diff --git a/cmake/platform.inc b/cmake/platform.inc
index 2e0e2d6af08fa529f7435b2e39fa4cdb4d293fae..7bad9bc4f36dd61d449054c8d85f9c5959cbf7d6 100755
--- a/cmake/platform.inc
+++ b/cmake/platform.inc
@@ -109,7 +109,7 @@ IF (TD_LINUX_64)
       IF (${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
         SET(COMMON_FLAGS "-std=gnu99 -Wall -Werror -Wno-missing-braces -fPIC -g3 -gdwarf-2 -msse4.2 -D_FILE_OFFSET_BITS=64 -D_LARGE_FILE")
       ELSE ()
-        SET(COMMON_FLAGS "-std=gnu99 -Wall -Werror -fPIC -malign-double -g3 -gdwarf-2 -malign-stringops -msse4.2 -D_FILE_OFFSET_BITS=64 -D_LARGE_FILE")
+        SET(COMMON_FLAGS "-std=gnu99 -Wall -Werror -pg -fPIC -malign-double -g3 -gdwarf-2 -malign-stringops -msse4.2 -D_FILE_OFFSET_BITS=64 -D_LARGE_FILE")
       ENDIF ()
     ELSE ()
       SET(COMMON_FLAGS "-std=gnu99 -Wall -Werror -fPIC -g -fsigned-char -fpack-struct=8 -D_FILE_OFFSET_BITS=64 -D_LARGE_FILE")
diff --git a/src/client/inc/tscUtil.h b/src/client/inc/tscUtil.h
index 499269210947d458cc0295d4c1d49e3117a046d6..4af929bf4146c14b180b7055b6e58a55aeb64863 100644
--- a/src/client/inc/tscUtil.h
+++ b/src/client/inc/tscUtil.h
@@ -160,7 +160,9 @@ void tscFieldInfoUpdateOffsetForInterResult(SQueryInfo* pQueryInfo);
 
 int16_t tscFieldInfoGetOffset(SQueryInfo* pQueryInfo, int32_t index);
 void    tscFieldInfoClear(SFieldInfo* pFieldInfo);
-int32_t tscNumOfFields(SQueryInfo* pQueryInfo);
+
+static FORCE_INLINE int32_t tscNumOfFields(SQueryInfo* pQueryInfo) { return pQueryInfo->fieldsInfo.numOfOutput; }
+
 int32_t tscFieldInfoCompare(const SFieldInfo* pFieldInfo1, const SFieldInfo* pFieldInfo2);
 
 void addExprParams(SSqlExpr* pExpr, char* argument, int32_t type, int32_t bytes, int16_t tableIndex);
diff --git a/src/client/inc/tsclient.h b/src/client/inc/tsclient.h
index c8754e5bebd7a460ed3e33a5e56a1e535dbf7035..b44ea3b32529b5ede460a25458fa1ebc92792fc0 100644
--- a/src/client/inc/tsclient.h
+++ b/src/client/inc/tsclient.h
@@ -411,7 +411,44 @@ char *tscGetErrorMsgPayload(SSqlCmd *pCmd);
 int32_t tscInvalidSQLErrMsg(char *msg, const char *additionalInfo, const char *sql);
 
 int32_t tscToSQLCmd(SSqlObj *pSql, struct SSqlInfo *pInfo);
-void    tscGetResultColumnChr(SSqlRes *pRes, SFieldInfo* pFieldInfo, int32_t column);
+//void    tscGetResultColumnChr(SSqlRes *pRes, SFieldInfo* pFieldInfo, int32_t column);
+
+static FORCE_INLINE void tscGetResultColumnChr(SSqlRes* pRes, SFieldInfo* pFieldInfo, int32_t columnIndex) {
+  SFieldSupInfo* pInfo = TARRAY_GET_ELEM(pFieldInfo->pSupportInfo, columnIndex);
+  assert(pInfo->pSqlExpr != NULL);
+
+  int32_t type = pInfo->pSqlExpr->resType;
+  int32_t bytes = pInfo->pSqlExpr->resBytes;
+
+  char* pData = pRes->data + pInfo->pSqlExpr->offset * pRes->numOfRows + bytes * pRes->row;
+
+  if (type == TSDB_DATA_TYPE_NCHAR || type == TSDB_DATA_TYPE_BINARY) {
+    int32_t realLen = varDataLen(pData);
+    assert(realLen <= bytes - VARSTR_HEADER_SIZE);
+
+    if (isNull(pData, type)) {
+      pRes->tsrow[columnIndex] = NULL;
+    } else {
+      pRes->tsrow[columnIndex] = ((tstr*)pData)->data;
+    }
+
+    if (realLen < pInfo->pSqlExpr->resBytes - VARSTR_HEADER_SIZE) { // todo refactor
+      *(pData + realLen + VARSTR_HEADER_SIZE) = 0;
+    }
+
+    pRes->length[columnIndex] = realLen;
+  } else {
+    assert(bytes == tDataTypeDesc[type].nSize);
+
+    if (isNull(pData, type)) {
+      pRes->tsrow[columnIndex] = NULL;
+    } else {
+      pRes->tsrow[columnIndex] = pData;
+    }
+
+    pRes->length[columnIndex] = bytes;
+  }
+}
 
 extern void *    tscCacheHandle;
 extern void *    tscTmr;
diff --git a/src/client/src/tscFunctionImpl.c b/src/client/src/tscFunctionImpl.c
index 909338aa4a20b50ef4a12c0fee22e39a2ea41cfc..bc37c16187c5c6a790779ea9f25f739daa545f27 100644
--- a/src/client/src/tscFunctionImpl.c
+++ b/src/client/src/tscFunctionImpl.c
@@ -2952,10 +2952,14 @@ static void tag_project_function(SQLFunctionCtx *pCtx) {
   INC_INIT_VAL(pCtx, pCtx->size);
   
   assert(pCtx->inputBytes == pCtx->outputBytes);
-  
-  for (int32_t i = 0; i < pCtx->size; ++i) {
-    tVariantDump(&pCtx->tag, pCtx->aOutputBuf, pCtx->outputType, true);
-    
+
+  tVariantDump(&pCtx->tag, pCtx->aOutputBuf, pCtx->outputType, true);
+  char* data = pCtx->aOutputBuf;
+  pCtx->aOutputBuf += pCtx->outputBytes;
+
+  // directly copy from the first one
+  for (int32_t i = 1; i < pCtx->size; ++i) {
+    memmove(pCtx->aOutputBuf, data, pCtx->outputBytes);
     pCtx->aOutputBuf += pCtx->outputBytes;
   }
 }
diff --git a/src/client/src/tscSql.c b/src/client/src/tscSql.c
index f9f93b3f894d400abced33141d468398fc9a20fc..6d75aef01f085226fb27ba81ceb611b7c97527cc 100644
--- a/src/client/src/tscSql.c
+++ b/src/client/src/tscSql.c
@@ -403,7 +403,7 @@ TAOS_ROW taos_fetch_row(TAOS_RES *res) {
     taos_fetch_rows_a(res, waitForRetrieveRsp, pSql->pTscObj);
     sem_wait(&pSql->rspSem);
   }
-  
+
   return doSetResultRowData(pSql, true);
 }
 
