIndex[left];
-  int32_t rightPos = supporter->rowIndex[right];
-
-  /* left source is exhausted */
-  if (leftPos == -1) {
-    return 1;
-  }
-
-  /* right source is exhausted*/
-  if (rightPos == -1) {
-    return -1;
-  }
-
-  STableQueryInfo** pList = supporter->pTableQueryInfo;
-
-  SResultRowInfo *pWindowResInfo1 = &(pList[left]->windowResInfo);
-  SResultRow * pWindowRes1 = getResultRow(pWindowResInfo1, leftPos);
-  TSKEY leftTimestamp = pWindowRes1->win.skey;
-
-  SResultRowInfo *pWindowResInfo2 = &(pList[right]->windowResInfo);
-  SResultRow * pWindowRes2 = getResultRow(pWindowResInfo2, rightPos);
-  TSKEY rightTimestamp = pWindowRes2->win.skey;
-
-  if (leftTimestamp == rightTimestamp) {
-    return 0;
-  }
-
-  if (supporter->order == TSDB_ORDER_ASC) {
-    return (leftTimestamp > rightTimestamp)? 1:-1;
-  } else {
-    return (leftTimestamp < rightTimestamp)? 1:-1;
-  }
-}
-
-int32_t mergeGroupResult(SQInfo *pQInfo) {
-  int64_t st = taosGetTimestampUs();
-
-  SGroupResInfo* pGroupResInfo = &pQInfo->groupResInfo;
-
-  int32_t numOfGroups = (int32_t)(GET_NUM_OF_TABLEGROUP(pQInfo));
-  while (pQInfo->groupIndex < numOfGroups) {
-    SArray *group = GET_TABLEGROUP(pQInfo, pQInfo->groupIndex);
-
-    int32_t ret = mergeIntoGroupResultImpl(pGroupResInfo, group, pQInfo);
-    if (ret != TSDB_CODE_SUCCESS) {
-      return ret;
-    }
-
-    // this group generates at least one result, return results
-    pQInfo->groupIndex += 1;
-    if (taosArrayGetSize(pGroupResInfo->pRows) > 0) {
-      break;
-    }
-
-    qDebug("QInfo:%p no result in group %d, continue", pQInfo, pQInfo->groupIndex - 1);
-    taosArrayClear(pGroupResInfo->pRows);
-
-    pGroupResInfo->index = 0;
-    pGroupResInfo->rowId = 0;
-  }
-
-  if (pQInfo->groupIndex == numOfGroups && taosArrayGetSize(pGroupResInfo->pRows) == 0) {
-    SET_STABLE_QUERY_OVER(pQInfo);
-  }
-
-  int64_t elapsedTime = taosGetTimestampUs() - st;
-  qDebug("QInfo:%p merge res data into group, index:%d, total group:%d, elapsed time:%" PRId64 "us", pQInfo,
-         pQInfo->groupIndex - 1, numOfGroups, elapsedTime);
-
-  pQInfo->runtimeEnv.summary.firstStageMergeTime += elapsedTime;
-  return TSDB_CODE_SUCCESS;
-}
-
-static int32_t doCopyToSData(SQInfo *pQInfo, SResultRow **pRows, int32_t numOfRows, int32_t* index, int32_t orderType);
+static int32_t doCopyToSData(SQueryRuntimeEnv* pRuntimeEnv, SGroupResInfo* pGroupResInfo, int32_t orderType);
 
 void copyResToQueryResultBuf(SQInfo *pQInfo, SQuery *pQuery) {
   SGroupResInfo* pGroupResInfo = &pQInfo->groupResInfo;
 
+  while(pGroupResInfo->currentGroup < pGroupResInfo->totalGroup) {
     // all results in current group have been returned to client, try next group
-  if (pGroupResInfo->index >= taosArrayGetSize(pGroupResInfo->pRows)) {
-    // current results of group has been sent to client, try next group
-    pGroupResInfo->index = 0;
-    pGroupResInfo->rowId = 0;
-    taosArrayClear(pGroupResInfo->pRows);
-
-    if (mergeGroupResult(pQInfo) != TSDB_CODE_SUCCESS) {
-      return;  // failed to save data in the disk
-    }
-
-    // check if all results has been sent to client
-    int32_t numOfGroup = (int32_t)(GET_NUM_OF_TABLEGROUP(pQInfo));
-    if (taosArrayGetSize(pGroupResInfo->pRows) == 0 && pQInfo->groupIndex == numOfGroup) {
-      SET_STABLE_QUERY_OVER(pQInfo);
+    if ((pGroupResInfo->pRows == NULL) || taosArrayGetSize(pGroupResInfo->pRows) == 0) {
+      assert(pGroupResInfo->index == 0);
+      if ((pQInfo->code = mergeIntoGroupResult(&pQInfo->groupResInfo, pQInfo)) != TSDB_CODE_SUCCESS) {
         return;
       }
     }
 
-  int32_t size = (int32_t) taosArrayGetSize(pGroupResInfo->pRows);
-  pQuery->rec.rows = doCopyToSData(pQInfo, pGroupResInfo->pRows->pData, size, &pGroupResInfo->index, TSDB_ORDER_ASC);
-}
-
-int64_t getNumOfResultWindowRes(SQueryRuntimeEnv* pRuntimeEnv, SResultRow *pResultRow) {
-  SQuery* pQuery = pRuntimeEnv->pQuery;
-
-  for (int32_t j = 0; j < pQuery->numOfOutput; ++j) {
-    int32_t functionId = pQuery->pExpr1[j].base.functionId;
-
-    /*
-     * ts, tag, tagprj function can not decide the output number of current query
-     * the number of output result is decided by main output
-     */
-    if (functionId == TSDB_FUNC_TS || functionId == TSDB_FUNC_TAG || functionId == TSDB_FUNC_TAGPRJ) {
-      continue;
-    }
-
-    SResultRowCellInfo *pResultInfo = getResultCell(pRuntimeEnv, pResultRow, j);
-    assert(pResultInfo != NULL);
-
-    if (pResultInfo->numOfRes > 0) {
-      return pResultInfo->numOfRes;
-    }
-  }
-
-  return 0;
-}
-
-int32_t mergeIntoGroupResultImpl(SGroupResInfo* pGroupResInfo, SArray *pTableList, SQInfo* pQInfo) {
-  SQueryRuntimeEnv *pRuntimeEnv = &pQInfo->runtimeEnv;
-  bool ascQuery = QUERY_IS_ASC_QUERY(pRuntimeEnv->pQuery);
-
-  int32_t code = TSDB_CODE_SUCCESS;
-
-  int32_t *posList = NULL;
-  SLoserTreeInfo *pTree = NULL;
-  STableQueryInfo **pTableQueryInfoList = NULL;
-
-  size_t size = taosArrayGetSize(pTableList);
-  if (pGroupResInfo->pRows == NULL) {
-    pGroupResInfo->pRows = taosArrayInit(100, POINTER_BYTES);
-  }
-
-  posList = calloc(size, sizeof(int32_t));
-  pTableQueryInfoList = malloc(POINTER_BYTES * size);
-
-  if (pTableQueryInfoList == NULL || posList == NULL || pGroupResInfo->pRows == NULL) {
-    qError("QInfo:%p failed alloc memory", pQInfo);
-    code = TSDB_CODE_QRY_OUT_OF_MEMORY;
-    goto _end;
-  }
-
-  int32_t numOfTables = 0;
-  for (int32_t i = 0; i < size; ++i) {
-    STableQueryInfo *item = taosArrayGetP(pTableList, i);
-    if (item->windowResInfo.size > 0) {
-      pTableQueryInfoList[numOfTables++] = item;
-    }
-  }
-
-  // there is no data in current group
-  // no need to merge results since only one table in each group
-  if (numOfTables == 0) {
-    goto _end;
-  }
-
-  SCompSupporter cs = {pTableQueryInfoList, posList, pRuntimeEnv->pQuery->order.order};
-
-  int32_t ret = tLoserTreeCreate(&pTree, numOfTables, &cs, tableResultComparFn);
-  if (ret != TSDB_CODE_SUCCESS) {
-    code = TSDB_CODE_QRY_OUT_OF_MEMORY;
-    goto _end;
-  }
-
-  int64_t lastTimestamp = ascQuery? INT64_MIN:INT64_MAX;
-  int64_t startt = taosGetTimestampMs();
-
-  while (1) {
-    if (isQueryKilled(pQInfo)) {
-      qDebug("QInfo:%p it is already killed, abort", pQInfo);
-      code = TSDB_CODE_TSC_QUERY_CANCELLED;
-      goto _end;
-    }
-
-    int32_t tableIndex = pTree->pNode[0].index;
-
-    SResultRowInfo *pWindowResInfo = &pTableQueryInfoList[tableIndex]->windowResInfo;
-    SResultRow  *pWindowRes = getResultRow(pWindowResInfo, cs.rowIndex[tableIndex]);
+    pQuery->rec.rows = doCopyToSData(&pQInfo->runtimeEnv, pGroupResInfo, TSDB_ORDER_ASC);
 
-    int64_t num = getNumOfResultWindowRes(pRuntimeEnv, pWindowRes);
-    if (num <= 0) {
-      cs.rowIndex[tableIndex] += 1;
-
-      if (cs.rowIndex[tableIndex] >= pWindowResInfo->size) {
-        cs.rowIndex[tableIndex] = -1;
-        if (--numOfTables == 0) { // all input sources are exhausted
-          break;
-        }
+    // current data are all dumped to result buffer, clear it
+    if (!hasRemainData(pGroupResInfo)) {
+      cleanupGroupResInfo(pGroupResInfo);
+      if (!incNextGroup(pGroupResInfo)) {
+        SET_STABLE_QUERY_OVER(pQInfo);
       }
-    } else {
-      assert((pWindowRes->win.skey >= lastTimestamp && ascQuery) || (pWindowRes->win.skey <= lastTimestamp && !ascQuery));
-
-      if (pWindowRes->win.skey != lastTimestamp) {
-        taosArrayPush(pGroupResInfo->pRows, &pWindowRes);
-        pWindowRes->numOfRows = (uint32_t) num;
     }
 
-      lastTimestamp = pWindowRes->win.skey;
-
-      // move to the next row of current entry
-      if ((++cs.rowIndex[tableIndex]) >= pWindowResInfo->size) {
-        cs.rowIndex[tableIndex] = -1;
-
-        // all input sources are exhausted
-        if ((--numOfTables) == 0) {
+    // enough results in data buffer, return
+    if (pQuery->rec.rows >= pQuery->rec.threshold) {
       break;
     }
   }
-    }
-
-    tLoserTreeAdjust(pTree, tableIndex + pTree->numOfEntries);
-  }
-
-  int64_t endt = taosGetTimestampMs();
-
-#ifdef _DEBUG_VIEW
-  displayInterResult(pQuery->sdata, pRuntimeEnv, pQuery->sdata[0]->num);
-#endif
-
-  qDebug("QInfo:%p result merge completed for group:%d, elapsed time:%" PRId64 " ms", pQInfo, pQInfo->groupIndex, endt - startt);
-
-  _end:
-  tfree(pTableQueryInfoList);
-  tfree(posList);
-  tfree(pTree);
-
-  return code;
 }
 
 static void updateTableQueryInfoForReverseScan(SQuery *pQuery, STableQueryInfo *pTableQueryInfo) {
@@ -3403,7 +3103,7 @@ static void updateTableQueryInfoForReverseScan(SQuery *pQuery, STableQueryInfo *
   pTableQueryInfo->cur.vgroupIndex = -1;
 
   // set the index at the end of time window
-  pTableQueryInfo->windowResInfo.curIndex = pTableQueryInfo->windowResInfo.size - 1;
+  pTableQueryInfo->resInfo.curIndex = pTableQueryInfo->resInfo.size - 1;
 }
 
 static void disableFuncInReverseScanImpl(SQueryRuntimeEnv* pRuntimeEnv, SResultRowInfo *pWindowResInfo, int32_t order) {
@@ -3438,7 +3138,7 @@ void disableFuncInReverseScan(SQInfo *pQInfo) {
   int32_t order = pQuery->order.order;
 
   // group by normal columns and interval query on normal table
-  SResultRowInfo *pWindowResInfo = &pRuntimeEnv->windowResInfo;
+  SResultRowInfo *pWindowResInfo = &pRuntimeEnv->resultRowInfo;
   if (pRuntimeEnv->groupbyColumn || QUERY_IS_INTERVAL_QUERY(pQuery)) {
     disableFuncInReverseScanImpl(pRuntimeEnv, pWindowResInfo, order);
   } else {  // for simple result of table query,
@@ -3502,7 +3202,7 @@ void resetDefaultResInfoOutputBuf(SQueryRuntimeEnv *pRuntimeEnv) {
 
   int32_t tid = 0;
   int64_t uid = 0;
-  SResultRow* pRow = doPrepareResultRowFromKey(pRuntimeEnv, &pRuntimeEnv->windowResInfo, (char *)&tid, sizeof(tid), true, uid);
+  SResultRow* pRow = doPrepareResultRowFromKey(pRuntimeEnv, &pRuntimeEnv->resultRowInfo, (char *)&tid, sizeof(tid), true, uid);
 
   for (int32_t i = 0; i < pQuery->numOfOutput; ++i) {
     SQLFunctionCtx *pCtx = &pRuntimeEnv->pCtx[i];
@@ -3629,7 +3329,7 @@ bool needRepeatScan(SQueryRuntimeEnv *pRuntimeEnv) {
   bool toContinue = false;
   if (pRuntimeEnv->groupbyColumn || QUERY_IS_INTERVAL_QUERY(pQuery)) {
     // for each group result, call the finalize function for each column
-    SResultRowInfo *pWindowResInfo = &pRuntimeEnv->windowResInfo;
+    SResultRowInfo *pWindowResInfo = &pRuntimeEnv->resultRowInfo;
 
     for (int32_t i = 0; i < pWindowResInfo->size; ++i) {
       SResultRow *pResult = getResultRow(pWindowResInfo, i);
@@ -3673,7 +3373,7 @@ static SQueryStatusInfo getQueryStatusInfo(SQueryRuntimeEnv *pRuntimeEnv, TSKEY
 
   SQueryStatusInfo info = {
       .status      = pQuery->status,
-      .windowIndex = pRuntimeEnv->windowResInfo.curIndex,
+      .windowIndex = pRuntimeEnv->resultRowInfo.curIndex,
       .lastKey     = start,
   };
 
@@ -3858,7 +3558,7 @@ void scanOneTableDataBlocks(SQueryRuntimeEnv *pRuntimeEnv, TSKEY start) {
       longjmp(pRuntimeEnv->env, terrno);
     }
 
-    pRuntimeEnv->windowResInfo.curIndex = qstatus.windowIndex;
+    pRuntimeEnv->resultRowInfo.curIndex = qstatus.windowIndex;
     setQueryStatus(pQuery, QUERY_NOT_COMPLETED);
     pRuntimeEnv->scanFlag = REPEAT_SCAN;
 