diff --git a/src/client/src/tscSubquery.c b/src/client/src/tscSubquery.c
index b5b659de0cccf69e71ee579e404378bedbbac4d0..044a2aba9d7782ecab43964e8e52ba8adbf7fb80 100644
--- a/src/client/src/tscSubquery.c
+++ b/src/client/src/tscSubquery.c
@@ -2108,9 +2108,9 @@ static char *getArithemicInputSrc(void *param, const char *name, int32_t colId)
 void **doSetResultRowData(SSqlObj *pSql, bool finalResult) {
   SSqlCmd *pCmd = &pSql->cmd;
   SSqlRes *pRes = &pSql->res;
-  
+
   assert(pRes->row >= 0 && pRes->row <= pRes->numOfRows);
-  
+
   if(pCmd->command == TSDB_SQL_TABLE_JOIN_RETRIEVE) {
     if (pRes->completed) {
       tfree(pRes->tsrow);
@@ -2118,29 +2118,31 @@ void **doSetResultRowData(SSqlObj *pSql, bool finalResult) {
 
     return pRes->tsrow;
   }
-  
+
   if (pRes->row >= pRes->numOfRows) {  // all the results has returned to invoker
     tfree(pRes->tsrow);
     return pRes->tsrow;
   }
-  
+
   SQueryInfo *pQueryInfo = tscGetQueryInfoDetail(pCmd, pCmd->clauseIndex);
-  
+
   size_t size = tscNumOfFields(pQueryInfo);
   for (int i = 0; i < size; ++i) {
-    SFieldSupInfo* pSup = tscFieldInfoGetSupp(&pQueryInfo->fieldsInfo, i);
+    SFieldSupInfo* pSup = TARRAY_GET_ELEM(pQueryInfo->fieldsInfo.pSupportInfo, i);
     if (pSup->pSqlExpr != NULL) {
       tscGetResultColumnChr(pRes, &pQueryInfo->fieldsInfo, i);
     }
-    
+
     // primary key column cannot be null in interval query, no need to check
     if (i == 0 && pQueryInfo->intervalTime > 0) {
       continue;
     }
-    
-    TAOS_FIELD *pField = tscFieldInfoGetField(&pQueryInfo->fieldsInfo, i);
-    transferNcharData(pSql, i, pField);
-    
+
+    TAOS_FIELD *pField = TARRAY_GET_ELEM(pQueryInfo->fieldsInfo.pFields, i);
+    if (pRes->tsrow[i] != NULL && pField->type == TSDB_DATA_TYPE_NCHAR) {
+      transferNcharData(pSql, i, pField);
+    }
+
     // calculate the result from several other columns
     if (pSup->pArithExprInfo != NULL) {
       if (pRes->pArithSup == NULL) {
@@ -2150,10 +2152,10 @@ void **doSetResultRowData(SSqlObj *pSql, bool finalResult) {
         sas->numOfCols  = tscSqlExprNumOfExprs(pQueryInfo);
         sas->exprList   = pQueryInfo->exprList;
         sas->data       = calloc(sas->numOfCols, POINTER_BYTES);
-        
+
         pRes->pArithSup = sas;
       }
-      
+
       if (pRes->buffer[i] == NULL) {
         TAOS_FIELD* field = tscFieldInfoGetField(&pQueryInfo->fieldsInfo, i);
         pRes->buffer[i] = malloc(field->bytes);
@@ -2163,13 +2165,13 @@ void **doSetResultRowData(SSqlObj *pSql, bool finalResult) {
         SSqlExpr* pExpr = tscSqlExprGet(pQueryInfo, k);
         pRes->pArithSup->data[k] = (pRes->data + pRes->numOfRows* pExpr->offset) + pRes->row*pExpr->resBytes;
       }
-  
+
       tExprTreeCalcTraverse(pRes->pArithSup->pArithExpr->pExpr, 1, pRes->buffer[i], pRes->pArithSup,
           TSDB_ORDER_ASC, getArithemicInputSrc);
       pRes->tsrow[i] = pRes->buffer[i];
     }
   }
-  
+
   pRes->row++;  // index increase one-step
   return pRes->tsrow;
 }
diff --git a/src/client/src/tscUtil.c b/src/client/src/tscUtil.c
index 54c9cdea65123da9aee5c2932aff11aa2ed1e59d..dea41d29321f288ee2f4ef7399d75eea0c149b61 100644
--- a/src/client/src/tscUtil.c
+++ b/src/client/src/tscUtil.c
@@ -794,7 +794,7 @@ SFieldSupInfo* tscFieldInfoAppend(SFieldInfo* pFieldInfo, TAOS_FIELD* pField) {
 }
 
 SFieldSupInfo* tscFieldInfoGetSupp(SFieldInfo* pFieldInfo, int32_t index) {
-  return taosArrayGet(pFieldInfo->pSupportInfo, index);
+  return TARRAY_GET_ELEM(pFieldInfo->pSupportInfo, index);
 }
 
 SFieldSupInfo* tscFieldInfoInsert(SFieldInfo* pFieldInfo, int32_t index, TAOS_FIELD* field) {
@@ -858,11 +858,9 @@ void tscFieldInfoCopy(SFieldInfo* dst, const SFieldInfo* src) {
 }
 
 TAOS_FIELD* tscFieldInfoGetField(SFieldInfo* pFieldInfo, int32_t index) {
-  return taosArrayGet(pFieldInfo->pFields, index);
+  return TARRAY_GET_ELEM(pFieldInfo->pFields, index);
 }
 
-int32_t tscNumOfFields(SQueryInfo* pQueryInfo) { return pQueryInfo->fieldsInfo.numOfOutput; }
-
 int16_t tscFieldInfoGetOffset(SQueryInfo* pQueryInfo, int32_t index) {
   SFieldSupInfo* pInfo = tscFieldInfoGetSupp(&pQueryInfo->fieldsInfo, index);
   assert(pInfo != NULL);
@@ -2086,42 +2084,42 @@ void tscTryQueryNextClause(SSqlObj* pSql, void (*queryFp)()) {
   }
 }
 
-void tscGetResultColumnChr(SSqlRes* pRes, SFieldInfo* pFieldInfo, int32_t columnIndex) {
-  SFieldSupInfo* pInfo = taosArrayGet(pFieldInfo->pSupportInfo, columnIndex);
-  assert(pInfo->pSqlExpr != NULL);
-
-  int32_t type = pInfo->pSqlExpr->resType;
-  int32_t bytes = pInfo->pSqlExpr->resBytes;
-  
-  char* pData = pRes->data + pInfo->pSqlExpr->offset * pRes->numOfRows + bytes * pRes->row;
-  
-  if (type == TSDB_DATA_TYPE_NCHAR || type == TSDB_DATA_TYPE_BINARY) {
-    int32_t realLen = varDataLen(pData);
-    assert(realLen <= bytes - VARSTR_HEADER_SIZE);
-    
-    if (isNull(pData, type)) {
-      pRes->tsrow[columnIndex] = NULL;
-    } else {
-      pRes->tsrow[columnIndex] = ((tstr*)pData)->data;
-    }
-  
-    if (realLen < pInfo->pSqlExpr->resBytes - VARSTR_HEADER_SIZE) { // todo refactor
-      *(pData + realLen + VARSTR_HEADER_SIZE) = 0;
-    }
-    
-    pRes->length[columnIndex] = realLen;
-  } else {
-    assert(bytes == tDataTypeDesc[type].nSize);
-    
-    if (isNull(pData, type)) {
-      pRes->tsrow[columnIndex] = NULL;
-    } else {
-      pRes->tsrow[columnIndex] = pData;
-    }
-    
-    pRes->length[columnIndex] = bytes;
-  }
-}
+//void tscGetResultColumnChr(SSqlRes* pRes, SFieldInfo* pFieldInfo, int32_t columnIndex) {
+//  SFieldSupInfo* pInfo = TARRAY_GET_ELEM(pFieldInfo->pSupportInfo, columnIndex);
+//  assert(pInfo->pSqlExpr != NULL);
+//
+//  int32_t type = pInfo->pSqlExpr->resType;
+//  int32_t bytes = pInfo->pSqlExpr->resBytes;
+//
+//  char* pData = pRes->data + pInfo->pSqlExpr->offset * pRes->numOfRows + bytes * pRes->row;
+//
+//  if (type == TSDB_DATA_TYPE_NCHAR || type == TSDB_DATA_TYPE_BINARY) {
+//    int32_t realLen = varDataLen(pData);
+//    assert(realLen <= bytes - VARSTR_HEADER_SIZE);
+//
+//    if (isNull(pData, type)) {
+//      pRes->tsrow[columnIndex] = NULL;
+//    } else {
+//      pRes->tsrow[columnIndex] = ((tstr*)pData)->data;
+//    }
+//
+//    if (realLen < pInfo->pSqlExpr->resBytes - VARSTR_HEADER_SIZE) { // todo refactor
+//      *(pData + realLen + VARSTR_HEADER_SIZE) = 0;
+//    }
+//
+//    pRes->length[columnIndex] = realLen;
+//  } else {
+//    assert(bytes == tDataTypeDesc[type].nSize);
+//
+//    if (isNull(pData, type)) {
+//      pRes->tsrow[columnIndex] = NULL;
+//    } else {
+//      pRes->tsrow[columnIndex] = pData;
+//    }
+//
+//    pRes->length[columnIndex] = bytes;
+//  }
+//}
 
 void* malloc_throw(size_t size) {
   void* p = malloc(size);
diff --git a/src/common/src/ttypes.c b/src/common/src/ttypes.c
index eff25e8c93047a174dddeb1f4b281942e94959cb..44b1d9fc385595638098ad098f51245a46a4bfef 100644
--- a/src/common/src/ttypes.c
+++ b/src/common/src/ttypes.c
@@ -367,31 +367,31 @@ bool isValidDataType(int32_t type) {
   return type >= TSDB_DATA_TYPE_NULL && type <= TSDB_DATA_TYPE_NCHAR;
 }
 
-bool isNull(const char *val, int32_t type) {
-  switch (type) {
-    case TSDB_DATA_TYPE_BOOL:
-      return *(uint8_t *)val == TSDB_DATA_BOOL_NULL;
-    case TSDB_DATA_TYPE_TINYINT:
-      return *(uint8_t *)val == TSDB_DATA_TINYINT_NULL;
-    case TSDB_DATA_TYPE_SMALLINT:
-      return *(uint16_t *)val == TSDB_DATA_SMALLINT_NULL;
-    case TSDB_DATA_TYPE_INT:
-      return *(uint32_t *)val == TSDB_DATA_INT_NULL;
-    case TSDB_DATA_TYPE_BIGINT:
-    case TSDB_DATA_TYPE_TIMESTAMP:
-      return *(uint64_t *)val == TSDB_DATA_BIGINT_NULL;
-    case TSDB_DATA_TYPE_FLOAT:
-      return *(uint32_t *)val == TSDB_DATA_FLOAT_NULL;
-    case TSDB_DATA_TYPE_DOUBLE:
-      return *(uint64_t *)val == TSDB_DATA_DOUBLE_NULL;
-    case TSDB_DATA_TYPE_NCHAR:
-      return *(uint32_t*) varDataVal(val) == TSDB_DATA_NCHAR_NULL;
-    case TSDB_DATA_TYPE_BINARY:
-      return *(uint8_t *) varDataVal(val) == TSDB_DATA_BINARY_NULL;
-    default:
-      return false;
-  };
-}
+//bool isNull(const char *val, int32_t type) {
+//  switch (type) {
+//    case TSDB_DATA_TYPE_BOOL:
+//      return *(uint8_t *)val == TSDB_DATA_BOOL_NULL;
+//    case TSDB_DATA_TYPE_TINYINT:
+//      return *(uint8_t *)val == TSDB_DATA_TINYINT_NULL;
+//    case TSDB_DATA_TYPE_SMALLINT:
+//      return *(uint16_t *)val == TSDB_DATA_SMALLINT_NULL;
+//    case TSDB_DATA_TYPE_INT:
+//      return *(uint32_t *)val == TSDB_DATA_INT_NULL;
+//    case TSDB_DATA_TYPE_BIGINT:
+//    case TSDB_DATA_TYPE_TIMESTAMP:
+//      return *(uint64_t *)val == TSDB_DATA_BIGINT_NULL;
+//    case TSDB_DATA_TYPE_FLOAT:
+//      return *(uint32_t *)val == TSDB_DATA_FLOAT_NULL;
+//    case TSDB_DATA_TYPE_DOUBLE:
+//      return *(uint64_t *)val == TSDB_DATA_DOUBLE_NULL;
+//    case TSDB_DATA_TYPE_NCHAR:
+//      return *(uint32_t*) varDataVal(val) == TSDB_DATA_NCHAR_NULL;
+//    case TSDB_DATA_TYPE_BINARY:
+//      return *(uint8_t *) varDataVal(val) == TSDB_DATA_BINARY_NULL;
+//    default:
+//      return false;
+//  };
+//}
 