@@ -3889,7 +3589,7 @@ void finalizeQueryResult(SQueryRuntimeEnv *pRuntimeEnv) {
 
   if (pRuntimeEnv->groupbyColumn || QUERY_IS_INTERVAL_QUERY(pQuery)) {
     // for each group result, call the finalize function for each column
-    SResultRowInfo *pWindowResInfo = &pRuntimeEnv->windowResInfo;
+    SResultRowInfo *pWindowResInfo = &pRuntimeEnv->resultRowInfo;
     if (pRuntimeEnv->groupbyColumn) {
       closeAllResultRows(pWindowResInfo);
     }
@@ -3944,7 +3644,7 @@ static STableQueryInfo *createTableQueryInfo(SQuery* pQuery, void* pTable, bool
   // set more initial size of interval/groupby query
   if (QUERY_IS_INTERVAL_QUERY(pQuery) || groupbyColumn) {
     int32_t initialSize = 128;
-    int32_t code = initResultRowInfo(&pTableQueryInfo->windowResInfo, initialSize, TSDB_DATA_TYPE_INT);
+    int32_t code = initResultRowInfo(&pTableQueryInfo->resInfo, initialSize, TSDB_DATA_TYPE_INT);
     if (code != TSDB_CODE_SUCCESS) {
       return NULL;
     }
@@ -3960,7 +3660,7 @@ void destroyTableQueryInfoImpl(STableQueryInfo *pTableQueryInfo) {
   }
 
   tVariantDestroy(&pTableQueryInfo->tag);
-  cleanupResultRowInfo(&pTableQueryInfo->windowResInfo);
+  cleanupResultRowInfo(&pTableQueryInfo->resInfo);
 }
 
 /**
@@ -3971,7 +3671,7 @@ void destroyTableQueryInfoImpl(STableQueryInfo *pTableQueryInfo) {
 void setExecutionContext(SQInfo *pQInfo, int32_t groupIndex, TSKEY nextKey) {
   SQueryRuntimeEnv *pRuntimeEnv     = &pQInfo->runtimeEnv;
   STableQueryInfo  *pTableQueryInfo = pRuntimeEnv->pQuery->current;
-  SResultRowInfo   *pWindowResInfo  = &pRuntimeEnv->windowResInfo;
+  SResultRowInfo   *pWindowResInfo  = &pRuntimeEnv->resultRowInfo;
 
   // lastKey needs to be updated
   pTableQueryInfo->lastKey = nextKey;
@@ -4168,7 +3868,7 @@ void setIntervalQueryRange(SQInfo *pQInfo, TSKEY key) {
      * operations involve.
      */
     STimeWindow     w = TSWINDOW_INITIALIZER;
-    SResultRowInfo *pWindowResInfo = &pTableQueryInfo->windowResInfo;
+    SResultRowInfo *pWindowResInfo = &pTableQueryInfo->resInfo;
 
     TSKEY sk = MIN(win.skey, win.ekey);
     TSKEY ek = MAX(win.skey, win.ekey);
@@ -4211,73 +3911,63 @@ bool needPrimaryTimestampCol(SQuery *pQuery, SDataBlockInfo *pDataBlockInfo) {
   return loadPrimaryTS;
 }
 
-static int32_t doCopyToSData(SQInfo *pQInfo, SResultRow **pRows, int32_t numOfRows, int32_t *index, int32_t orderType) {
-  SQueryRuntimeEnv *pRuntimeEnv = &pQInfo->runtimeEnv;
-  SQuery *          pQuery = pRuntimeEnv->pQuery;
+static int32_t doCopyToSData(SQueryRuntimeEnv* pRuntimeEnv, SGroupResInfo* pGroupResInfo, int32_t orderType) {
+  void* qinfo = GET_QINFO_ADDR(pRuntimeEnv);
+  SQuery *pQuery = pRuntimeEnv->pQuery;
+
+  int32_t numOfRows = getNumOfTotalRes(pGroupResInfo);
+  int32_t numOfResult = pQuery->rec.rows; // there are already exists result rows
 
-  int32_t numOfResult = 0;
   int32_t start = 0;
   int32_t step = -1;
 
-  qDebug("QInfo:%p start to copy data from windowResInfo to query buf", pQInfo);
+  qDebug("QInfo:%p start to copy data from windowResInfo to output buf", qinfo);
   if (orderType == TSDB_ORDER_ASC) {
-    start = (*index);
+    start = pGroupResInfo->index;
     step = 1;
   } else {  // desc order copy all data
-    start = numOfRows - (*index) - 1;
+    start = numOfRows - pGroupResInfo->index - 1;
     step = -1;
   }
 
-  SGroupResInfo* pGroupResInfo = &pQInfo->groupResInfo;
-
   for (int32_t i = start; (i < numOfRows) && (i >= 0); i += step) {
-    if (pRows[i]->numOfRows == 0) {
-      (*index) += 1;
-      pGroupResInfo->rowId = 0;
+    SResultRow* pRow = taosArrayGetP(pGroupResInfo->pRows, i);
+    if (pRow->numOfRows == 0) {
+      pGroupResInfo->index += 1;
       continue;
     }
 
-    int32_t numOfRowsToCopy = pRows[i]->numOfRows - pGroupResInfo->rowId;
-    int32_t oldOffset = pGroupResInfo->rowId;
+    int32_t numOfRowsToCopy = pRow->numOfRows;
 
-    /*
-     * current output space is not enough to accommodate all data of this page, only partial results
-     * will be copied to SQuery object's result buffer
-     */
-    if (numOfRowsToCopy > pQuery->rec.capacity - numOfResult) {
-      numOfRowsToCopy = (int32_t) pQuery->rec.capacity - numOfResult;
-      pGroupResInfo->rowId += numOfRowsToCopy;
-    } else {
-      pGroupResInfo->rowId = 0;
-      (*index) += 1;
+    //current output space is not enough to accommodate all data of this page, prepare more space
+    if (numOfRowsToCopy > (pQuery->rec.capacity - numOfResult)) {
+      int32_t newSize = pQuery->rec.capacity + (numOfRowsToCopy - numOfResult);
+      expandBuffer(pRuntimeEnv, newSize, GET_QINFO_ADDR(pRuntimeEnv));
     }
 
-    tFilePage *page = getResBufPage(pRuntimeEnv->pResultBuf, pRows[i]->pageId);
+    pGroupResInfo->index += 1;
 
+    tFilePage *page = getResBufPage(pRuntimeEnv->pResultBuf, pRow->pageId);
     for (int32_t j = 0; j < pQuery->numOfOutput; ++j) {
       int32_t size = pRuntimeEnv->pCtx[j].outputBytes;
 
       char *out = pQuery->sdata[j]->data + numOfResult * size;
-      char *in  = getPosInResultPage(pRuntimeEnv, j, pRows[i], page);
-      memcpy(out, in + oldOffset * size, size * numOfRowsToCopy);
+      char *in  = getPosInResultPage(pRuntimeEnv, j, pRow, page);
+      memcpy(out, in, size * numOfRowsToCopy);
     }
 
     numOfResult += numOfRowsToCopy;
-    if (numOfResult == pQuery->rec.capacity) {
+    if (numOfResult == pQuery->rec.capacity) {  // output buffer is full
       break;
     }
   }
 
-  qDebug("QInfo:%p copy data to query buf completed", pQInfo);
-
-#ifdef _DEBUG_VIEW
-  displayInterResult(pQuery->sdata, pRuntimeEnv, numOfResult);
-#endif
+  qDebug("QInfo:%p copy data to query buf completed", qinfo);
   return numOfResult;
 }
 
 /**
- * copyFromWindowResToSData support copy data in ascending/descending order
+ * copyToOutputBuf support copy data in ascending/descending order
  * For interval query of both super table and table, copy the data in ascending order, since the output results are
  * ordered in SWindowResutl already. While handling the group by query for both table and super table,
  * all group result are completed already.
@@ -4285,14 +3975,17 @@ static int32_t doCopyToSData(SQInfo *pQInfo, SResultRow **pRows, int32_t numOfRo
  * @param pQInfo
  * @param result
  */
-void copyFromWindowResToSData(SQInfo *pQInfo, SResultRowInfo *pResultInfo) {
+void copyToOutputBuf(SQInfo *pQInfo, SResultRowInfo *pResultInfo) {
   SQuery        *pQuery = pQInfo->runtimeEnv.pQuery;
+  SGroupResInfo *pGroupResInfo = &pQInfo->groupResInfo;
 
-  int32_t orderType = (pQuery->pGroupbyExpr != NULL) ? pQuery->pGroupbyExpr->orderType : TSDB_ORDER_ASC;
-  int32_t numOfResult = doCopyToSData(pQInfo, pResultInfo->pResult, pResultInfo->size, &pQInfo->groupIndex, orderType);
+  assert(pQuery->rec.rows == 0 && pGroupResInfo->currentGroup <= pGroupResInfo->totalGroup);
+  if (pGroupResInfo->index >= taosArrayGetSize(pGroupResInfo->pRows)) {
+    return;
+  }
 
-  pQuery->rec.rows += numOfResult;
-  assert(pQuery->rec.rows <= pQuery->rec.capacity);
+  int32_t orderType = (pQuery->pGroupbyExpr != NULL) ? pQuery->pGroupbyExpr->orderType : TSDB_ORDER_ASC;
+  pQuery->rec.rows = doCopyToSData(&pQInfo->runtimeEnv, pGroupResInfo, orderType);
 }
 
 static void updateWindowResNumOfRes(SQueryRuntimeEnv *pRuntimeEnv) {
@@ -4303,8 +3996,8 @@ static void updateWindowResNumOfRes(SQueryRuntimeEnv *pRuntimeEnv) {
     return;
   }
 
-  for (int32_t i = 0; i < pRuntimeEnv->windowResInfo.size; ++i) {
-    SResultRow *pResult = pRuntimeEnv->windowResInfo.pResult[i];
+  for (int32_t i = 0; i < pRuntimeEnv->resultRowInfo.size; ++i) {
+    SResultRow *pResult = pRuntimeEnv->resultRowInfo.pResult[i];
 
     for (int32_t j = 0; j < pQuery->numOfOutput; ++j) {
       int32_t functionId = pRuntimeEnv->pCtx[j].functionId;
@@ -4323,7 +4016,7 @@ static void stableApplyFunctionsOnBlock(SQueryRuntimeEnv *pRuntimeEnv, SDataBloc
   SQuery *         pQuery = pRuntimeEnv->pQuery;
   STableQueryInfo* pTableQueryInfo = pQuery->current;
 
-  SResultRowInfo * pResultRowInfo = &pTableQueryInfo->windowResInfo;
+  SResultRowInfo * pResultRowInfo = &pTableQueryInfo->resInfo;
   pQuery->pos = QUERY_IS_ASC_QUERY(pQuery)? 0 : pDataBlockInfo->rows - 1;
 
   if (pQuery->numOfFilterCols > 0 || pRuntimeEnv->pTsBuf != NULL || pRuntimeEnv->groupbyColumn) {
@@ -4337,7 +4030,7 @@ static void stableApplyFunctionsOnBlock(SQueryRuntimeEnv *pRuntimeEnv, SDataBloc
   }
 }
 
-bool hasNotReturnedResults(SQueryRuntimeEnv* pRuntimeEnv) {
+bool hasNotReturnedResults(SQueryRuntimeEnv* pRuntimeEnv, SGroupResInfo* pGroupResInfo) {
   SQuery *pQuery = pRuntimeEnv->pQuery;
   SFillInfo *pFillInfo = pRuntimeEnv->pFillInfo;
 
@@ -4360,18 +4053,13 @@ bool hasNotReturnedResults(SQueryRuntimeEnv* pRuntimeEnv) {
      * set is the FIRST result block, the gap between the start time of query time window and the timestamp of the
      * first result row in the actual result set will fill nothing.
      */
-    if (Q_STATUS_EQUAL(pQuery->status, QUERY_COMPLETED)) {
     int32_t numOfTotal = (int32_t)getNumOfResultsAfterFillGap(pFillInfo, pQuery->window.ekey, (int32_t)pQuery->rec.capacity);
     return numOfTotal > 0;
-    }
-
-  } else {
-    // there are results waiting for returned to client.
-    if (Q_STATUS_EQUAL(pQuery->status, QUERY_COMPLETED) &&
-        (pRuntimeEnv->groupbyColumn || QUERY_IS_INTERVAL_QUERY(pQuery)) &&
-        (pRuntimeEnv->windowResInfo.size > 0)) {
-      return true;
-    }
+  } else { // there are results waiting for returned to client.
+    if (Q_STATUS_EQUAL(pQuery->status, QUERY_COMPLETED) && hasRemainData(pGroupResInfo) &&
+        (pRuntimeEnv->groupbyColumn || QUERY_IS_INTERVAL_QUERY(pQuery))) {
+      return true;
+    }
   }
 
   return false;
@@ -4429,15 +4117,15 @@ static void doCopyQueryResultToMsg(SQInfo *pQInfo, int32_t numOfRows, char *data
         setQueryStatus(pQuery, QUERY_OVER);
       }
     } else {
-      if (!hasNotReturnedResults(&pQInfo->runtimeEnv)) {
+      if (!hasNotReturnedResults(&pQInfo->runtimeEnv, &pQInfo->groupResInfo)) {
         setQueryStatus(pQuery, QUERY_OVER);
       }
     }
   }
 }
 
-int32_t doFillGapsInResults(SQueryRuntimeEnv* pRuntimeEnv, tFilePage **pDst, int32_t *numOfFilled) {
-  SQInfo* pQInfo = GET_QINFO_ADDR(pRuntimeEnv);
+int32_t doFillGapsInResults(SQueryRuntimeEnv* pRuntimeEnv, tFilePage **pDst) {
+  SQInfo *pQInfo = GET_QINFO_ADDR(pRuntimeEnv);
   SQuery *pQuery = pRuntimeEnv->pQuery;
   SFillInfo* pFillInfo = pRuntimeEnv->pFillInfo;
 
@@ -4456,7 +4144,7 @@ int32_t doFillGapsInResults(SQueryRuntimeEnv* pRuntimeEnv, tFilePage **pDst, int
              pQInfo, pFillInfo->numOfRows, ret, pQuery->limit.offset, ret - pQuery->limit.offset, 0);
 
       ret -= (int32_t)pQuery->limit.offset;
-      for (int32_t i = 0; i < pQuery->numOfOutput; ++i) {
+      for (int32_t i = 0; i < pQuery->numOfOutput; ++i) { //???pExpr1 or pExpr2
         memmove(pDst[i]->data, pDst[i]->data + pQuery->pExpr1[i].bytes * pQuery->limit.offset,
                 ret * pQuery->pExpr1[i].bytes);
       }
@@ -4469,17 +4157,18 @@ int32_t doFillGapsInResults(SQueryRuntimeEnv* pRuntimeEnv, tFilePage **pDst, int
           pQuery->limit.offset - ret);
 
       pQuery->limit.offset -= ret;
-      pQuery->rec.rows = 0;
       ret = 0;
     }
 
-    if (!hasNotReturnedResults(pRuntimeEnv)) {
-      return ret;
+    // no data in current data after fill
+    int32_t numOfTotal = (int32_t)getNumOfResultsAfterFillGap(pFillInfo, pFillInfo->end, (int32_t)pQuery->rec.capacity);
+    if (numOfTotal == 0) {
+      return 0;
     }
   }
 }
 
-static void queryCostStatis(SQInfo *pQInfo) {
+void queryCostStatis(SQInfo *pQInfo) {
   SQueryRuntimeEnv *pRuntimeEnv = &pQInfo->runtimeEnv;
   SQueryCostInfo *pSummary = &pRuntimeEnv->summary;
 