 void setVardataNull(char* val, int32_t type) {
   if (type == TSDB_DATA_TYPE_BINARY) {
diff --git a/src/inc/taosdef.h b/src/inc/taosdef.h
index e4ee058cef2987819640a43768edf7f483b8c1bc..ac5b92b4abb25d37af1fac7ecb47855a50fe80e4 100644
--- a/src/inc/taosdef.h
+++ b/src/inc/taosdef.h
@@ -160,7 +160,32 @@ extern tDataTypeDescriptor tDataTypeDesc[11];
 #define POINTER_BYTES sizeof(void *)  // 8 by default  assert(sizeof(ptrdiff_t) == sizseof(void*)
 
 bool isValidDataType(int32_t type);
-bool isNull(const char *val, int32_t type);
+//bool isNull(const char *val, int32_t type);
+static inline __attribute__((always_inline)) bool isNull(const char *val, int32_t type) {
+  switch (type) {
+    case TSDB_DATA_TYPE_BOOL:
+      return *(uint8_t *)val == TSDB_DATA_BOOL_NULL;
+    case TSDB_DATA_TYPE_TINYINT:
+      return *(uint8_t *)val == TSDB_DATA_TINYINT_NULL;
+    case TSDB_DATA_TYPE_SMALLINT:
+      return *(uint16_t *)val == TSDB_DATA_SMALLINT_NULL;
+    case TSDB_DATA_TYPE_INT:
+      return *(uint32_t *)val == TSDB_DATA_INT_NULL;
+    case TSDB_DATA_TYPE_BIGINT:
+    case TSDB_DATA_TYPE_TIMESTAMP:
+      return *(uint64_t *)val == TSDB_DATA_BIGINT_NULL;
+    case TSDB_DATA_TYPE_FLOAT:
+      return *(uint32_t *)val == TSDB_DATA_FLOAT_NULL;
+    case TSDB_DATA_TYPE_DOUBLE:
+      return *(uint64_t *)val == TSDB_DATA_DOUBLE_NULL;
+    case TSDB_DATA_TYPE_NCHAR:
+      return *(uint32_t*) varDataVal(val) == TSDB_DATA_NCHAR_NULL;
+    case TSDB_DATA_TYPE_BINARY:
+      return *(uint8_t *) varDataVal(val) == TSDB_DATA_BINARY_NULL;
+    default:
+      return false;
+  };
+}
 
 void setVardataNull(char* val, int32_t type);
 void setNull(char *val, int32_t type, int32_t bytes);
diff --git a/src/query/src/qExecutor.c b/src/query/src/qExecutor.c
index dfe489e1e8df4233365fcc36e6f88c242472247c..f55678582460af4618204cc255c8e0a121d416b6 100644
--- a/src/query/src/qExecutor.c
+++ b/src/query/src/qExecutor.c
@@ -44,7 +44,7 @@
 #define SET_MASTER_SCAN_FLAG(runtime)  ((runtime)->scanFlag = MASTER_SCAN)
 #define SET_REVERSE_SCAN_FLAG(runtime) ((runtime)->scanFlag = REVERSE_SCAN)
 
-#define GET_QINFO_ADDR(x) ((void *)((char *)(x)-offsetof(SQInfo, runtimeEnv)))
+#define GET_QINFO_ADDR(x) ((SQInfo *)((char *)(x)-offsetof(SQInfo, runtimeEnv)))
 
 #define GET_COL_DATA_POS(query, index, step) ((query)->pos + (index) * (step))
 #define SWITCH_ORDER(n) (((n) = ((n) == TSDB_ORDER_ASC) ? TSDB_ORDER_DESC : TSDB_ORDER_ASC))
@@ -351,27 +351,6 @@ static bool hasTagValOutput(SQuery* pQuery) {
   return false;
 }
 
-static SDataStatis *getStatisInfo(SQuery *pQuery, SDataStatis *pStatis, int32_t numOfCols, int32_t index) {
-  // for a tag column, no corresponding field info
-  SColIndex *pColIndex = &pQuery->pSelectExpr[index].base.colInfo;
-  if (TSDB_COL_IS_TAG(pColIndex->flag)) {
-    return NULL;
-  }
-  
-  /*
-   * Choose the right column field info by field id, since the file block may be out of date,
-   * which means the newest table schema is not equalled to the schema of this block.
-   * TODO: speedup by using bsearch
-   */
-  for (int32_t i = 0; i < numOfCols; ++i) {
-    if (pColIndex->colId == pStatis[i].colId) {
-      return &pStatis[i];
-    }
-  }
-  
-  return NULL;
-}
-
 /**
  * @param pQuery
  * @param col
@@ -380,19 +359,14 @@ static SDataStatis *getStatisInfo(SQuery *pQuery, SDataStatis *pStatis, int32_t
  * @param pColStatis
  * @return
  */
-static bool hasNullValue(SQuery *pQuery, int32_t col, int32_t numOfCols, SDataStatis *pStatis, SDataStatis **pColStatis) {
-  SColIndex *pColIndex = &pQuery->pSelectExpr[col].base.colInfo;
-  if (TSDB_COL_IS_TAG(pColIndex->flag)) {
-    return false;
-  }
-
-  // query on primary timestamp column, not null value at all
-  if (pColIndex->colId == PRIMARYKEY_TIMESTAMP_COL_INDEX) {
+static bool hasNullValue(SColIndex* pColIndex, SDataStatis *pStatis, SDataStatis **pColStatis) {
+  if (TSDB_COL_IS_TAG(pColIndex->flag) || pColIndex->colId == PRIMARYKEY_TIMESTAMP_COL_INDEX) {
     return false;
   }
 
   if (pStatis != NULL) {
-    *pColStatis = getStatisInfo(pQuery, pStatis, numOfCols, col);
+    *pColStatis = &pStatis[pColIndex->colIndex];
+    assert((*pColStatis)->colId == pColIndex->colId);
   } else {
     *pColStatis = NULL;
   }
@@ -842,8 +816,8 @@ static char *getDataBlock(SQueryRuntimeEnv *pRuntimeEnv, SArithmeticSupport *sas
   if (pDataBlock == NULL) {
     return NULL;
   }
-  char *dataBlock = NULL;
 
+  char *dataBlock = NULL;
   SQuery *pQuery = pRuntimeEnv->pQuery;
   SQLFunctionCtx *pCtx = pRuntimeEnv->pCtx;
 
@@ -887,10 +861,14 @@ static char *getDataBlock(SQueryRuntimeEnv *pRuntimeEnv, SArithmeticSupport *sas
 
   } else {  // other type of query function
     SColIndex *pCol = &pQuery->pSelectExpr[col].base.colInfo;
-    if (TSDB_COL_IS_TAG(pCol->flag) || pDataBlock == NULL) {
+    if (TSDB_COL_IS_TAG(pCol->flag)) {
       dataBlock = NULL;
     } else {
-      dataBlock = getDataBlockImpl(pDataBlock, pCol->colId);
+      SColIndex* pColIndex = &pQuery->pSelectExpr[col].base.colInfo;
+      SColumnInfoData *p = taosArrayGet(pDataBlock, pColIndex->colIndex);
+      assert(p->info.colId == pColIndex->colId);
+
+      dataBlock = p->pData;
     }
   }
 
@@ -1365,7 +1343,7 @@ void setExecParams(SQuery *pQuery, SQLFunctionCtx *pCtx, void* inputData, TSKEY
   int32_t colId = pQuery->pSelectExpr[colIndex].base.colInfo.colId;
   
   SDataStatis *tpField = NULL;
-  pCtx->hasNull = hasNullValue(pQuery, colIndex, pBlockInfo->numOfCols, pStatis, &tpField);
+  pCtx->hasNull = hasNullValue(&pQuery->pSelectExpr[colIndex].base.colInfo, pStatis, &tpField);
   pCtx->aInputElemBuf = inputData;
 
   if (tpField != NULL) {
@@ -1622,19 +1600,18 @@ static void teardownQueryRuntimeEnv(SQueryRuntimeEnv *pRuntimeEnv) {
   pRuntimeEnv->pTSBuf = tsBufDestory(pRuntimeEnv->pTSBuf);
 }
 
-static bool isQueryKilled(SQInfo *pQInfo) {
-  return (pQInfo->code == TSDB_CODE_TSC_QUERY_CANCELLED);
-}
+#define IS_QUERY_KILLED(_q) ((_q)->code == TSDB_CODE_TSC_QUERY_CANCELLED)
 
 static void setQueryKilled(SQInfo *pQInfo) { pQInfo->code = TSDB_CODE_TSC_QUERY_CANCELLED; }
 
-static bool isFixedOutputQuery(SQuery *pQuery) {
-  if (pQuery->intervalTime != 0) {
+static bool isFixedOutputQuery(SQueryRuntimeEnv* pRuntimeEnv) {
+  SQuery* pQuery = pRuntimeEnv->pQuery;
+  if (QUERY_IS_INTERVAL_QUERY(pQuery)) {
     return false;
   }
 
   // Note:top/bottom query is fixed output query
-  if (isTopBottomQuery(pQuery) || isGroupbyNormalCol(pQuery->pGroupbyExpr)) {
+  if (pRuntimeEnv->topBotQuery || pRuntimeEnv->groupbyNormalCol) {
     return true;
   }
 
@@ -2054,7 +2031,7 @@ SArray *loadDataBlockOnDemand(SQueryRuntimeEnv *pRuntimeEnv, void* pQueryHandle,
     // check if this data block is required to load
     for (int32_t i = 0; i < pQuery->numOfOutput; ++i) {
       SSqlFuncMsg* pSqlFunc = &pQuery->pSelectExpr[i].base;
-      
+
       int32_t functionId = pSqlFunc->functionId;
       int32_t colId = pSqlFunc->colInfo.colId;
       r |= aAggs[functionId].dataReqFunc(&pRuntimeEnv->pCtx[i], pQuery->window.skey, pQuery->window.ekey, colId);
@@ -2066,8 +2043,7 @@ SArray *loadDataBlockOnDemand(SQueryRuntimeEnv *pRuntimeEnv, void* pQueryHandle,
   }
 
   if (r == BLK_DATA_NO_NEEDED) {
-    qDebug("QInfo:%p data block discard, brange:%" PRId64 "-%" PRId64 ", rows:%d", GET_QINFO_ADDR(pRuntimeEnv),
-           pBlockInfo->window.skey, pBlockInfo->window.ekey, pBlockInfo->rows);
+    qDebug("QInfo:%p data block discard, rows:%d", GET_QINFO_ADDR(pRuntimeEnv), pBlockInfo->rows);
     pRuntimeEnv->summary.discardBlocks += 1;
   } else if (r == BLK_DATA_STATIS_NEEDED) {
     if (tsdbRetrieveDataBlockStatisInfo(pQueryHandle, pStatis) != TSDB_CODE_SUCCESS) {
@@ -2199,7 +2175,7 @@ static void ensureOutputBufferSimple(SQueryRuntimeEnv* pRuntimeEnv, int32_t capa
 static void ensureOutputBuffer(SQueryRuntimeEnv* pRuntimeEnv, SDataBlockInfo* pBlockInfo) {
   // 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->groupbyNormalCol && !isFixedOutputQuery(pQuery)) {
+  if (!QUERY_IS_INTERVAL_QUERY(pQuery) && !pRuntimeEnv->groupbyNormalCol && !isFixedOutputQuery(pRuntimeEnv)) {
     SResultRec *pRec = &pQuery->rec;
     
     if (pQuery->rec.capacity - pQuery->rec.rows < pBlockInfo->rows) {
@@ -2249,7 +2225,7 @@ static int64_t doScanAllDataBlocks(SQueryRuntimeEnv *pRuntimeEnv) {
   SDataBlockInfo blockInfo = SDATA_BLOCK_INITIALIZER;
   while (tsdbNextDataBlock(pQueryHandle)) {
     summary->totalBlocks += 1;
-    if (isQueryKilled(GET_QINFO_ADDR(pRuntimeEnv))) {
+    if (IS_QUERY_KILLED(GET_QINFO_ADDR(pRuntimeEnv))) {
       return 0;
     }
 
@@ -3304,7 +3280,7 @@ void scanOneTableDataBlocks(SQueryRuntimeEnv *pRuntimeEnv, TSKEY start) {
         cond.twindow.skey, cond.twindow.ekey);
 
     // check if query is killed or not
-    if (isQueryKilled(pQInfo)) {
+    if (IS_QUERY_KILLED(pQInfo)) {
       return;
     }
   }
@@ -3695,23 +3671,24 @@ void copyFromWindowResToSData(SQInfo *pQInfo, SWindowResInfo *pResultInfo) {
   assert(pQuery->rec.rows <= pQuery->rec.capacity);
 }
 
-static void updateWindowResNumOfRes(SQueryRuntimeEnv *pRuntimeEnv, STableQueryInfo *pTableQueryInfo) {
+static void updateWindowResNumOfRes(SQueryRuntimeEnv *pRuntimeEnv) {
   SQuery *pQuery = pRuntimeEnv->pQuery;
 
   // update the number of result for each, only update the number of rows for the corresponding window result.
-  if (!QUERY_IS_INTERVAL_QUERY(pQuery)) {
-
-    for (int32_t i = 0; i < pRuntimeEnv->windowResInfo.size; ++i) {
-      SWindowResult *pResult = &pRuntimeEnv->windowResInfo.pResult[i];
+  if (QUERY_IS_INTERVAL_QUERY(pQuery)) {
+    return;
+  }
 
-      for (int32_t j = 0; j < pQuery->numOfOutput; ++j) {
-        int32_t functionId = pRuntimeEnv->pCtx[j].functionId;
-        if (functionId == TSDB_FUNC_TS || functionId == TSDB_FUNC_TAG || functionId == TSDB_FUNC_TAGPRJ) {
-          continue;
-        }
+  for (int32_t i = 0; i < pRuntimeEnv->windowResInfo.size; ++i) {
+    SWindowResult *pResult = &pRuntimeEnv->windowResInfo.pResult[i];
 