@@ -4637,7 +4326,7 @@ void skipBlocks(SQueryRuntimeEnv *pRuntimeEnv) {
 
 static TSKEY doSkipIntervalProcess(SQueryRuntimeEnv* pRuntimeEnv, STimeWindow* win, SDataBlockInfo* pBlockInfo, STableQueryInfo* pTableQueryInfo) {
   SQuery *pQuery = pRuntimeEnv->pQuery;
-  SResultRowInfo *pWindowResInfo = &pRuntimeEnv->windowResInfo;
+  SResultRowInfo *pWindowResInfo = &pRuntimeEnv->resultRowInfo;
 
   assert(pQuery->limit.offset == 0);
   STimeWindow tw = *win;
@@ -4665,10 +4354,10 @@ static TSKEY doSkipIntervalProcess(SQueryRuntimeEnv* pRuntimeEnv, STimeWindow* w
     TSKEY key = pTableQueryInfo->win.skey;
 
     pWindowResInfo->prevSKey = tw.skey;
-    int32_t index = pRuntimeEnv->windowResInfo.curIndex;
+    int32_t index = pRuntimeEnv->resultRowInfo.curIndex;
 
     int32_t numOfRes = tableApplyFunctionsOnBlock(pRuntimeEnv, pBlockInfo, NULL, binarySearchForKey, pDataBlock);
-    pRuntimeEnv->windowResInfo.curIndex = index;  // restore the window index
+    pRuntimeEnv->resultRowInfo.curIndex = index;  // restore the window index
 
     qDebug("QInfo:%p check data block, brange:%" PRId64 "-%" PRId64 ", numOfRows:%d, numOfRes:%d, lastKey:%" PRId64,
            GET_QINFO_ADDR(pRuntimeEnv), pBlockInfo->window.skey, pBlockInfo->window.ekey, pBlockInfo->rows, numOfRes,
@@ -4704,12 +4393,12 @@ static bool skipTimeInterval(SQueryRuntimeEnv *pRuntimeEnv, TSKEY* start) {
    *    pQuery->limit.offset times. Since hole exists, pQuery->interval.interval*pQuery->limit.offset value is
    *    not valid. otherwise, we only forward pQuery->limit.offset number of points
    */
-  assert(pRuntimeEnv->windowResInfo.prevSKey == TSKEY_INITIAL_VAL);
+  assert(pRuntimeEnv->resultRowInfo.prevSKey == TSKEY_INITIAL_VAL);
 
   STimeWindow w = TSWINDOW_INITIALIZER;
   bool ascQuery = QUERY_IS_ASC_QUERY(pQuery);
 
-  SResultRowInfo *pWindowResInfo = &pRuntimeEnv->windowResInfo;
+  SResultRowInfo *pWindowResInfo = &pRuntimeEnv->resultRowInfo;
   STableQueryInfo *pTableQueryInfo = pQuery->current;
 
   SDataBlockInfo blockInfo = SDATA_BLOCK_INITIALIZER;
@@ -4890,7 +4579,6 @@ int32_t doInitQInfo(SQInfo *pQInfo, STSBuf *pTsBuf, SArray* prevResult, void *ts
   pRuntimeEnv->topBotQuery = isTopBottomQuery(pQuery);
   pRuntimeEnv->hasTagResults = hasTagValOutput(pQuery);
   pRuntimeEnv->timeWindowInterpo = timeWindowInterpoRequired(pQuery);
-
   pRuntimeEnv->prevResult = prevResult;
 
   setScanLimitationByResultBuffer(pQuery);
@@ -4902,6 +4590,7 @@ int32_t doInitQInfo(SQInfo *pQInfo, STSBuf *pTsBuf, SArray* prevResult, void *ts
 
   pQInfo->tsdb = tsdb;
   pQInfo->vgId = vgId;
+  pQInfo->groupResInfo.totalGroup = isSTableQuery? GET_NUM_OF_TABLEGROUP(pQInfo):0;
 
   pRuntimeEnv->pQuery = pQuery;
   pRuntimeEnv->pTsBuf = pTsBuf;
@@ -4934,7 +4623,7 @@ int32_t doInitQInfo(SQInfo *pQInfo, STSBuf *pTsBuf, SArray* prevResult, void *ts
         type = TSDB_DATA_TYPE_INT;  // group id
       }
 
-      code = initResultRowInfo(&pRuntimeEnv->windowResInfo, 8, type);
+      code = initResultRowInfo(&pRuntimeEnv->resultRowInfo, 8, type);
       if (code != TSDB_CODE_SUCCESS) {
         return code;
       }
@@ -4954,14 +4643,14 @@ int32_t doInitQInfo(SQInfo *pQInfo, STSBuf *pTsBuf, SArray* prevResult, void *ts
       type = TSDB_DATA_TYPE_TIMESTAMP;
     }
 
-    code = initResultRowInfo(&pRuntimeEnv->windowResInfo, numOfResultRows, type);
+    code = initResultRowInfo(&pRuntimeEnv->resultRowInfo, numOfResultRows, type);
     if (code != TSDB_CODE_SUCCESS) {
       return code;
     }
   }
 
   // create runtime environment
-  code = setupQueryRuntimeEnv(pRuntimeEnv, (int32_t) pQInfo->tableGroupInfo.numOfTables, pQuery->order.order);
+  code = setupQueryRuntimeEnv(pRuntimeEnv, (int32_t) pQInfo->tableGroupInfo.numOfTables, pQuery->order.order, pQInfo->vgId);
   if (code != TSDB_CODE_SUCCESS) {
     return code;
   }
@@ -5072,7 +4761,7 @@ static int64_t scanMultiTableDataBlocks(SQInfo *pQInfo) {
     SDataStatis *pStatis = NULL;
     SArray      *pDataBlock = NULL;
 
-    int32_t ret = loadDataBlockOnDemand(pRuntimeEnv, &pQuery->current->windowResInfo, pQueryHandle, &blockInfo, &pStatis, &pDataBlock, &status);
+    int32_t ret = loadDataBlockOnDemand(pRuntimeEnv, &pQuery->current->resInfo, pQueryHandle, &blockInfo, &pStatis, &pDataBlock, &status);
     if (ret != TSDB_CODE_SUCCESS) {
       break;
     }
@@ -5366,7 +5055,7 @@ static void sequentialTableProcess(SQInfo *pQInfo) {
       taosArrayDestroy(s);
 
       // no results generated for current group, continue to try the next group
-      SResultRowInfo *pWindowResInfo = &pRuntimeEnv->windowResInfo;
+      SResultRowInfo *pWindowResInfo = &pRuntimeEnv->resultRowInfo;
       if (pWindowResInfo->size <= 0) {
         continue;
       }
@@ -5383,17 +5072,18 @@ static void sequentialTableProcess(SQInfo *pQInfo) {
 
       qDebug("QInfo:%p generated groupby columns results %d rows for group %d completed", pQInfo, pWindowResInfo->size,
              pQInfo->groupIndex);
-      int32_t currentGroupIndex = pQInfo->groupIndex;
 
       pQuery->rec.rows = 0;
-      pQInfo->groupIndex = 0;
-
-      ensureOutputBufferSimple(pRuntimeEnv, pWindowResInfo->size);
-      copyFromWindowResToSData(pQInfo, pWindowResInfo);
+      if (pWindowResInfo->size > pQuery->rec.capacity) {
+        expandBuffer(pRuntimeEnv, pWindowResInfo->size, pQInfo);
+      }
 
-      pQInfo->groupIndex = currentGroupIndex;  // restore the group index
+      initGroupResInfo(&pQInfo->groupResInfo, &pRuntimeEnv->resultRowInfo, 0);
+      copyToOutputBuf(pQInfo, pWindowResInfo);
       assert(pQuery->rec.rows == pWindowResInfo->size);
-      resetResultRowInfo(pRuntimeEnv, &pRuntimeEnv->windowResInfo);
+
+      resetResultRowInfo(pRuntimeEnv, &pRuntimeEnv->resultRowInfo);
+      cleanupGroupResInfo(&pQInfo->groupResInfo);
       break;
     }
   } else if (pRuntimeEnv->queryWindowIdentical && pRuntimeEnv->pTsBuf == NULL && !isTsCompQuery(pQuery)) {
@@ -5449,7 +5139,7 @@ static void sequentialTableProcess(SQInfo *pQInfo) {
         setTagVal(pRuntimeEnv, pQuery->current->pTable);
       }
 
-      if (pQuery->prjInfo.vgroupLimit > 0 && pQuery->current->windowResInfo.size > pQuery->prjInfo.vgroupLimit) {
+      if (pQuery->prjInfo.vgroupLimit > 0 && pQuery->current->resInfo.size > pQuery->prjInfo.vgroupLimit) {
         pQuery->current->lastKey =
                 QUERY_IS_ASC_QUERY(pQuery) ? blockInfo.window.ekey + step : blockInfo.window.skey + step;
         continue;
@@ -5472,7 +5162,7 @@ static void sequentialTableProcess(SQInfo *pQInfo) {
       SDataStatis *pStatis = NULL;
       SArray      *pDataBlock = NULL;
 
-      int32_t ret = loadDataBlockOnDemand(pRuntimeEnv, &pQuery->current->windowResInfo, pQueryHandle, &blockInfo,
+      int32_t ret = loadDataBlockOnDemand(pRuntimeEnv, &pQuery->current->resInfo, pQueryHandle, &blockInfo,
                                           &pStatis, &pDataBlock, &status);
       if (ret != TSDB_CODE_SUCCESS) {
         break;
@@ -5484,7 +5174,7 @@ static void sequentialTableProcess(SQInfo *pQInfo) {
         continue;
       }
 
-      ensureOutputBuffer(pRuntimeEnv, &blockInfo);
+      ensureOutputBuffer(pRuntimeEnv, blockInfo.rows);
       int64_t prev = getNumOfResult(pRuntimeEnv);
 
       pQuery->pos = QUERY_IS_ASC_QUERY(pQuery) ? 0 : blockInfo.rows - 1;
@@ -5498,7 +5188,7 @@ static void sequentialTableProcess(SQInfo *pQInfo) {
       pQuery->rec.rows = getNumOfResult(pRuntimeEnv);
 
       int64_t inc = pQuery->rec.rows - prev;
-      pQuery->current->windowResInfo.size += (int32_t) inc;
+      pQuery->current->resInfo.size += (int32_t) inc;
 
       // the flag may be set by tableApplyFunctionsOnBlock, clear it here
       CLEAR_QUERY_STATUS(pQuery, QUERY_COMPLETED);
@@ -5516,8 +5206,7 @@ static void sequentialTableProcess(SQInfo *pQInfo) {
       } else {
         // the limitation of output result is reached, set the query completed
         skipResults(pRuntimeEnv);
-        if (limitOperator(pRuntimeEnv)) {
-          setQueryStatus(pQuery, QUERY_COMPLETED);
+        if (limitOperator(pQuery, pQInfo)) {
           SET_STABLE_QUERY_OVER(pQInfo);
           break;
         }
@@ -5540,8 +5229,8 @@ static void sequentialTableProcess(SQInfo *pQInfo) {
      * If the subgroup index is larger than 0, results generated by group by tbname,k is existed.
      * we need to return it to client in the first place.
      */
-    if (pQInfo->groupIndex > 0) {
-      copyFromWindowResToSData(pQInfo, &pRuntimeEnv->windowResInfo);
+    if (hasRemainData(&pQInfo->groupResInfo)) {
+      copyToOutputBuf(pQInfo, &pRuntimeEnv->resultRowInfo);
       pQuery->rec.total += pQuery->rec.rows;
 
       if (pQuery->rec.rows > 0) {
@@ -5555,7 +5244,7 @@ static void sequentialTableProcess(SQInfo *pQInfo) {
     }
 
     resetDefaultResInfoOutputBuf(pRuntimeEnv);
-    resetResultRowInfo(pRuntimeEnv, &pRuntimeEnv->windowResInfo);
+    resetResultRowInfo(pRuntimeEnv, &pRuntimeEnv->resultRowInfo);
 
     SArray *group = GET_TABLEGROUP(pQInfo, 0);
     assert(taosArrayGetSize(group) == pQInfo->tableqinfoGroupInfo.numOfTables &&
@@ -5584,7 +5273,7 @@ static void sequentialTableProcess(SQInfo *pQInfo) {
       skipResults(pRuntimeEnv);
 
       // the limitation of output result is reached, set the query completed
-      if (limitOperator(pRuntimeEnv)) {
+      if (limitOperator(pQuery, pQInfo)) {
         SET_STABLE_QUERY_OVER(pQInfo);
         break;
       }
@@ -5709,11 +5398,11 @@ static void doCloseAllTimeWindow(SQInfo *pQInfo) {
       size_t num = taosArrayGetSize(group);
       for (int32_t j = 0; j < num; ++j) {
         STableQueryInfo* item = taosArrayGetP(group, j);
-        closeAllResultRows(&item->windowResInfo);
+        closeAllResultRows(&item->resInfo);
       }
     }
   } else {  // close results for group result
-    closeAllResultRows(&pQInfo->runtimeEnv.windowResInfo);
+    closeAllResultRows(&pQInfo->runtimeEnv.resultRowInfo);
   }
 }
 
@@ -5721,18 +5410,14 @@ static void multiTableQueryProcess(SQInfo *pQInfo) {
   SQueryRuntimeEnv *pRuntimeEnv = &pQInfo->runtimeEnv;
   SQuery           *pQuery = pRuntimeEnv->pQuery;
 
-  if (pQInfo->groupIndex > 0) {
-    /*
-     * if the groupIndex > 0, the query process must be completed yet, we only need to
-     * copy the data into output buffer
-     */
+  if (Q_STATUS_EQUAL(pQuery->status, QUERY_COMPLETED)) {
     if (QUERY_IS_INTERVAL_QUERY(pQuery)) {
       copyResToQueryResultBuf(pQInfo, pQuery);
     } else {
-      copyFromWindowResToSData(pQInfo, &pRuntimeEnv->windowResInfo);
+      copyToOutputBuf(pQInfo, &pRuntimeEnv->resultRowInfo);
     }
 
-    qDebug("QInfo:%p current:%"PRId64", total:%"PRId64"", pQInfo, pQuery->rec.rows, pQuery->rec.total);
+    qDebug("QInfo:%p current:%"PRId64", total:%"PRId64, pQInfo, pQuery->rec.rows, pQuery->rec.total);
     return;
   }
 
@@ -5774,18 +5459,10 @@ static void multiTableQueryProcess(SQInfo *pQInfo) {
   }
 
   if (QUERY_IS_INTERVAL_QUERY(pQuery) || isSumAvgRateQuery(pQuery)) {
-    int32_t code = mergeGroupResult(pQInfo);
-    if (code == TSDB_CODE_SUCCESS) {
     copyResToQueryResultBuf(pQInfo, pQuery);
-
-#ifdef _DEBUG_VIEW
-      displayInterResult(pQuery->sdata, pRuntimeEnv, pQuery->sdata[0]->num);
-#endif
-    } else {  // set the error code
-      pQInfo->code = code;
-    }
   } else {  // not a interval query
-    copyFromWindowResToSData(pQInfo, &pRuntimeEnv->windowResInfo);
+    initGroupResInfo(&pQInfo->groupResInfo, &pRuntimeEnv->resultRowInfo, 0);
+    copyToOutputBuf(pQInfo, &pRuntimeEnv->resultRowInfo);
   }
 
   // handle the limitation of output buffer
@@ -5889,7 +5566,7 @@ static void tableAggregationProcess(SQInfo *pQInfo, STableQueryInfo* pTableInfo)
 
   // TODO limit/offset refactor to be one operator
   skipResults(pRuntimeEnv);
-  limitOperator(pRuntimeEnv);
+  limitOperator(pQuery, pQInfo);
 }
 
 static void tableProjectionProcess(SQInfo *pQInfo, STableQueryInfo* pTableInfo) {
@@ -5931,7 +5608,7 @@ static void tableProjectionProcess(SQInfo *pQInfo, STableQueryInfo* pTableInfo)
     resetDefaultResInfoOutputBuf(pRuntimeEnv);
   }
 
-  limitOperator(pRuntimeEnv);
+  limitOperator(pQuery, pQInfo);
   if (Q_STATUS_EQUAL(pQuery->status, QUERY_RESBUF_FULL)) {
     qDebug("QInfo:%p query paused due to output limitation, next qrange:%" PRId64 "-%" PRId64, pQInfo,
         pQuery->current->lastKey, pQuery->window.ekey);
@@ -5945,6 +5622,40 @@ static void tableProjectionProcess(SQInfo *pQInfo, STableQueryInfo* pTableInfo)
   }
 }
 
+static void copyAndFillResult(SQInfo* pQInfo) {
+  SQueryRuntimeEnv* pRuntimeEnv = &pQInfo->runtimeEnv;
+  SQuery* pQuery = pRuntimeEnv->pQuery;
+
+  while(1) {
+    copyToOutputBuf(pQInfo, &pRuntimeEnv->resultRowInfo);
+    doSecondaryArithmeticProcess(pQuery);
+
+    TSKEY lastKey = 0;
+    if (!hasRemainData(&pQInfo->groupResInfo)) {
+      lastKey = pQuery->window.ekey;
+    } else {
+      lastKey = ((TSKEY*)pQuery->sdata[0]->data)[pQuery->rec.rows - 1];
+    }
+
+    assert(lastKey <= pQuery->window.ekey);
+
+    taosFillSetStartInfo(pRuntimeEnv->pFillInfo, (int32_t)pQuery->rec.rows, lastKey);
+    taosFillSetDataBlockFromFilePage(pRuntimeEnv->pFillInfo, (const tFilePage **)pQuery->sdata);
+
+    pQuery->rec.rows = doFillGapsInResults(pRuntimeEnv, (tFilePage **)pQuery->sdata);
+
+    if (pQuery->rec.rows > 0) {
+      limitOperator(pQuery, pQInfo);
+      break;
+    }
+
+    // here the pQuery->rec.rows == 0
+    if (!hasRemainData(&pQInfo->groupResInfo) && !taosFillHasMoreResults(pRuntimeEnv->pFillInfo)) {
+      break;
+    }
+  }
+}
+
 // handle time interval query on table
 static void tableIntervalProcess(SQInfo *pQInfo, STableQueryInfo* pTableInfo) {
   SQueryRuntimeEnv *pRuntimeEnv = &(pQInfo->runtimeEnv);
@@ -5968,69 +5679,56 @@ static void tableIntervalProcess(SQInfo *pQInfo, STableQueryInfo* pTableInfo) {
   pQuery->rec.rows = 0;
 
   // not fill or no result generated during this query
-  if (pQuery->fillType == TSDB_FILL_NONE || pRuntimeEnv->windowResInfo.size == 0 || isPointInterpoQuery(pQuery)) {
+  if (pQuery->fillType == TSDB_FILL_NONE || pRuntimeEnv->resultRowInfo.size == 0 || isPointInterpoQuery(pQuery)) {
     // all data scanned, the group by normal column can return
-    int32_t numOfClosed = numOfClosedResultRows(&pRuntimeEnv->windowResInfo);
+    int32_t numOfClosed = numOfClosedResultRows(&pRuntimeEnv->resultRowInfo);
     if (pQuery->limit.offset > numOfClosed) {
       return;
     }
 
-    pQInfo->groupIndex = (int32_t) pQuery->limit.offset;
-
-    copyFromWindowResToSData(pQInfo, &pRuntimeEnv->windowResInfo);
+    initGroupResInfo(&pQInfo->groupResInfo, &pRuntimeEnv->resultRowInfo, pQuery->limit.offset);
+    copyToOutputBuf(pQInfo, &pRuntimeEnv->resultRowInfo);
     doSecondaryArithmeticProcess(pQuery);
 
-    limitOperator(pRuntimeEnv);
+    limitOperator(pQuery, pQInfo);
   } else {
-
-    copyFromWindowResToSData(pQInfo, &pRuntimeEnv->windowResInfo);
-    doSecondaryArithmeticProcess(pQuery);
-
-    taosFillSetStartInfo(pRuntimeEnv->pFillInfo, (int32_t)pQuery->rec.rows, pQuery->window.ekey);
-    taosFillSetDataBlockFromFilePage(pRuntimeEnv->pFillInfo, (const tFilePage **)pQuery->sdata);
-
-    int32_t numOfFilled = 0;
-    pQuery->rec.rows = doFillGapsInResults(pRuntimeEnv, (tFilePage **)pQuery->sdata, &numOfFilled);
-
-    if (pQuery->rec.rows > 0 || Q_STATUS_EQUAL(pQuery->status, QUERY_COMPLETED)) {
-      limitOperator(pRuntimeEnv);
-    }
+    initGroupResInfo(&pQInfo->groupResInfo, &pRuntimeEnv->resultRowInfo, 0);
+    return copyAndFillResult(pQInfo);
   }
 }
 
-static void tableQueryImpl(SQInfo *pQInfo) {
+void tableQueryImpl(SQInfo *pQInfo) {
   SQueryRuntimeEnv *pRuntimeEnv = &pQInfo->runtimeEnv;
   SQuery *          pQuery = pRuntimeEnv->pQuery;
 
-  if (hasNotReturnedResults(pRuntimeEnv)) {
+  if (hasNotReturnedResults(pRuntimeEnv, &pQInfo->groupResInfo)) {
     if (pQuery->fillType != TSDB_FILL_NONE && !isPointInterpoQuery(pQuery)) {
       /*
        * There are remain results that are not returned due to result interpolation
        * So, we do keep in this procedure instead of launching retrieve procedure for next results.
        */
-      int32_t numOfFilled = 0;
-      pQuery->rec.rows = doFillGapsInResults(pRuntimeEnv, (tFilePage **)pQuery->sdata, &numOfFilled);
-
+      pQuery->rec.rows = doFillGapsInResults(pRuntimeEnv, (tFilePage **)pQuery->sdata);
       if (pQuery->rec.rows > 0) {
-        limitOperator(pRuntimeEnv);
+        limitOperator(pQuery, pQInfo);
+        qDebug("QInfo:%p current:%" PRId64 " returned, total:%" PRId64, pQInfo, pQuery->rec.rows, pQuery->rec.total);
+      } else {
+        return copyAndFillResult(pQInfo);
       }
 
-      qDebug("QInfo:%p current:%" PRId64 " returned, total:%" PRId64, pQInfo, pQuery->rec.rows, pQuery->rec.total);
     } else {
       pQuery->rec.rows = 0;
-      assert(pRuntimeEnv->windowResInfo.size > 0);
+      assert(pRuntimeEnv->resultRowInfo.size > 0);
+      copyToOutputBuf(pQInfo, &pRuntimeEnv->resultRowInfo);
+      doSecondaryArithmeticProcess(pQuery);
 
-      if (pQInfo->groupIndex < pRuntimeEnv->windowResInfo.size) {
-        copyFromWindowResToSData(pQInfo, &pRuntimeEnv->windowResInfo);
+      if (pQuery->rec.rows > 0) {
+        limitOperator(pQuery, pQInfo);
       }
 
       if (pQuery->rec.rows > 0) {
         qDebug("QInfo:%p %" PRId64 " rows returned from group results, total:%" PRId64 "", pQInfo, pQuery->rec.rows,
                pQuery->rec.total);
-      }
-
-      // there are not data remains
-      if (pQuery->rec.rows <= 0 || pRuntimeEnv->windowResInfo.size <= pQInfo->groupIndex) {
+      } else {
         qDebug("QInfo:%p query over, %" PRId64 " rows are returned", pQInfo, pQuery->rec.total);
       }
     }
@@ -6062,7 +5760,8 @@ static void tableQueryImpl(SQInfo *pQInfo) {
   pRuntimeEnv->summary.elapsedTime += (taosGetTimestampUs() - st);
   assert(pQInfo->tableqinfoGroupInfo.numOfTables == 1);
 }
-static void buildTableBlockDistResult(SQInfo *pQInfo) {
+
+void buildTableBlockDistResult(SQInfo *pQInfo) {
   SQueryRuntimeEnv *pRuntimeEnv = &pQInfo->runtimeEnv;
   SQuery *pQuery = pRuntimeEnv->pQuery;
   pQuery->pos = 0;
@@ -6115,7 +5814,7 @@ static void buildTableBlockDistResult(SQInfo *pQInfo) {
   return;
 }
 
-static void stableQueryImpl(SQInfo *pQInfo) {
+void stableQueryImpl(SQInfo *pQInfo) {
   SQueryRuntimeEnv* pRuntimeEnv = &pQInfo->runtimeEnv;
   SQuery *pQuery = pRuntimeEnv->pQuery;
   pQuery->rec.rows = 0;
@@ -6127,7 +5826,6 @@ static void stableQueryImpl(SQInfo *pQInfo) {
     multiTableQueryProcess(pQInfo);
   } else {
     assert(pQuery->checkResultBuf == 1 || isPointInterpoQuery(pQuery) || pRuntimeEnv->groupbyColumn);
-
     sequentialTableProcess(pQInfo);
   }
 
@@ -6247,37 +5945,6 @@ static char *createTableIdList(SQueryTableMsg *pQueryMsg, char *pMsg, SArray **p
   return pMsg;
 }
 
-typedef struct SQueryParam {
-  char            *sql;
-  char            *tagCond;
-  char            *tbnameCond;
-  char            *prevResult;
-  SArray          *pTableIdList;
-  SSqlFuncMsg    **pExprMsg;
-  SSqlFuncMsg    **pSecExprMsg;
-  SExprInfo       *pExprs;
-  SExprInfo       *pSecExprs;
-
-  SColIndex       *pGroupColIndex;
-  SColumnInfo     *pTagColumnInfo;
-  SSqlGroupbyExpr *pGroupbyExpr;
-} SQueryParam;
-
-static void freeParam(SQueryParam *param) {
-  tfree(param->sql);
-  tfree(param->tagCond);
-  tfree(param->tbnameCond);
-  tfree(param->pTableIdList);
-  tfree(param->pExprMsg);
-  tfree(param->pSecExprMsg);
-  tfree(param->pExprs);
-  tfree(param->pSecExprs);
-  tfree(param->pGroupColIndex);
-  tfree(param->pTagColumnInfo);
-  tfree(param->pGroupbyExpr);
-  tfree(param->prevResult);
-}
-
 /**
  * pQueryMsg->head has been converted before this function is called.
  *
@@ -6286,7 +5953,7 @@ static void freeParam(SQueryParam *param) {
  * @param pExpr
  * @return
  */
-static int32_t convertQueryMsg(SQueryTableMsg *pQueryMsg, SQueryParam* param) {
+int32_t convertQueryMsg(SQueryTableMsg *pQueryMsg, SQueryParam* param) {
   int32_t code = TSDB_CODE_SUCCESS;
 
   if (taosCheckVersion(pQueryMsg->version, version, 3) != 0) {
@@ -6604,7 +6271,7 @@ static int32_t buildArithmeticExprFromMsg(SExprInfo *pArithExprInfo, SQueryTable
   return TSDB_CODE_SUCCESS;
 }
 
-static int32_t createQueryFuncExprFromMsg(SQueryTableMsg *pQueryMsg, int32_t numOfOutput, SExprInfo **pExprInfo, SSqlFuncMsg **pExprMsg,
+int32_t createQueryFuncExprFromMsg(SQueryTableMsg *pQueryMsg, int32_t numOfOutput, SExprInfo **pExprInfo, SSqlFuncMsg **pExprMsg,
     SColumnInfo* pTagCols) {
   *pExprInfo = NULL;
   int32_t code = TSDB_CODE_SUCCESS;
@@ -6714,7 +6381,7 @@ static int32_t createQueryFuncExprFromMsg(SQueryTableMsg *pQueryMsg, int32_t num
   return TSDB_CODE_SUCCESS;
 }
 
-static SSqlGroupbyExpr *createGroupbyExprFromMsg(SQueryTableMsg *pQueryMsg, SColIndex *pColIndex, int32_t *code) {
+SSqlGroupbyExpr *createGroupbyExprFromMsg(SQueryTableMsg *pQueryMsg, SColIndex *pColIndex, int32_t *code) {
   if (pQueryMsg->numOfGroupCols == 0) {
     return NULL;
   }
@@ -6831,8 +6498,6 @@ static void doUpdateExprColumnIndex(SQuery *pQuery) {
   }
 }
 
-static void freeQInfo(SQInfo *pQInfo);
-
 static void calResultBufSize(SQuery* pQuery) {
   const int32_t RESULT_MSG_MIN_SIZE  = 1024 * (1024 + 512);  // bytes
   const int32_t RESULT_MSG_MIN_ROWS  = 8192;
@@ -6852,7 +6517,7 @@ static void calResultBufSize(SQuery* pQuery) {
   }
 }
 
-static SQInfo *createQInfoImpl(SQueryTableMsg *pQueryMsg, SSqlGroupbyExpr *pGroupbyExpr, SExprInfo *pExprs,
+SQInfo *createQInfoImpl(SQueryTableMsg *pQueryMsg, SSqlGroupbyExpr *pGroupbyExpr, SExprInfo *pExprs,
                                SExprInfo *pSecExprs, STableGroupInfo *pTableGroupInfo, SColumnInfo* pTagCols, bool stableQuery, char* sql) {
   int16_t numOfCols = pQueryMsg->numOfCols;
   int16_t numOfOutput = pQueryMsg->numOfOutput;
@@ -7060,7 +6725,7 @@ _cleanup:
   return NULL;
 }
 
-static bool isValidQInfo(void *param) {
+bool isValidQInfo(void *param) {
   SQInfo *pQInfo = (SQInfo *)param;
   if (pQInfo == NULL) {
     return false;
@@ -7074,7 +6739,7 @@ static bool isValidQInfo(void *param) {
   return (sig == (uint64_t)pQInfo);
 }
 
-static int32_t initQInfo(SQueryTableMsg *pQueryMsg, void *tsdb, int32_t vgId, SQInfo *pQInfo, SQueryParam* param, bool isSTable) {
+int32_t initQInfo(SQueryTableMsg *pQueryMsg, void *tsdb, int32_t vgId, SQInfo *pQInfo, SQueryParam* param, bool isSTable) {
   int32_t code = TSDB_CODE_SUCCESS;
   SQuery *pQuery = pQInfo->runtimeEnv.pQuery;
 
@@ -7125,7 +6790,7 @@ _error:
   return code;
 }
 
-static void freeColumnFilterInfo(SColumnFilterInfo* pFilter, int32_t numOfFilters) {
+void freeColumnFilterInfo(SColumnFilterInfo* pFilter, int32_t numOfFilters) {
     if (pFilter == NULL || numOfFilters == 0) {
       return;
     }
@@ -7179,7 +6844,7 @@ static void* destroyQueryFuncExpr(SExprInfo* pExprInfo, int32_t numOfExpr) {
   return NULL;
 }
 
-static void freeQInfo(SQInfo *pQInfo) {
+void freeQInfo(SQInfo *pQInfo) {
   if (!isValidQInfo(pQInfo)) {
     return;
   }
@@ -7248,7 +6913,7 @@ static void freeQInfo(SQInfo *pQInfo) {
   tfree(pQInfo);
 }
 
-static size_t getResultSize(SQInfo *pQInfo, int64_t *numOfRows) {
+size_t getResultSize(SQInfo *pQInfo, int64_t *numOfRows) {
   SQuery *pQuery = pQInfo->runtimeEnv.pQuery;
 
   /*
@@ -7271,7 +6936,7 @@ static size_t getResultSize(SQInfo *pQInfo, int64_t *numOfRows) {
   }
 }
 
-static int32_t doDumpQueryResult(SQInfo *pQInfo, char *data) {
+int32_t doDumpQueryResult(SQInfo *pQInfo, char *data) {
   // the remained number of retrieved rows, not the interpolated result
   SQuery *pQuery = pQInfo->runtimeEnv.pQuery;
 