-        pResult->numOfRows = MAX(pResult->numOfRows, pResult->resultInfo[j].numOfRes);
+    for (int32_t j = 0; j < pQuery->numOfOutput; ++j) {
+      int32_t functionId = pRuntimeEnv->pCtx[j].functionId;
+      if (functionId == TSDB_FUNC_TS || functionId == TSDB_FUNC_TAG || functionId == TSDB_FUNC_TAGPRJ) {
+        continue;
       }
+
+      pResult->numOfRows = MAX(pResult->numOfRows, pResult->resultInfo[j].numOfRes);
     }
   }
 }
@@ -3729,8 +3706,6 @@ void stableApplyFunctionsOnBlock(SQueryRuntimeEnv *pRuntimeEnv, SDataBlockInfo *
   } else {
     blockwiseApplyFunctions(pRuntimeEnv, pStatis, pDataBlockInfo, pWindowResInfo, searchFn, pDataBlock);
   }
-
-  updateWindowResNumOfRes(pRuntimeEnv, pTableQueryInfo);
 }
 
 bool queryHasRemainResults(SQueryRuntimeEnv* pRuntimeEnv) {
@@ -3950,7 +3925,7 @@ void skipBlocks(SQueryRuntimeEnv *pRuntimeEnv) {
 
   SDataBlockInfo blockInfo = SDATA_BLOCK_INITIALIZER;
   while (tsdbNextDataBlock(pQueryHandle)) {
-    if (isQueryKilled(GET_QINFO_ADDR(pRuntimeEnv))) {
+    if (IS_QUERY_KILLED(GET_QINFO_ADDR(pRuntimeEnv))) {
       return;
     }
 
@@ -4099,7 +4074,7 @@ static void setupQueryHandle(void* tsdb, SQInfo* pQInfo, bool isSTableQuery) {
     return;
   }
 
-  if (isSTableQuery && (!QUERY_IS_INTERVAL_QUERY(pQuery)) && (!isFixedOutputQuery(pQuery))) {
+  if (isSTableQuery && (!QUERY_IS_INTERVAL_QUERY(pQuery)) && (!isFixedOutputQuery(pRuntimeEnv))) {
     return;
   }
 
@@ -4115,7 +4090,7 @@ static void setupQueryHandle(void* tsdb, SQInfo* pQInfo, bool isSTableQuery) {
     && (cond.order == TSDB_ORDER_ASC) 
     && (!QUERY_IS_INTERVAL_QUERY(pQuery))
     && (!isGroupbyNormalCol(pQuery->pGroupbyExpr))
-    && (!isFixedOutputQuery(pQuery))
+    && (!isFixedOutputQuery(pRuntimeEnv))
   ) {
     SArray* pa = GET_TABLEGROUP(pQInfo, 0);
     STableQueryInfo* pCheckInfo = taosArrayGetP(pa, 0);
@@ -4267,7 +4242,7 @@ static int64_t scanMultiTableDataBlocks(SQInfo *pQInfo) {
 
   while (tsdbNextDataBlock(pQueryHandle)) {
     summary->totalBlocks += 1;
-    if (isQueryKilled(pQInfo)) {
+    if (IS_QUERY_KILLED(pQInfo)) {
       break;
     }
 
@@ -4304,6 +4279,8 @@ static int64_t scanMultiTableDataBlocks(SQInfo *pQInfo) {
            pQInfo, blockInfo.uid, blockInfo.tid, blockInfo.window.skey, blockInfo.window.ekey, blockInfo.rows, pQuery->current->lastKey);
   }
 
+  updateWindowResNumOfRes(pRuntimeEnv);
+
   int64_t et = taosGetTimestampMs();
   return et - st;
 }
@@ -4316,7 +4293,9 @@ static bool multiTableMultioutputHelper(SQInfo *pQInfo, int32_t index) {
   SArray *group = GET_TABLEGROUP(pQInfo, 0);
   STableQueryInfo* pCheckInfo = taosArrayGetP(group, index);
 
-  setTagVal(pRuntimeEnv, pCheckInfo->pTable, pQInfo->tsdb);
+  if (pRuntimeEnv->hasTagResults || pRuntimeEnv->pTSBuf != NULL) {
+    setTagVal(pRuntimeEnv, pCheckInfo->pTable, pQInfo->tsdb);
+  }
 
   STableId* id = TSDB_TABLEID(pCheckInfo->pTable);
   qDebug("QInfo:%p query on (%d): uid:%" PRIu64 ", tid:%d, qrange:%" PRId64 "-%" PRId64, pQInfo, index,
@@ -4547,7 +4526,7 @@ static void sequentialTableProcess(SQInfo *pQInfo) {
            1 == taosArrayGetSize(pQInfo->tableqinfoGroupInfo.pGroupList));
 
     while (pQInfo->tableIndex < pQInfo->tableqinfoGroupInfo.numOfTables) {
-      if (isQueryKilled(pQInfo)) {
+      if (IS_QUERY_KILLED(pQInfo)) {
         return;
       }
 
@@ -4742,7 +4721,7 @@ static void multiTableQueryProcess(SQInfo *pQInfo) {
   qDebug("QInfo:%p master scan completed, elapsed time: %" PRId64 "ms, reverse scan start", pQInfo, el);
 
   // query error occurred or query is killed, abort current execution
-  if (pQInfo->code != TSDB_CODE_SUCCESS || isQueryKilled(pQInfo)) {
+  if (pQInfo->code != TSDB_CODE_SUCCESS || IS_QUERY_KILLED(pQInfo)) {
     qDebug("QInfo:%p query killed or error occurred, code:%s, abort", pQInfo, tstrerror(pQInfo->code));
     return;
   }
@@ -4764,7 +4743,7 @@ static void multiTableQueryProcess(SQInfo *pQInfo) {
 
   setQueryStatus(pQuery, QUERY_COMPLETED);
 
-  if (pQInfo->code != TSDB_CODE_SUCCESS || isQueryKilled(pQInfo)) {
+  if (pQInfo->code != TSDB_CODE_SUCCESS || IS_QUERY_KILLED(pQInfo)) {
     qDebug("QInfo:%p query killed or error occurred, code:%s, abort", pQInfo, tstrerror(pQInfo->code));
     return;
   }
@@ -4804,7 +4783,7 @@ static void tableFixedOutputProcess(SQInfo *pQInfo, STableQueryInfo* pTableInfo)
   scanOneTableDataBlocks(pRuntimeEnv, pTableInfo->lastKey);
   finalizeQueryResult(pRuntimeEnv);
 
-  if (isQueryKilled(pQInfo)) {
+  if (IS_QUERY_KILLED(pQInfo)) {
     return;
   }
 