@@ -7326,154 +6991,7 @@ static int32_t doDumpQueryResult(SQInfo *pQInfo, char *data) {
   return TSDB_CODE_SUCCESS;
 }
 
-typedef struct SQueryMgmt {
-  pthread_mutex_t lock;
-  SCacheObj      *qinfoPool;      // query handle pool
-  int32_t         vgId;
-  bool            closed;
-} SQueryMgmt;
-
-int32_t qCreateQueryInfo(void* tsdb, int32_t vgId, SQueryTableMsg* pQueryMsg, qinfo_t* pQInfo) {
-  assert(pQueryMsg != NULL && tsdb != NULL);
-
-  int32_t code = TSDB_CODE_SUCCESS;
-
-  SQueryParam param = {0};
-  code = convertQueryMsg(pQueryMsg, &param);
-  if (code != TSDB_CODE_SUCCESS) {
-    goto _over;
-  }
-
-  if (pQueryMsg->numOfTables <= 0) {
-    qError("Invalid number of tables to query, numOfTables:%d", pQueryMsg->numOfTables);
-    code = TSDB_CODE_QRY_INVALID_MSG;
-    goto _over;
-  }
-
-  if (param.pTableIdList == NULL || taosArrayGetSize(param.pTableIdList) == 0) {
-    qError("qmsg:%p, SQueryTableMsg wrong format", pQueryMsg);
-    code = TSDB_CODE_QRY_INVALID_MSG;
-    goto _over;
-  }
-
-  if ((code = createQueryFuncExprFromMsg(pQueryMsg, pQueryMsg->numOfOutput, &param.pExprs, param.pExprMsg, param.pTagColumnInfo)) != TSDB_CODE_SUCCESS) {
-    goto _over;
-  }
-
-  if (param.pSecExprMsg != NULL) {
-    if ((code = createQueryFuncExprFromMsg(pQueryMsg, pQueryMsg->secondStageOutput, &param.pSecExprs, param.pSecExprMsg, param.pTagColumnInfo)) != TSDB_CODE_SUCCESS) {
-      goto _over;
-    }
-  }
-
-  param.pGroupbyExpr = createGroupbyExprFromMsg(pQueryMsg, param.pGroupColIndex, &code);
-  if ((param.pGroupbyExpr == NULL && pQueryMsg->numOfGroupCols != 0) || code != TSDB_CODE_SUCCESS) {
-    goto _over;
-  }
-
-  bool isSTableQuery = false;
-  STableGroupInfo tableGroupInfo = {0};
-  int64_t st = taosGetTimestampUs();
-
-  if (TSDB_QUERY_HAS_TYPE(pQueryMsg->queryType, TSDB_QUERY_TYPE_TABLE_QUERY)) {
-    STableIdInfo *id = taosArrayGet(param.pTableIdList, 0);
-
-    qDebug("qmsg:%p query normal table, uid:%"PRId64", tid:%d", pQueryMsg, id->uid, id->tid);
-    if ((code = tsdbGetOneTableGroup(tsdb, id->uid, pQueryMsg->window.skey, &tableGroupInfo)) != TSDB_CODE_SUCCESS) {
-      goto _over;
-    }
-  } else if (TSDB_QUERY_HAS_TYPE(pQueryMsg->queryType, TSDB_QUERY_TYPE_MULTITABLE_QUERY|TSDB_QUERY_TYPE_STABLE_QUERY)) {
-    isSTableQuery = true;
-
-    // also note there's possibility that only one table in the super table
-    if (!TSDB_QUERY_HAS_TYPE(pQueryMsg->queryType, TSDB_QUERY_TYPE_MULTITABLE_QUERY)) {
-      STableIdInfo *id = taosArrayGet(param.pTableIdList, 0);
-
-      // group by normal column, do not pass the group by condition to tsdb to group table into different group
-      int32_t numOfGroupByCols = pQueryMsg->numOfGroupCols;
-      if (pQueryMsg->numOfGroupCols == 1 && !TSDB_COL_IS_TAG(param.pGroupColIndex->flag)) {
-        numOfGroupByCols = 0;
-      }
-
-      qDebug("qmsg:%p query stable, uid:%"PRId64", tid:%d", pQueryMsg, id->uid, id->tid);
-      code = tsdbQuerySTableByTagCond(tsdb, id->uid, pQueryMsg->window.skey, param.tagCond, pQueryMsg->tagCondLen,
-          pQueryMsg->tagNameRelType, param.tbnameCond, &tableGroupInfo, param.pGroupColIndex, numOfGroupByCols);
-
-      if (code != TSDB_CODE_SUCCESS) {
-        qError("qmsg:%p failed to query stable, reason: %s", pQueryMsg, tstrerror(code));
-        goto _over;
-      }
-    } else {
-      code = tsdbGetTableGroupFromIdList(tsdb, param.pTableIdList, &tableGroupInfo);
-      if (code != TSDB_CODE_SUCCESS) {
-        goto _over;
-      }
-
-      qDebug("qmsg:%p query on %" PRIzu " tables in one group from client", pQueryMsg, tableGroupInfo.numOfTables);
-    }
-
-    int64_t el = taosGetTimestampUs() - st;
-    qDebug("qmsg:%p tag filter completed, numOfTables:%" PRIzu ", elapsed time:%"PRId64"us", pQueryMsg, tableGroupInfo.numOfTables, el);
-  } else {
-    assert(0);
-  }
-
-  code = checkForQueryBuf(tableGroupInfo.numOfTables);
-  if (code != TSDB_CODE_SUCCESS) {  // not enough query buffer, abort
-    goto _over;
-  }
-
-  (*pQInfo) = createQInfoImpl(pQueryMsg, param.pGroupbyExpr, param.pExprs, param.pSecExprs, &tableGroupInfo, param.pTagColumnInfo, isSTableQuery, param.sql);
-
-  param.sql    = NULL;
-  param.pExprs = NULL;
-  param.pSecExprs = NULL;
-  param.pGroupbyExpr = NULL;
-  param.pTagColumnInfo = NULL;
-
-  if ((*pQInfo) == NULL) {
-    code = TSDB_CODE_QRY_OUT_OF_MEMORY;
-    goto _over;
-  }
-
-  code = initQInfo(pQueryMsg, tsdb, vgId, *pQInfo, &param, isSTableQuery);
-
-_over:
-  if (param.pGroupbyExpr != NULL) {
-    taosArrayDestroy(param.pGroupbyExpr->columnInfo);
-  }
-
-  taosArrayDestroy(param.pTableIdList);
-  param.pTableIdList = NULL;
-
-  freeParam(&param);
-
-  for (int32_t i = 0; i < pQueryMsg->numOfCols; i++) {
-    SColumnInfo* column = pQueryMsg->colList + i;
-    freeColumnFilterInfo(column->filters, column->numOfFilters);
-  }
-
-  //pQInfo already freed in initQInfo, but *pQInfo may not pointer to null;
-  if (code != TSDB_CODE_SUCCESS) {
-    *pQInfo = NULL;
-  }
-
-  // if failed to add ref for all tables in this query, abort current query
-  return code;
-}
-
-void qDestroyQueryInfo(qinfo_t qHandle) {
-  SQInfo* pQInfo = (SQInfo*) qHandle;
-  if (!isValidQInfo(pQInfo)) {
-    return;
-  }
-
-  qDebug("QInfo:%p query completed", pQInfo);
-  queryCostStatis(pQInfo);   // print the query cost summary
-  freeQInfo(pQInfo);
-}
-
-static bool doBuildResCheck(SQInfo* pQInfo) {
+bool doBuildResCheck(SQInfo* pQInfo) {
   bool buildRes = false;
 
   pthread_mutex_lock(&pQInfo->lock);
@@ -7493,212 +7011,20 @@ static bool doBuildResCheck(SQInfo* pQInfo) {
   return buildRes;
 }
 
-bool qTableQuery(qinfo_t qinfo) {
-  SQInfo *pQInfo = (SQInfo *)qinfo;
-  assert(pQInfo && pQInfo->signature == pQInfo);
-  int64_t threadId = taosGetSelfPthreadId();
-
-  int64_t curOwner = 0;
-  if ((curOwner = atomic_val_compare_exchange_64(&pQInfo->owner, 0, threadId)) != 0) {
-    qError("QInfo:%p qhandle is now executed by thread:%p", pQInfo, (void*) curOwner);
-    pQInfo->code = TSDB_CODE_QRY_IN_EXEC;
-    return false;
-  }
-
-  pQInfo->startExecTs = taosGetTimestampSec();
-
-  if (isQueryKilled(pQInfo)) {
-    qDebug("QInfo:%p it is already killed, abort", pQInfo);
-    return doBuildResCheck(pQInfo);
-  }
-
-  if (pQInfo->tableqinfoGroupInfo.numOfTables == 0) {
-    qDebug("QInfo:%p no table exists for query, abort", pQInfo);
-    setQueryStatus(pQInfo->runtimeEnv.pQuery, QUERY_COMPLETED);
-    return doBuildResCheck(pQInfo);
-  }
-
-  // error occurs, record the error code and return to client
-  int32_t ret = setjmp(pQInfo->runtimeEnv.env);
-  if (ret != TSDB_CODE_SUCCESS) {
-    pQInfo->code = ret;
-    qDebug("QInfo:%p query abort due to error/cancel occurs, code:%s", pQInfo, tstrerror(pQInfo->code));
-    return doBuildResCheck(pQInfo);
-  }
-
-  qDebug("QInfo:%p query task is launched", pQInfo);
-
-  SQueryRuntimeEnv* pRuntimeEnv = &pQInfo->runtimeEnv;
-  if (onlyQueryTags(pQInfo->runtimeEnv.pQuery)) {
-    assert(pQInfo->runtimeEnv.pQueryHandle == NULL);
-    buildTagQueryResult(pQInfo);
-  } else if (pQInfo->runtimeEnv.stableQuery) {
-    stableQueryImpl(pQInfo);
-  } else if (pQInfo->runtimeEnv.queryBlockDist){
-    buildTableBlockDistResult(pQInfo);  
-  } else {
-    tableQueryImpl(pQInfo);
-  }
-
-  SQuery* pQuery = pRuntimeEnv->pQuery;
-  if (isQueryKilled(pQInfo)) {
-    qDebug("QInfo:%p query is killed", pQInfo);
-  } else if (pQuery->rec.rows == 0) {
-    qDebug("QInfo:%p over, %" PRIzu " tables queried, %"PRId64" rows are returned", pQInfo, pQInfo->tableqinfoGroupInfo.numOfTables, pQuery->rec.total);
-  } else {
-    qDebug("QInfo:%p query paused, %" PRId64 " rows returned, numOfTotal:%" PRId64 " rows",
-           pQInfo, pQuery->rec.rows, pQuery->rec.total + pQuery->rec.rows);
-  }
-
-  return doBuildResCheck(pQInfo);
-}
-
-int32_t qRetrieveQueryResultInfo(qinfo_t qinfo, bool* buildRes, void* pRspContext) {
-  SQInfo *pQInfo = (SQInfo *)qinfo;
-
-  if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
-    qError("QInfo:%p invalid qhandle", pQInfo);
-    return TSDB_CODE_QRY_INVALID_QHANDLE;
-  }
-
-  *buildRes = false;
-  if (IS_QUERY_KILLED(pQInfo)) {
-    qDebug("QInfo:%p query is killed, code:0x%08x", pQInfo, pQInfo->code);
-    return pQInfo->code;
-  }
-
-  int32_t code = TSDB_CODE_SUCCESS;
-
-  if (tsRetrieveBlockingModel) {
-    pQInfo->rspContext = pRspContext;
-    tsem_wait(&pQInfo->ready);
-    *buildRes = true;
-    code = pQInfo->code;
-  } else {
-    SQuery *pQuery = pQInfo->runtimeEnv.pQuery;
-
-    pthread_mutex_lock(&pQInfo->lock);
-
-    assert(pQInfo->rspContext == NULL);
-    if (pQInfo->dataReady == QUERY_RESULT_READY) {
-      *buildRes = true;
-      qDebug("QInfo:%p retrieve result info, rowsize:%d, rows:%" PRId64 ", code:%s", pQInfo, pQuery->resultRowSize,
-             pQuery->rec.rows, tstrerror(pQInfo->code));
-    } else {
-      *buildRes = false;
-      qDebug("QInfo:%p retrieve req set query return result after paused", pQInfo);
-      pQInfo->rspContext = pRspContext;
-      assert(pQInfo->rspContext != NULL);
-    }
-
-    code = pQInfo->code;
-    pthread_mutex_unlock(&pQInfo->lock);
-  }
-
-  return code;
-}
-
-int32_t qDumpRetrieveResult(qinfo_t qinfo, SRetrieveTableRsp **pRsp, int32_t *contLen, bool* continueExec) {
-  SQInfo *pQInfo = (SQInfo *)qinfo;
-
-  if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
-    return TSDB_CODE_QRY_INVALID_QHANDLE;
-  }
-
-  SQueryRuntimeEnv* pRuntimeEnv = &pQInfo->runtimeEnv;
-  SQuery *pQuery = pQInfo->runtimeEnv.pQuery;
-  size_t  size = getResultSize(pQInfo, &pQuery->rec.rows);
-
-  size += sizeof(int32_t);
-  size += sizeof(STableIdInfo) * taosHashGetSize(pQInfo->arrTableIdInfo);
-
-  *contLen = (int32_t)(size + sizeof(SRetrieveTableRsp));
-
-  // todo proper handle failed to allocate memory,
-  // current solution only avoid crash, but cannot return error code to client
-  *pRsp = (SRetrieveTableRsp *)rpcMallocCont(*contLen);
-  if (*pRsp == NULL) {
-    return TSDB_CODE_QRY_OUT_OF_MEMORY;
-  }
-
-  (*pRsp)->numOfRows = htonl((int32_t)pQuery->rec.rows);
-
-  if (pQInfo->code == TSDB_CODE_SUCCESS) {
-    (*pRsp)->offset   = htobe64(pQuery->limit.offset);
-    (*pRsp)->useconds = htobe64(pRuntimeEnv->summary.elapsedTime);
-  } else {
-    (*pRsp)->offset   = 0;
-    (*pRsp)->useconds = htobe64(pRuntimeEnv->summary.elapsedTime);
-  }
-
-  (*pRsp)->precision = htons(pQuery->precision);
-  if (pQuery->rec.rows > 0 && pQInfo->code == TSDB_CODE_SUCCESS) {
-    doDumpQueryResult(pQInfo, (*pRsp)->data);
-  } else {
-    setQueryStatus(pQuery, QUERY_OVER);
-  }
-
-  pQInfo->rspContext = NULL;
-  pQInfo->dataReady  = QUERY_RESULT_NOT_READY;
-
-  if (IS_QUERY_KILLED(pQInfo) || Q_STATUS_EQUAL(pQuery->status, QUERY_OVER)) {
-    // here current thread hold the refcount, so it is safe to free tsdbQueryHandle.
-    *continueExec = false;
-    (*pRsp)->completed = 1;  // notify no more result to client
-  } else {
-    *continueExec = true;
-    qDebug("QInfo:%p has more results to retrieve", pQInfo);
-  }
-
-  return pQInfo->code;
-}
-
-int32_t qQueryCompleted(qinfo_t qinfo) {
-  SQInfo *pQInfo = (SQInfo *)qinfo;
-
-  if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
-    return TSDB_CODE_QRY_INVALID_QHANDLE;
-  }
-
-  SQuery* pQuery = pQInfo->runtimeEnv.pQuery;
-  return isQueryKilled(pQInfo) || Q_STATUS_EQUAL(pQuery->status, QUERY_OVER);
-}
-
-int32_t qKillQuery(qinfo_t qinfo) {
-  SQInfo *pQInfo = (SQInfo *)qinfo;
-
-  if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
-    return TSDB_CODE_QRY_INVALID_QHANDLE;
-  }
-
-  setQueryKilled(pQInfo);
-
-  // Wait for the query executing thread being stopped/
-  // Once the query is stopped, the owner of qHandle will be cleared immediately.
-  while (pQInfo->owner != 0) {
-    taosMsleep(100);
+static void doSetTagValueToResultBuf(char* output, const char* val, int16_t type, int16_t bytes) {
+  if (val == NULL) {
+    setNull(output, type, bytes);
+    return;
   }
 
-  return TSDB_CODE_SUCCESS;
-}
-
-static void doSetTagValueToResultBuf(char* output, const char* val, int16_t type, int16_t bytes) {
   if (type == TSDB_DATA_TYPE_BINARY || type == TSDB_DATA_TYPE_NCHAR) {
-    if (val == NULL) {
-      setVardataNull(output, type);
-    } else {
     memcpy(output, val, varDataTLen(val));
-    }
   } else {
-    if (val == NULL) {
-      setNull(output, type, bytes);
-    } else {  // todo here stop will cause client crash
     memcpy(output, val, bytes);
   }
-  }
 }
 
-static void buildTagQueryResult(SQInfo* pQInfo) {
+void buildTagQueryResult(SQInfo* pQInfo) {
   SQueryRuntimeEnv *pRuntimeEnv = &pQInfo->runtimeEnv;
   SQuery *          pQuery = pRuntimeEnv->pQuery;
 
@@ -7872,157 +7198,3 @@ void releaseQueryBuf(size_t numOfTables) {
   // restore value is not enough buffer available
   atomic_add_fetch_64(&tsQueryBufferSizeBytes, t);
 }
-
-void* qGetResultRetrieveMsg(qinfo_t qinfo) {
-  SQInfo* pQInfo = (SQInfo*) qinfo;
-  assert(pQInfo != NULL);
-
-  return pQInfo->rspContext;
-}
-
-void freeqinfoFn(void *qhandle) {
-  void** handle = qhandle;
-  if (handle == NULL || *handle == NULL) {
-    return;
-  }
-
-  qKillQuery(*handle);
-  qDestroyQueryInfo(*handle);
-}
-
-void* qOpenQueryMgmt(int32_t vgId) {
-  const int32_t REFRESH_HANDLE_INTERVAL = 30; // every 30 seconds, refresh handle pool
-
-  char cacheName[128] = {0};
-  sprintf(cacheName, "qhandle_%d", vgId);
-
-  SQueryMgmt* pQueryMgmt = calloc(1, sizeof(SQueryMgmt));
-  if (pQueryMgmt == NULL) {
-    terrno = TSDB_CODE_QRY_OUT_OF_MEMORY;
-    return NULL;
-  }
-
-  pQueryMgmt->qinfoPool = taosCacheInit(TSDB_CACHE_PTR_KEY, REFRESH_HANDLE_INTERVAL, true, freeqinfoFn, cacheName);
-  pQueryMgmt->closed    = false;
-  pQueryMgmt->vgId      = vgId;
-
-  pthread_mutex_init(&pQueryMgmt->lock, NULL);
-
-  qDebug("vgId:%d, open querymgmt success", vgId);
-  return pQueryMgmt;
-}
-
-static void queryMgmtKillQueryFn(void* handle) {
-  void** fp = (void**)handle;
-  qKillQuery(*fp);
-}
-
-void qQueryMgmtNotifyClosed(void* pQMgmt) {
-  if (pQMgmt == NULL) {
-    return;
-  }
-
-  SQueryMgmt* pQueryMgmt = pQMgmt;
-  qDebug("vgId:%d, set querymgmt closed, wait for all queries cancelled", pQueryMgmt->vgId);
-
-  pthread_mutex_lock(&pQueryMgmt->lock);
-  pQueryMgmt->closed = true;
-  pthread_mutex_unlock(&pQueryMgmt->lock);
-
-  taosCacheRefresh(pQueryMgmt->qinfoPool, queryMgmtKillQueryFn);
-}
-
-void qQueryMgmtReOpen(void *pQMgmt) {
-  if (pQMgmt == NULL) {
-    return;
-  }
-
-  SQueryMgmt *pQueryMgmt = pQMgmt;
-  qDebug("vgId:%d, set querymgmt reopen", pQueryMgmt->vgId);
-
-  pthread_mutex_lock(&pQueryMgmt->lock);
-  pQueryMgmt->closed = false;
-  pthread_mutex_unlock(&pQueryMgmt->lock);
-}
-
-void qCleanupQueryMgmt(void* pQMgmt) {
-  if (pQMgmt == NULL) {
-    return;
-  }
-
-  SQueryMgmt* pQueryMgmt = pQMgmt;
-  int32_t vgId = pQueryMgmt->vgId;
-
-  assert(pQueryMgmt->closed);
-
-  SCacheObj* pqinfoPool = pQueryMgmt->qinfoPool;
-  pQueryMgmt->qinfoPool = NULL;
-
-  taosCacheCleanup(pqinfoPool);
-  pthread_mutex_destroy(&pQueryMgmt->lock);
-  tfree(pQueryMgmt);
-
-  qDebug("vgId:%d, queryMgmt cleanup completed", vgId);
-}
-
-void** qRegisterQInfo(void* pMgmt, uint64_t qInfo) {
-  if (pMgmt == NULL) {
-    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
-    return NULL;
-  }
-
-  SQueryMgmt *pQueryMgmt = pMgmt;
-  if (pQueryMgmt->qinfoPool == NULL) {
-    qError("QInfo:%p failed to add qhandle into qMgmt, since qMgmt is closed", (void *)qInfo);
-    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
-    return NULL;
-  }
-
-  pthread_mutex_lock(&pQueryMgmt->lock);
-  if (pQueryMgmt->closed) {
-    pthread_mutex_unlock(&pQueryMgmt->lock);
-    qError("QInfo:%p failed to add qhandle into cache, since qMgmt is colsing", (void *)qInfo);
-    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
-    return NULL;
-  } else {
-    TSDB_CACHE_PTR_TYPE handleVal = (TSDB_CACHE_PTR_TYPE) qInfo;
-    void** handle = taosCachePut(pQueryMgmt->qinfoPool, &handleVal, sizeof(TSDB_CACHE_PTR_TYPE), &qInfo, sizeof(TSDB_CACHE_PTR_TYPE),
-        (getMaximumIdleDurationSec()*1000));
-    pthread_mutex_unlock(&pQueryMgmt->lock);
-
-    return handle;
-  }
-}
-
-void** qAcquireQInfo(void* pMgmt, uint64_t _key) {
-  SQueryMgmt *pQueryMgmt = pMgmt;
-
-  if (pQueryMgmt->closed) {
-    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
-    return NULL;
-  }
-
-  if (pQueryMgmt->qinfoPool == NULL) {
-    terrno = TSDB_CODE_QRY_INVALID_QHANDLE;
-    return NULL;
-  }
-
-  TSDB_CACHE_PTR_TYPE key = (TSDB_CACHE_PTR_TYPE)_key;
-  void** handle = taosCacheAcquireByKey(pQueryMgmt->qinfoPool, &key, sizeof(TSDB_CACHE_PTR_TYPE));
-  if (handle == NULL || *handle == NULL) {
-    terrno = TSDB_CODE_QRY_INVALID_QHANDLE;
-    return NULL;
-  } else {
-    return handle;
-  }
-}
-
-void** qReleaseQInfo(void* pMgmt, void* pQInfo, bool freeHandle) {
-  SQueryMgmt *pQueryMgmt = pMgmt;
-  if (pQueryMgmt->qinfoPool == NULL) {
-    return NULL;
-  }
-
-  taosCacheRelease(pQueryMgmt->qinfoPool, pQInfo, freeHandle);
-  return 0;
-}

src/query/src/qUtil.c

@@ -20,6 +20,14 @@
 #include "qExecutor.h"
 #include "qUtil.h"
 #include "tbuffer.h"
+#include "tlosertree.h"
+#include "queryLog.h"
+
+typedef struct SCompSupporter {
+  STableQueryInfo **pTableQueryInfo;
+  int32_t          *rowIndex;
+  int32_t           order;
+} SCompSupporter;
 
 int32_t getOutputInterResultBufSize(SQuery* pQuery) {
   int32_t size = 0;
@@ -323,3 +331,242 @@ void freeInterResult(void* param) {
 
   taosArrayDestroy(pResult->pResult);
 }
+
+void cleanupGroupResInfo(SGroupResInfo* pGroupResInfo) {
+  assert(pGroupResInfo != NULL);
+
+  taosArrayDestroy(pGroupResInfo->pRows);
+  pGroupResInfo->pRows     = NULL;
+  pGroupResInfo->index     = 0;
+}
+
+void initGroupResInfo(SGroupResInfo* pGroupResInfo, SResultRowInfo* pResultInfo, int32_t offset) {
+  if (pGroupResInfo->pRows != NULL) {
+    taosArrayDestroy(pGroupResInfo->pRows);
+  }
+
+  pGroupResInfo->pRows = taosArrayFromList(pResultInfo->pResult, pResultInfo->size, POINTER_BYTES);
+  pGroupResInfo->index = offset;
+
+  assert(pGroupResInfo->index <= getNumOfTotalRes(pGroupResInfo));
+}
+
+bool hasRemainData(SGroupResInfo* pGroupResInfo) {
+  if (pGroupResInfo->pRows == NULL) {
+    return false;
+  }
+
+  return pGroupResInfo->index < taosArrayGetSize(pGroupResInfo->pRows);
+}
+
+bool incNextGroup(SGroupResInfo* pGroupResInfo) {
+  return (++pGroupResInfo->currentGroup) < pGroupResInfo->totalGroup;
+}
+
+int32_t getNumOfTotalRes(SGroupResInfo* pGroupResInfo) {
+  assert(pGroupResInfo != NULL);
+  if (pGroupResInfo->pRows == 0) {
+    return 0;
+  }
+
+  return taosArrayGetSize(pGroupResInfo->pRows);
+}
+
+static int64_t getNumOfResultWindowRes(SQueryRuntimeEnv* pRuntimeEnv, SResultRow *pResultRow) {
+  SQuery* pQuery = pRuntimeEnv->pQuery;
+
+  for (int32_t j = 0; j < pQuery->numOfOutput; ++j) {
+    int32_t functionId = pQuery->pExpr1[j].base.functionId;
+
+    /*
+     * ts, tag, tagprj function can not decide the output number of current query
+     * the number of output result is decided by main output
+     */
+    if (functionId == TSDB_FUNC_TS || functionId == TSDB_FUNC_TAG || functionId == TSDB_FUNC_TAGPRJ) {
+      continue;
+    }
+
+    SResultRowCellInfo *pResultInfo = getResultCell(pRuntimeEnv, pResultRow, j);
+    assert(pResultInfo != NULL);
+
+    if (pResultInfo->numOfRes > 0) {
+      return pResultInfo->numOfRes;
+    }
+  }
+
+  return 0;
+}
+
+static int32_t tableResultComparFn(const void *pLeft, const void *pRight, void *param) {
+  int32_t left  = *(int32_t *)pLeft;
+  int32_t right = *(int32_t *)pRight;
+
+  SCompSupporter *  supporter = (SCompSupporter *)param;
+
+  int32_t leftPos  = supporter->rowIndex[left];
+  int32_t rightPos = supporter->rowIndex[right];
+
+  /* left source is exhausted */
+  if (leftPos == -1) {
+    return 1;
+  }
+
+  /* right source is exhausted*/
+  if (rightPos == -1) {
+    return -1;
+  }
+
+  STableQueryInfo** pList = supporter->pTableQueryInfo;
+
+  SResultRowInfo *pWindowResInfo1 = &(pList[left]->resInfo);
+  SResultRow * pWindowRes1 = getResultRow(pWindowResInfo1, leftPos);
+  TSKEY leftTimestamp = pWindowRes1->win.skey;
+
+  SResultRowInfo *pWindowResInfo2 = &(pList[right]->resInfo);
+  SResultRow * pWindowRes2 = getResultRow(pWindowResInfo2, rightPos);
+  TSKEY rightTimestamp = pWindowRes2->win.skey;
+
+  if (leftTimestamp == rightTimestamp) {
+    return 0;
+  }
+
+  if (supporter->order == TSDB_ORDER_ASC) {
+    return (leftTimestamp > rightTimestamp)? 1:-1;
+  } else {
+    return (leftTimestamp < rightTimestamp)? 1:-1;
+  }
+}
+
+static int32_t mergeIntoGroupResultImpl(SQueryRuntimeEnv *pRuntimeEnv, SGroupResInfo* pGroupResInfo, SArray *pTableList, void* qinfo) {
+  bool ascQuery = QUERY_IS_ASC_QUERY(pRuntimeEnv->pQuery);
+
+  int32_t code = TSDB_CODE_SUCCESS;
+
+  int32_t *posList = NULL;
+  SLoserTreeInfo *pTree = NULL;
+  STableQueryInfo **pTableQueryInfoList = NULL;
+
+  size_t size = taosArrayGetSize(pTableList);
+  if (pGroupResInfo->pRows == NULL) {
+    pGroupResInfo->pRows = taosArrayInit(100, POINTER_BYTES);
+  }
+
+  posList = calloc(size, sizeof(int32_t));
+  pTableQueryInfoList = malloc(POINTER_BYTES * size);
+
+  if (pTableQueryInfoList == NULL || posList == NULL || pGroupResInfo->pRows == NULL || pGroupResInfo->pRows == NULL) {
+    qError("QInfo:%p failed alloc memory", qinfo);
+    code = TSDB_CODE_QRY_OUT_OF_MEMORY;
+    goto _end;
+  }
+
+  int32_t numOfTables = 0;
+  for (int32_t i = 0; i < size; ++i) {
+    STableQueryInfo *item = taosArrayGetP(pTableList, i);
+    if (item->resInfo.size > 0) {
+      pTableQueryInfoList[numOfTables++] = item;
+    }
+  }
+
+  // there is no data in current group
+  // no need to merge results since only one table in each group
+  if (numOfTables == 0) {
+    goto _end;
+  }
+
+  SCompSupporter cs = {pTableQueryInfoList, posList, pRuntimeEnv->pQuery->order.order};
+
+  int32_t ret = tLoserTreeCreate(&pTree, numOfTables, &cs, tableResultComparFn);
+  if (ret != TSDB_CODE_SUCCESS) {
+    code = TSDB_CODE_QRY_OUT_OF_MEMORY;
+    goto _end;
+  }
+
+  int64_t lastTimestamp = ascQuery? INT64_MIN:INT64_MAX;
+  int64_t startt = taosGetTimestampMs();
+
+  while (1) {
+    int32_t tableIndex = pTree->pNode[0].index;
+
+    SResultRowInfo *pWindowResInfo = &pTableQueryInfoList[tableIndex]->resInfo;
+    SResultRow  *pWindowRes = getResultRow(pWindowResInfo, cs.rowIndex[tableIndex]);
+
+    int64_t num = getNumOfResultWindowRes(pRuntimeEnv, pWindowRes);
+    if (num <= 0) {
+      cs.rowIndex[tableIndex] += 1;
+
+      if (cs.rowIndex[tableIndex] >= pWindowResInfo->size) {
+        cs.rowIndex[tableIndex] = -1;
+        if (--numOfTables == 0) { // all input sources are exhausted
+          break;
+        }
+      }
+    } else {
+      assert((pWindowRes->win.skey >= lastTimestamp && ascQuery) || (pWindowRes->win.skey <= lastTimestamp && !ascQuery));
+
+      if (pWindowRes->win.skey != lastTimestamp) {
+        taosArrayPush(pGroupResInfo->pRows, &pWindowRes);
+        pWindowRes->numOfRows = (uint32_t) num;
+      }
+
+      lastTimestamp = pWindowRes->win.skey;
+
+      // move to the next row of current entry
+      if ((++cs.rowIndex[tableIndex]) >= pWindowResInfo->size) {
+        cs.rowIndex[tableIndex] = -1;
+
+        // all input sources are exhausted
+        if ((--numOfTables) == 0) {
+          break;
+        }
+      }
+    }
+
+    tLoserTreeAdjust(pTree, tableIndex + pTree->numOfEntries);
+  }
+
+  int64_t endt = taosGetTimestampMs();
+
+  qDebug("QInfo:%p result merge completed for group:%d, elapsed time:%" PRId64 " ms", qinfo,
+         pGroupResInfo->currentGroup, endt - startt);
+
+  _end:
+  tfree(pTableQueryInfoList);
+  tfree(posList);
+  tfree(pTree);
+
+  return code;
+}
+
+int32_t mergeIntoGroupResult(SGroupResInfo* pGroupResInfo, SQInfo *pQInfo) {
+  int64_t st = taosGetTimestampUs();
+
+  while (pGroupResInfo->currentGroup < pGroupResInfo->totalGroup) {
+    SArray *group = GET_TABLEGROUP(pQInfo, pGroupResInfo->currentGroup);
+
+    int32_t ret = mergeIntoGroupResultImpl(&pQInfo->runtimeEnv, pGroupResInfo, group, pQInfo);
+    if (ret != TSDB_CODE_SUCCESS) {
+      return ret;
+    }
+
+    // this group generates at least one result, return results
+    if (taosArrayGetSize(pGroupResInfo->pRows) > 0) {
+      break;
+    }
+
+    qDebug("QInfo:%p no result in group %d, continue", pQInfo, pGroupResInfo->currentGroup);
+    cleanupGroupResInfo(pGroupResInfo);
+    incNextGroup(pGroupResInfo);
+  }
+
+  if (pGroupResInfo->currentGroup >= pGroupResInfo->totalGroup && !hasRemainData(pGroupResInfo)) {
+    SET_STABLE_QUERY_OVER(pQInfo);
+  }
+
+  int64_t elapsedTime = taosGetTimestampUs() - st;
+  qDebug("QInfo:%p merge res data into group, index:%d, total group:%d, elapsed time:%" PRId64 "us", pQInfo,
+         pGroupResInfo->currentGroup, pGroupResInfo->totalGroup, elapsedTime);
+
+  pQInfo->runtimeEnv.summary.firstStageMergeTime += elapsedTime;
+  return TSDB_CODE_SUCCESS;
+}