@@ -4837,7 +4816,7 @@ static void tableMultiOutputProcess(SQInfo *pQInfo, STableQueryInfo* pTableInfo)
     scanOneTableDataBlocks(pRuntimeEnv, pQuery->current->lastKey);
     finalizeQueryResult(pRuntimeEnv);
 
-    if (isQueryKilled(pQInfo)) {
+    if (IS_QUERY_KILLED(pQInfo)) {
       return;
     }
 
@@ -4885,7 +4864,7 @@ static void tableIntervalProcessImpl(SQueryRuntimeEnv *pRuntimeEnv, TSKEY start)
   while (1) {
     scanOneTableDataBlocks(pRuntimeEnv, start);
 
-    if (isQueryKilled(GET_QINFO_ADDR(pRuntimeEnv))) {
+    if (IS_QUERY_KILLED(GET_QINFO_ADDR(pRuntimeEnv))) {
       return;
     }
 
@@ -5022,7 +5001,7 @@ static void tableQueryImpl(SQInfo *pQInfo) {
   // group by normal column, sliding window query, interval query are handled by interval query processor
   if (QUERY_IS_INTERVAL_QUERY(pQuery) || pRuntimeEnv->groupbyNormalCol) {  // interval (down sampling operation)
     tableIntervalProcess(pQInfo, item);
-  } else if (isFixedOutputQuery(pQuery)) {
+  } else if (isFixedOutputQuery(pRuntimeEnv)) {
     tableFixedOutputProcess(pQInfo, item);
   } else {  // diff/add/multiply/subtract/division
     assert(pQuery->checkBuffer == 1);
@@ -5042,7 +5021,7 @@ static void stableQueryImpl(SQInfo *pQInfo) {
   int64_t st = taosGetTimestampUs();
 
   if (QUERY_IS_INTERVAL_QUERY(pQuery) ||
-      (isFixedOutputQuery(pQuery) && (!isPointInterpoQuery(pQuery)) && !pRuntimeEnv->groupbyNormalCol &&
+      (isFixedOutputQuery(pRuntimeEnv) && (!isPointInterpoQuery(pQuery)) && !pRuntimeEnv->groupbyNormalCol &&
       !isFirstLastRowQuery(pQuery))) {
     multiTableQueryProcess(pQInfo);
   } else {
@@ -6173,7 +6152,7 @@ void qTableQuery(qinfo_t qinfo) {
     return;
   }
 
-  if (isQueryKilled(pQInfo)) {
+  if (IS_QUERY_KILLED(pQInfo)) {
     qDebug("QInfo:%p it is already killed, abort", pQInfo);
 
     sem_post(&pQInfo->dataReady);
@@ -6214,7 +6193,7 @@ void qTableQuery(qinfo_t qinfo) {
   }
 
   SQuery* pQuery = pRuntimeEnv->pQuery;
-  if (isQueryKilled(pQInfo)) {
+  if (IS_QUERY_KILLED(pQInfo)) {
     qDebug("QInfo:%p query is killed", pQInfo);
   } else if (pQuery->rec.rows == 0) {
     qDebug("QInfo:%p over, %zu tables queried, %"PRId64" rows are returned", pQInfo, pQInfo->tableqinfoGroupInfo.numOfTables, pQuery->rec.total);
@@ -6235,7 +6214,7 @@ int32_t qRetrieveQueryResultInfo(qinfo_t qinfo) {
   }
 
   SQuery *pQuery = pQInfo->runtimeEnv.pQuery;
-  if (isQueryKilled(pQInfo)) {
+  if (IS_QUERY_KILLED(pQInfo)) {
     qDebug("QInfo:%p query is killed, code:%d", pQInfo, pQInfo->code);
     return pQInfo->code;
   }
@@ -6310,7 +6289,7 @@ int32_t qDumpRetrieveResult(qinfo_t qinfo, SRetrieveTableRsp **pRsp, int32_t *co
     code = pQInfo->code;
   }
 
-  if (isQueryKilled(pQInfo) || Q_STATUS_EQUAL(pQuery->status, QUERY_OVER)) {
+  if (IS_QUERY_KILLED(pQInfo) || Q_STATUS_EQUAL(pQuery->status, QUERY_OVER)) {
     (*pRsp)->completed = 1;  // notify no more result to client
   }
 
diff --git a/src/tsdb/src/tsdbRead.c b/src/tsdb/src/tsdbRead.c
index a4e0151f8992753bf4726408c30bdf3910c3da75..c055a27c39fdcc9791f07bb37d4fde8577e545f0 100644
--- a/src/tsdb/src/tsdbRead.c
+++ b/src/tsdb/src/tsdbRead.c
@@ -110,9 +110,10 @@ typedef struct STsdbQueryHandle {
   SFileGroupIter fileIter;
   SRWHelper      rhelper;
   STableBlockInfo* pDataBlockInfo;
+  int32_t        allocSize;        // allocated data block size
   SMemTable*     mem;              // mem-table
   SMemTable*     imem;             // imem-table, acquired from snapshot
-  
+  SArray*        defaultLoadColumn;// default load column
   SDataBlockLoadInfo dataBlockLoadInfo; /* record current block load information */
   SLoadCompBlockInfo compBlockLoadInfo; /* record current compblock information in SQuery */
 } STsdbQueryHandle;
@@ -136,6 +137,34 @@ static void tsdbInitCompBlockLoadInfo(SLoadCompBlockInfo* pCompBlockLoadInfo) {
   pCompBlockLoadInfo->fileId = -1;
 }
 
+static SArray* getColumnIdList(STsdbQueryHandle* pQueryHandle) {
+  size_t numOfCols = QH_GET_NUM_OF_COLS(pQueryHandle);
+  assert(numOfCols <= TSDB_MAX_COLUMNS);
+
+  SArray* pIdList = taosArrayInit(numOfCols, sizeof(int16_t));
+  for (int32_t i = 0; i < numOfCols; ++i) {
+    SColumnInfoData* pCol = taosArrayGet(pQueryHandle->pColumns, i);
+    taosArrayPush(pIdList, &pCol->info.colId);
+  }
+
+  return pIdList;
+}
+
+static SArray* getDefaultLoadColumns(STsdbQueryHandle* pQueryHandle, bool loadTS) {
+  SArray* pLocalIdList = getColumnIdList(pQueryHandle);
+
+  // check if the primary time stamp column needs to load
+  int16_t colId = *(int16_t*)taosArrayGet(pLocalIdList, 0);
+
+  // the primary timestamp column does not be included in the the specified load column list, add it
+  if (loadTS && colId != 0) {
+    int16_t columnId = 0;
+    taosArrayInsert(pLocalIdList, 0, &columnId);
+  }
+
+  return pLocalIdList;
+}
+
 TsdbQueryHandleT* tsdbQueryTables(TSDB_REPO_T* tsdb, STsdbQueryCond* pCond, STableGroupInfo* groupList, void* qinfo) {
   STsdbQueryHandle* pQueryHandle = calloc(1, sizeof(STsdbQueryHandle));
   pQueryHandle->order       = pCond->order;
@@ -148,7 +177,8 @@ TsdbQueryHandleT* tsdbQueryTables(TSDB_REPO_T* tsdb, STsdbQueryCond* pCond, STab
   pQueryHandle->activeIndex = 0;   // current active table index
   pQueryHandle->qinfo       = qinfo;
   pQueryHandle->outputCapacity = ((STsdbRepo*)tsdb)->config.maxRowsPerFileBlock;
-  
+  pQueryHandle->allocSize   = 0;
+
   tsdbInitReadHelper(&pQueryHandle->rhelper, (STsdbRepo*) tsdb);
   tsdbTakeMemSnapshot(pQueryHandle->pTsdb, &pQueryHandle->mem, &pQueryHandle->imem);
 