src/query/src/queryMain.c

+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "os.h"
+#include "qFill.h"
+#include "taosmsg.h"
+#include "tcache.h"
+#include "tglobal.h"
+
+#include "exception.h"
+#include "hash.h"
+#include "texpr.h"
+#include "qExecutor.h"
+#include "qResultbuf.h"
+#include "qUtil.h"
+#include "query.h"
+#include "queryLog.h"
+#include "tlosertree.h"
+#include "ttype.h"
+#include "tcompare.h"
+
+typedef struct SQueryMgmt {
+  pthread_mutex_t lock;
+  SCacheObj      *qinfoPool;      // query handle pool
+  int32_t         vgId;
+  bool            closed;
+} SQueryMgmt;
+
+static void queryMgmtKillQueryFn(void* handle) {
+  void** fp = (void**)handle;
+  qKillQuery(*fp);
+}
+
+static void freeqinfoFn(void *qhandle) {
+  void** handle = qhandle;
+  if (handle == NULL || *handle == NULL) {
+    return;
+  }
+
+  qKillQuery(*handle);
+  qDestroyQueryInfo(*handle);
+}
+
+void freeParam(SQueryParam *param) {
+  tfree(param->sql);
+  tfree(param->tagCond);
+  tfree(param->tbnameCond);
+  tfree(param->pTableIdList);
+  tfree(param->pExprMsg);
+  tfree(param->pSecExprMsg);
+  tfree(param->pExprs);
+  tfree(param->pSecExprs);
+  tfree(param->pGroupColIndex);
+  tfree(param->pTagColumnInfo);
+  tfree(param->pGroupbyExpr);
+  tfree(param->prevResult);
+}
+
+int32_t qCreateQueryInfo(void* tsdb, int32_t vgId, SQueryTableMsg* pQueryMsg, qinfo_t* pQInfo) {
+  assert(pQueryMsg != NULL && tsdb != NULL);
+
+  int32_t code = TSDB_CODE_SUCCESS;
+
+  SQueryParam param = {0};
+  code = convertQueryMsg(pQueryMsg, &param);
+  if (code != TSDB_CODE_SUCCESS) {
+    goto _over;
+  }
+
+  if (pQueryMsg->numOfTables <= 0) {
+    qError("Invalid number of tables to query, numOfTables:%d", pQueryMsg->numOfTables);
+    code = TSDB_CODE_QRY_INVALID_MSG;
+    goto _over;
+  }
+
+  if (param.pTableIdList == NULL || taosArrayGetSize(param.pTableIdList) == 0) {
+    qError("qmsg:%p, SQueryTableMsg wrong format", pQueryMsg);
+    code = TSDB_CODE_QRY_INVALID_MSG;
+    goto _over;
+  }
+
+  if ((code = createQueryFuncExprFromMsg(pQueryMsg, pQueryMsg->numOfOutput, &param.pExprs, param.pExprMsg, param.pTagColumnInfo)) != TSDB_CODE_SUCCESS) {
+    goto _over;
+  }
+
+  if (param.pSecExprMsg != NULL) {
+    if ((code = createQueryFuncExprFromMsg(pQueryMsg, pQueryMsg->secondStageOutput, &param.pSecExprs, param.pSecExprMsg, param.pTagColumnInfo)) != TSDB_CODE_SUCCESS) {
+      goto _over;
+    }
+  }
+
+  param.pGroupbyExpr = createGroupbyExprFromMsg(pQueryMsg, param.pGroupColIndex, &code);
+  if ((param.pGroupbyExpr == NULL && pQueryMsg->numOfGroupCols != 0) || code != TSDB_CODE_SUCCESS) {
+    goto _over;
+  }
+
+  bool isSTableQuery = false;
+  STableGroupInfo tableGroupInfo = {0};
+  int64_t st = taosGetTimestampUs();
+
+  if (TSDB_QUERY_HAS_TYPE(pQueryMsg->queryType, TSDB_QUERY_TYPE_TABLE_QUERY)) {
+    STableIdInfo *id = taosArrayGet(param.pTableIdList, 0);
+
+    qDebug("qmsg:%p query normal table, uid:%"PRId64", tid:%d", pQueryMsg, id->uid, id->tid);
+    if ((code = tsdbGetOneTableGroup(tsdb, id->uid, pQueryMsg->window.skey, &tableGroupInfo)) != TSDB_CODE_SUCCESS) {
+      goto _over;
+    }
+  } else if (TSDB_QUERY_HAS_TYPE(pQueryMsg->queryType, TSDB_QUERY_TYPE_MULTITABLE_QUERY|TSDB_QUERY_TYPE_STABLE_QUERY)) {
+    isSTableQuery = true;
+
+    // also note there's possibility that only one table in the super table
+    if (!TSDB_QUERY_HAS_TYPE(pQueryMsg->queryType, TSDB_QUERY_TYPE_MULTITABLE_QUERY)) {
+      STableIdInfo *id = taosArrayGet(param.pTableIdList, 0);
+
+      // group by normal column, do not pass the group by condition to tsdb to group table into different group
+      int32_t numOfGroupByCols = pQueryMsg->numOfGroupCols;
+      if (pQueryMsg->numOfGroupCols == 1 && !TSDB_COL_IS_TAG(param.pGroupColIndex->flag)) {
+        numOfGroupByCols = 0;
+      }
+
+      qDebug("qmsg:%p query stable, uid:%"PRId64", tid:%d", pQueryMsg, id->uid, id->tid);
+      code = tsdbQuerySTableByTagCond(tsdb, id->uid, pQueryMsg->window.skey, param.tagCond, pQueryMsg->tagCondLen,
+                                      pQueryMsg->tagNameRelType, param.tbnameCond, &tableGroupInfo, param.pGroupColIndex, numOfGroupByCols);
+
+      if (code != TSDB_CODE_SUCCESS) {
+        qError("qmsg:%p failed to query stable, reason: %s", pQueryMsg, tstrerror(code));
+        goto _over;
+      }
+    } else {
+      code = tsdbGetTableGroupFromIdList(tsdb, param.pTableIdList, &tableGroupInfo);
+      if (code != TSDB_CODE_SUCCESS) {
+        goto _over;
+      }
+
+      qDebug("qmsg:%p query on %" PRIzu " tables in one group from client", pQueryMsg, tableGroupInfo.numOfTables);
+    }
+
+    int64_t el = taosGetTimestampUs() - st;
+    qDebug("qmsg:%p tag filter completed, numOfTables:%" PRIzu ", elapsed time:%"PRId64"us", pQueryMsg, tableGroupInfo.numOfTables, el);
+  } else {
+    assert(0);
+  }
+
+  code = checkForQueryBuf(tableGroupInfo.numOfTables);
+  if (code != TSDB_CODE_SUCCESS) {  // not enough query buffer, abort
+    goto _over;
+  }
+
+  (*pQInfo) = createQInfoImpl(pQueryMsg, param.pGroupbyExpr, param.pExprs, param.pSecExprs, &tableGroupInfo, param.pTagColumnInfo, isSTableQuery, param.sql);
+
+  param.sql    = NULL;
+  param.pExprs = NULL;
+  param.pSecExprs = NULL;
+  param.pGroupbyExpr = NULL;
+  param.pTagColumnInfo = NULL;
+
+  if ((*pQInfo) == NULL) {
+    code = TSDB_CODE_QRY_OUT_OF_MEMORY;
+    goto _over;
+  }
+
+  code = initQInfo(pQueryMsg, tsdb, vgId, *pQInfo, &param, isSTableQuery);
+
+  _over:
+  if (param.pGroupbyExpr != NULL) {
+    taosArrayDestroy(param.pGroupbyExpr->columnInfo);
+  }
+
+  taosArrayDestroy(param.pTableIdList);
+  param.pTableIdList = NULL;
+
+  freeParam(&param);
+
+  for (int32_t i = 0; i < pQueryMsg->numOfCols; i++) {
+    SColumnInfo* column = pQueryMsg->colList + i;
+    freeColumnFilterInfo(column->filters, column->numOfFilters);
+  }
+
+  //pQInfo already freed in initQInfo, but *pQInfo may not pointer to null;
+  if (code != TSDB_CODE_SUCCESS) {
+    *pQInfo = NULL;
+  }
+
+  // if failed to add ref for all tables in this query, abort current query
+  return code;
+}
+
+bool qTableQuery(qinfo_t qinfo) {
+  SQInfo *pQInfo = (SQInfo *)qinfo;
+  assert(pQInfo && pQInfo->signature == pQInfo);
+  int64_t threadId = taosGetSelfPthreadId();
+
+  int64_t curOwner = 0;
+  if ((curOwner = atomic_val_compare_exchange_64(&pQInfo->owner, 0, threadId)) != 0) {
+    qError("QInfo:%p qhandle is now executed by thread:%p", pQInfo, (void*) curOwner);
+    pQInfo->code = TSDB_CODE_QRY_IN_EXEC;
+    return false;
+  }
+
+  pQInfo->startExecTs = taosGetTimestampSec();
+
+  if (isQueryKilled(pQInfo)) {
+    qDebug("QInfo:%p it is already killed, abort", pQInfo);
+    return doBuildResCheck(pQInfo);
+  }
+
+  if (pQInfo->tableqinfoGroupInfo.numOfTables == 0) {
+    qDebug("QInfo:%p no table exists for query, abort", pQInfo);
+    setQueryStatus(pQInfo->runtimeEnv.pQuery, QUERY_COMPLETED);
+    return doBuildResCheck(pQInfo);
+  }
+
+  // error occurs, record the error code and return to client
+  int32_t ret = setjmp(pQInfo->runtimeEnv.env);
+  if (ret != TSDB_CODE_SUCCESS) {
+    pQInfo->code = ret;
+    qDebug("QInfo:%p query abort due to error/cancel occurs, code:%s", pQInfo, tstrerror(pQInfo->code));
+    return doBuildResCheck(pQInfo);
+  }
+
+  qDebug("QInfo:%p query task is launched", pQInfo);
+
+  SQueryRuntimeEnv* pRuntimeEnv = &pQInfo->runtimeEnv;
+  if (onlyQueryTags(pQInfo->runtimeEnv.pQuery)) {
+    assert(pQInfo->runtimeEnv.pQueryHandle == NULL);
+    buildTagQueryResult(pQInfo);
+  } else if (pQInfo->runtimeEnv.stableQuery) {
+    stableQueryImpl(pQInfo);
+  } else if (pQInfo->runtimeEnv.queryBlockDist){
+    buildTableBlockDistResult(pQInfo);
+  } else {
+    tableQueryImpl(pQInfo);
+  }
+
+  SQuery* pQuery = pRuntimeEnv->pQuery;
+  if (isQueryKilled(pQInfo)) {
+    qDebug("QInfo:%p query is killed", pQInfo);
+  } else if (pQuery->rec.rows == 0) {
+    qDebug("QInfo:%p over, %" PRIzu " tables queried, %"PRId64" rows are returned", pQInfo, pQInfo->tableqinfoGroupInfo.numOfTables, pQuery->rec.total);
+  } else {
+    qDebug("QInfo:%p query paused, %" PRId64 " rows returned, numOfTotal:%" PRId64 " rows",
+           pQInfo, pQuery->rec.rows, pQuery->rec.total + pQuery->rec.rows);
+  }
+
+  return doBuildResCheck(pQInfo);
+}
+
+int32_t qRetrieveQueryResultInfo(qinfo_t qinfo, bool* buildRes, void* pRspContext) {
+  SQInfo *pQInfo = (SQInfo *)qinfo;
+
+  if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
+    qError("QInfo:%p invalid qhandle", pQInfo);
+    return TSDB_CODE_QRY_INVALID_QHANDLE;
+  }
+
+  *buildRes = false;
+  if (IS_QUERY_KILLED(pQInfo)) {
+    qDebug("QInfo:%p query is killed, code:0x%08x", pQInfo, pQInfo->code);
+    return pQInfo->code;
+  }
+
+  int32_t code = TSDB_CODE_SUCCESS;
+
+  if (tsRetrieveBlockingModel) {
+    pQInfo->rspContext = pRspContext;
+    tsem_wait(&pQInfo->ready);
+    *buildRes = true;
+    code = pQInfo->code;
+  } else {
+    SQuery *pQuery = pQInfo->runtimeEnv.pQuery;
+
+    pthread_mutex_lock(&pQInfo->lock);
+
+    assert(pQInfo->rspContext == NULL);
+    if (pQInfo->dataReady == QUERY_RESULT_READY) {
+      *buildRes = true;
+      qDebug("QInfo:%p retrieve result info, rowsize:%d, rows:%" PRId64 ", code:%s", pQInfo, pQuery->resultRowSize,
+             pQuery->rec.rows, tstrerror(pQInfo->code));
+    } else {
+      *buildRes = false;
+      qDebug("QInfo:%p retrieve req set query return result after paused", pQInfo);
+      pQInfo->rspContext = pRspContext;
+      assert(pQInfo->rspContext != NULL);
+    }
+
+    code = pQInfo->code;
+    pthread_mutex_unlock(&pQInfo->lock);
+  }
+
+  return code;
+}
+
+int32_t qDumpRetrieveResult(qinfo_t qinfo, SRetrieveTableRsp **pRsp, int32_t *contLen, bool* continueExec) {
+  SQInfo *pQInfo = (SQInfo *)qinfo;
+
+  if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
+    return TSDB_CODE_QRY_INVALID_QHANDLE;
+  }
+
+  SQueryRuntimeEnv* pRuntimeEnv = &pQInfo->runtimeEnv;
+  SQuery *pQuery = pQInfo->runtimeEnv.pQuery;
+  size_t  size = getResultSize(pQInfo, &pQuery->rec.rows);
+
+  size += sizeof(int32_t);
+  size += sizeof(STableIdInfo) * taosHashGetSize(pQInfo->arrTableIdInfo);
+
+  *contLen = (int32_t)(size + sizeof(SRetrieveTableRsp));
+
+  // current solution only avoid crash, but cannot return error code to client
+  *pRsp = (SRetrieveTableRsp *)rpcMallocCont(*contLen);
+  if (*pRsp == NULL) {
+    return TSDB_CODE_QRY_OUT_OF_MEMORY;
+  }
+
+  (*pRsp)->numOfRows = htonl((int32_t)pQuery->rec.rows);
+
+  if (pQInfo->code == TSDB_CODE_SUCCESS) {
+    (*pRsp)->offset   = htobe64(pQuery->limit.offset);
+    (*pRsp)->useconds = htobe64(pRuntimeEnv->summary.elapsedTime);
+  } else {
+    (*pRsp)->offset   = 0;
+    (*pRsp)->useconds = htobe64(pRuntimeEnv->summary.elapsedTime);
+  }
+
+  (*pRsp)->precision = htons(pQuery->precision);
+  if (pQuery->rec.rows > 0 && pQInfo->code == TSDB_CODE_SUCCESS) {
+    doDumpQueryResult(pQInfo, (*pRsp)->data);
+  } else {
+    setQueryStatus(pQuery, QUERY_OVER);
+  }
+
+  pQInfo->rspContext = NULL;
+  pQInfo->dataReady  = QUERY_RESULT_NOT_READY;
+
+  if (IS_QUERY_KILLED(pQInfo) || Q_STATUS_EQUAL(pQuery->status, QUERY_OVER)) {
+    // here current thread hold the refcount, so it is safe to free tsdbQueryHandle.
+    *continueExec = false;
+    (*pRsp)->completed = 1;  // notify no more result to client
+  } else {
+    *continueExec = true;
+    qDebug("QInfo:%p has more results to retrieve", pQInfo);
+  }
+
+  return pQInfo->code;
+}
+
+void* qGetResultRetrieveMsg(qinfo_t qinfo) {
+  SQInfo* pQInfo = (SQInfo*) qinfo;
+  assert(pQInfo != NULL);
+
+  return pQInfo->rspContext;
+}
+
+int32_t qKillQuery(qinfo_t qinfo) {
+  SQInfo *pQInfo = (SQInfo *)qinfo;
+
+  if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
+    return TSDB_CODE_QRY_INVALID_QHANDLE;
+  }
+
+  setQueryKilled(pQInfo);
+
+  // Wait for the query executing thread being stopped/
+  // Once the query is stopped, the owner of qHandle will be cleared immediately.
+  while (pQInfo->owner != 0) {
+    taosMsleep(100);
+  }
+
+  return TSDB_CODE_SUCCESS;
+}
+
+int32_t qQueryCompleted(qinfo_t qinfo) {
+  SQInfo *pQInfo = (SQInfo *)qinfo;
+
+  if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
+    return TSDB_CODE_QRY_INVALID_QHANDLE;
+  }
+
+  SQuery* pQuery = pQInfo->runtimeEnv.pQuery;
+  return isQueryKilled(pQInfo) || Q_STATUS_EQUAL(pQuery->status, QUERY_OVER);
+}
+
+void qDestroyQueryInfo(qinfo_t qHandle) {
+  SQInfo* pQInfo = (SQInfo*) qHandle;
+  if (!isValidQInfo(pQInfo)) {
+    return;
+  }
+
+  qDebug("QInfo:%p query completed", pQInfo);
+  queryCostStatis(pQInfo);   // print the query cost summary
+  freeQInfo(pQInfo);
+}
+
+void* qOpenQueryMgmt(int32_t vgId) {
+  const int32_t refreshHandleInterval = 30; // every 30 seconds, refresh handle pool
+
+  char cacheName[128] = {0};
+  sprintf(cacheName, "qhandle_%d", vgId);
+
+  SQueryMgmt* pQueryMgmt = calloc(1, sizeof(SQueryMgmt));
+  if (pQueryMgmt == NULL) {
+    terrno = TSDB_CODE_QRY_OUT_OF_MEMORY;
+    return NULL;
+  }
+
+  pQueryMgmt->qinfoPool = taosCacheInit(TSDB_CACHE_PTR_KEY, refreshHandleInterval, true, freeqinfoFn, cacheName);
+  pQueryMgmt->closed    = false;
+  pQueryMgmt->vgId      = vgId;
+
+  pthread_mutex_init(&pQueryMgmt->lock, NULL);
+
+  qDebug("vgId:%d, open querymgmt success", vgId);
+  return pQueryMgmt;
+}
+
+void qQueryMgmtNotifyClosed(void* pQMgmt) {
+  if (pQMgmt == NULL) {
+    return;
+  }
+
+  SQueryMgmt* pQueryMgmt = pQMgmt;
+  qDebug("vgId:%d, set querymgmt closed, wait for all queries cancelled", pQueryMgmt->vgId);
+
+  pthread_mutex_lock(&pQueryMgmt->lock);
+  pQueryMgmt->closed = true;
+  pthread_mutex_unlock(&pQueryMgmt->lock);
+
+  taosCacheRefresh(pQueryMgmt->qinfoPool, queryMgmtKillQueryFn);
+}
+
+void qQueryMgmtReOpen(void *pQMgmt) {
+  if (pQMgmt == NULL) {
+    return;
+  }
+
+  SQueryMgmt *pQueryMgmt = pQMgmt;
+  qDebug("vgId:%d, set querymgmt reopen", pQueryMgmt->vgId);
+
+  pthread_mutex_lock(&pQueryMgmt->lock);
+  pQueryMgmt->closed = false;
+  pthread_mutex_unlock(&pQueryMgmt->lock);
+}
+
+void qCleanupQueryMgmt(void* pQMgmt) {
+  if (pQMgmt == NULL) {
+    return;
+  }
+
+  SQueryMgmt* pQueryMgmt = pQMgmt;
+  int32_t vgId = pQueryMgmt->vgId;
+
+  assert(pQueryMgmt->closed);
+
+  SCacheObj* pqinfoPool = pQueryMgmt->qinfoPool;
+  pQueryMgmt->qinfoPool = NULL;
+
+  taosCacheCleanup(pqinfoPool);
+  pthread_mutex_destroy(&pQueryMgmt->lock);
+  tfree(pQueryMgmt);
+
+  qDebug("vgId:%d, queryMgmt cleanup completed", vgId);
+}
+
+void** qRegisterQInfo(void* pMgmt, uint64_t qInfo) {
+  if (pMgmt == NULL) {
+    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
+    return NULL;
+  }
+
+  SQueryMgmt *pQueryMgmt = pMgmt;
+  if (pQueryMgmt->qinfoPool == NULL) {
+    qError("QInfo:%p failed to add qhandle into qMgmt, since qMgmt is closed", (void *)qInfo);
+    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
+    return NULL;
+  }
+
+  pthread_mutex_lock(&pQueryMgmt->lock);
+  if (pQueryMgmt->closed) {
+    pthread_mutex_unlock(&pQueryMgmt->lock);
+    qError("QInfo:%p failed to add qhandle into cache, since qMgmt is colsing", (void *)qInfo);
+    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
+    return NULL;
+  } else {
+    TSDB_CACHE_PTR_TYPE handleVal = (TSDB_CACHE_PTR_TYPE) qInfo;
+    void** handle = taosCachePut(pQueryMgmt->qinfoPool, &handleVal, sizeof(TSDB_CACHE_PTR_TYPE), &qInfo, sizeof(TSDB_CACHE_PTR_TYPE),
+                                 (getMaximumIdleDurationSec()*1000));
+    pthread_mutex_unlock(&pQueryMgmt->lock);
+
+    return handle;
+  }
+}
+
+void** qAcquireQInfo(void* pMgmt, uint64_t _key) {
+  SQueryMgmt *pQueryMgmt = pMgmt;
+
+  if (pQueryMgmt->closed) {
+    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
+    return NULL;
+  }
+
+  if (pQueryMgmt->qinfoPool == NULL) {
+    terrno = TSDB_CODE_QRY_INVALID_QHANDLE;
+    return NULL;
+  }
+
+  TSDB_CACHE_PTR_TYPE key = (TSDB_CACHE_PTR_TYPE)_key;
+  void** handle = taosCacheAcquireByKey(pQueryMgmt->qinfoPool, &key, sizeof(TSDB_CACHE_PTR_TYPE));
+  if (handle == NULL || *handle == NULL) {
+    terrno = TSDB_CODE_QRY_INVALID_QHANDLE;
+    return NULL;
+  } else {
+    return handle;
+  }
+}
+
+void** qReleaseQInfo(void* pMgmt, void* pQInfo, bool freeHandle) {
+  SQueryMgmt *pQueryMgmt = pMgmt;
+  if (pQueryMgmt->qinfoPool == NULL) {
+    return NULL;
+  }
+
+  taosCacheRelease(pQueryMgmt->qinfoPool, pQInfo, freeHandle);
+  return 0;
+}

src/util/inc/tarray.h

@@ -125,7 +125,7 @@ void taosArrayRemove(SArray* pArray, size_t index);
  * @param pDst
  * @param pSrc
  */
-void taosArrayCopy(SArray* pDst, const SArray* pSrc);
+SArray* taosArrayFromList(const void* src, size_t size, size_t elemSize);
 
 /**
  * clone a new array

src/util/src/tarray.c

@@ -156,23 +156,14 @@ void taosArrayRemove(SArray* pArray, size_t index) {
   pArray->size -= 1;
 }
 
-void taosArrayCopy(SArray* pDst, const SArray* pSrc) {
-  assert(pSrc != NULL && pDst != NULL);
+SArray* taosArrayFromList(const void* src, size_t size, size_t elemSize) {
+  assert(src != NULL && elemSize > 0);
+  SArray* pDst = taosArrayInit(size, elemSize);
 
-  if (pDst->capacity < pSrc->size) {
-    void* pData = realloc(pDst->pData, pSrc->size * pSrc->elemSize);
-    if (pData == NULL) { // todo handle oom
+  memcpy(pDst->pData, src, elemSize * size);
+  pDst->size = size;
 
-    } else {
-      pDst->pData = pData;
-      pDst->capacity = pSrc->size;
-    }
-  }
-  
-  memcpy(pDst->pData, pSrc->pData, pSrc->elemSize * pSrc->size);
-  pDst->elemSize = pSrc->elemSize;
-  pDst->capacity = pSrc->size;
-  pDst->size = pSrc->size;
+  return pDst;
 }
 
 SArray* taosArrayDup(const SArray* pSrc) {

tests/script/general/parser/limit2_query.sim

@@ -143,6 +143,97 @@ if $data11 != -1 then
   return -1
 endi
 
+sql select max(c1) from lm2_tb0 where ts >= 1537146000000 and ts <= 1543145400000 interval(5m) fill(value, -1000, -2) limit 8200
+if $rows != 8200 then
+  return -1
+endi
+
+sql select max(c1) from lm2_tb0 where ts >= 1537146000000 and ts <= 1543145400000 interval(5m) fill(value, -1000, -2) limit 10 offset 8190;
+if $rows != 10 then
+  return -1
+endi
+
+if $data00 != @18-10-15 19:30:00.000@ then
+  return -1
+endi
+
+if $data01 != 5 then
+  return -1
+endi
+
+if $data10 != @18-10-15 19:35:00.000@ then
+  return -1
+endi
+
+if $data11 != -1000 then
+  return -1
+endi
+
+if $data20 != @18-10-15 19:40:00.000@ then
+  return -1
+endi
+
+if $data21 != 6 then
+  return -1
+endi
+
+if $data30 != @18-10-15 19:45:00.000@ then
+  return -1
+endi
+
+if $data31 != -1000 then
+  return -1
+endi
+
+sql select max(c1) from lm2_tb0 where ts >= 1537146000000 and ts <= 1543145400000 interval(5m) fill(value, -1000, -2) limit 10 offset 10001;
+if $rows != 10 then
+  return -1
+endi
+
+if $data00 != @18-10-22 02:25:00.000@ then
+  return -1
+endi
+
+if $data01 != -1000 then
+  return -1
+endi
+
+if $data10 != @18-10-22 02:30:00.000@ then
+  return -1
+endi
+
+if $data11 != 1 then
+  return -1
+endi
+
+if $data20 != @18-10-22 02:35:00.000@ then
+  return -1
+endi
+
+if $data21 != -1000 then
+  return -1
+endi
+
+if $data30 != @18-10-22 02:40:00.000@ then
+  return -1
+endi
+
+if $data31 != 2 then
+  return -1
+endi
+
+sql select max(c1) from lm2_tb0 where ts >= 1537146000000 and ts <= 1543145400000 interval(5m) fill(value, -1000, -2) limit 10000 offset 10001;
+print ====> needs to validate the last row result
+if $rows != 9998 then
+  return -1
+endi
+
+
+sql select max(c1) from lm2_tb0 where ts >= 1537146000000 and ts <= 1543145400000 interval(5m) fill(value, -1000, -2) limit 100 offset 20001;
+if $rows != 0 then
+  return -1
+endi
+
 # tb + interval + fill(linear) + limit offset
 $limit = $rowNum
 $offset = $limit / 2

tests/script/general/parser/testSuite.sim

@@ -53,32 +53,32 @@
 #run general/parser/limit1_tblocks100.sim
 #sleep 100
 #run general/parser/limit2.sim
-#sleep 100
-#run general/parser/mixed_blocks.sim
-#sleep 100
-#run general/parser/nchar.sim
-#sleep 100
-#run general/parser/null_char.sim
-#sleep 100
-#run general/parser/selectResNum.sim
-#sleep 100
-#run general/parser/select_across_vnodes.sim
-#sleep 100
-#run general/parser/select_from_cache_disk.sim
-#sleep 100
-#run general/parser/set_tag_vals.sim
-#sleep 100
-#run general/parser/single_row_in_tb.sim
-#sleep 100
-#run general/parser/slimit.sim
-#sleep 100
-#run general/parser/slimit1.sim
-#sleep 100
-#run general/parser/slimit_alter_tags.sim
-#sleep 100
-#run general/parser/tbnameIn.sim
-#sleep 100
-#run general/parser/slimit_alter_tags.sim  # persistent failed
+sleep 100
+run general/parser/mixed_blocks.sim
+sleep 100
+run general/parser/nchar.sim
+sleep 100
+run general/parser/null_char.sim
+sleep 100
+run general/parser/selectResNum.sim
+sleep 100
+run general/parser/select_across_vnodes.sim
+sleep 100
+run general/parser/select_from_cache_disk.sim
+sleep 100
+run general/parser/set_tag_vals.sim
+sleep 100
+run general/parser/single_row_in_tb.sim
+sleep 100
+run general/parser/slimit.sim
+sleep 100
+run general/parser/slimit1.sim
+sleep 100
+run general/parser/slimit_alter_tags.sim
+sleep 100
+run general/parser/tbnameIn.sim
+sleep 100
+run general/parser/slimit_alter_tags.sim  # persistent failed
 sleep 100
 run general/parser/join.sim
 sleep 100