@@ -195,7 +225,9 @@ TsdbQueryHandleT* tsdbQueryTables(TSDB_REPO_T* tsdb, STsdbQueryCond* pCond, STab
       taosArrayPush(pQueryHandle->pTableCheckInfo, &info);
     }
   }
-  
+
+  pQueryHandle->defaultLoadColumn = getDefaultLoadColumns(pQueryHandle, true);
+
   tsdbDebug("%p total numOfTable:%zu in query", pQueryHandle, taosArrayGetSize(pQueryHandle->pTableCheckInfo));
 
   tsdbInitDataBlockLoadInfo(&pQueryHandle->dataBlockLoadInfo);
@@ -546,33 +578,7 @@ static int32_t getFileCompInfo(STsdbQueryHandle* pQueryHandle, int32_t* numOfBlo
                     .tid = (_checkInfo)->tableId.tid,                                  \
                     .uid = (_checkInfo)->tableId.uid})
 
-static SArray* getColumnIdList(STsdbQueryHandle* pQueryHandle) {
-  size_t numOfCols = QH_GET_NUM_OF_COLS(pQueryHandle);
-  assert(numOfCols <= TSDB_MAX_COLUMNS);
-  
-  SArray* pIdList = taosArrayInit(numOfCols, sizeof(int16_t));
-  for (int32_t i = 0; i < numOfCols; ++i) {
-    SColumnInfoData* pCol = taosArrayGet(pQueryHandle->pColumns, i);
-    taosArrayPush(pIdList, &pCol->info.colId);
-  }
-  
-  return pIdList;
-}
 
-static SArray* getDefaultLoadColumns(STsdbQueryHandle* pQueryHandle, bool loadTS) {
-  SArray* pLocalIdList = getColumnIdList(pQueryHandle);
-  
-  // check if the primary time stamp column needs to load
-  int16_t colId = *(int16_t*)taosArrayGet(pLocalIdList, 0);
-  
-  // the primary timestamp column does not be included in the the specified load column list, add it
-  if (loadTS && colId != 0) {
-    int16_t columnId = 0;
-    taosArrayInsert(pLocalIdList, 0, &columnId);
-  }
-  
-  return pLocalIdList;
-}
 
 static bool doLoadFileDataBlock(STsdbQueryHandle* pQueryHandle, SCompBlock* pBlock, STableCheckInfo* pCheckInfo) {
   STsdbRepo *pRepo = pQueryHandle->pTsdb;
@@ -584,8 +590,6 @@ static bool doLoadFileDataBlock(STsdbQueryHandle* pQueryHandle, SCompBlock* pBlo
   data->uid = pCheckInfo->pTableObj->tableId.uid;
 
   bool    blockLoaded = false;
-  SArray* sa = getDefaultLoadColumns(pQueryHandle, true);
-
   int64_t st = taosGetTimestampUs();
 
   if (pCheckInfo->pDataCols == NULL) {
@@ -613,7 +617,6 @@ static bool doLoadFileDataBlock(STsdbQueryHandle* pQueryHandle, SCompBlock* pBlo
   assert(pCols->numOfRows != 0 && pCols->numOfRows <= pBlock->numOfRows);
 
   pBlock->numOfRows = pCols->numOfRows;
-  taosArrayDestroy(sa);
   tfree(data);
 
   int64_t et = taosGetTimestampUs() - st;
@@ -656,12 +659,8 @@ static void handleDataMergeIfNeeded(STsdbQueryHandle* pQueryHandle, SCompBlock*
       return;
     }
   
-    SArray* sa = getDefaultLoadColumns(pQueryHandle, true);
-  
     doLoadFileDataBlock(pQueryHandle, pBlock, pCheckInfo);
-    doMergeTwoLevelData(pQueryHandle, pCheckInfo, pBlock, sa);
-    taosArrayDestroy(sa);
-    
+    doMergeTwoLevelData(pQueryHandle, pCheckInfo, pBlock, pQueryHandle->defaultLoadColumn);
   } else {
     /*
      * no data in cache, only load data from file
@@ -681,14 +680,12 @@ static void handleDataMergeIfNeeded(STsdbQueryHandle* pQueryHandle, SCompBlock*
 }
 
 static bool loadFileDataBlock(STsdbQueryHandle* pQueryHandle, SCompBlock* pBlock, STableCheckInfo* pCheckInfo) {
-  SArray*        sa = getDefaultLoadColumns(pQueryHandle, true);
   SQueryFilePos* cur = &pQueryHandle->cur;
 
   if (ASCENDING_TRAVERSE(pQueryHandle->order)) {
     // query ended in current block
     if (pQueryHandle->window.ekey < pBlock->keyLast || pCheckInfo->lastKey > pBlock->keyFirst) {
       if (!doLoadFileDataBlock(pQueryHandle, pBlock, pCheckInfo)) {
-        taosArrayDestroy(sa);
         return false;
       }
 
@@ -702,7 +699,7 @@ static bool loadFileDataBlock(STsdbQueryHandle* pQueryHandle, SCompBlock* pBlock
         cur->pos = 0;
       }
       
-      doMergeTwoLevelData(pQueryHandle, pCheckInfo, pBlock, sa);
+      doMergeTwoLevelData(pQueryHandle, pCheckInfo, pBlock, pQueryHandle->defaultLoadColumn);
     } else {  // the whole block is loaded in to buffer
       handleDataMergeIfNeeded(pQueryHandle, pBlock, pCheckInfo);
     }
@@ -719,13 +716,12 @@ static bool loadFileDataBlock(STsdbQueryHandle* pQueryHandle, SCompBlock* pBlock
         cur->pos = pBlock->numOfRows - 1;
       }
       
-      doMergeTwoLevelData(pQueryHandle, pCheckInfo, pBlock, sa);
+      doMergeTwoLevelData(pQueryHandle, pCheckInfo, pBlock, pQueryHandle->defaultLoadColumn);
     } else {
       handleDataMergeIfNeeded(pQueryHandle, pBlock, pCheckInfo);
     }
   }
 
-  taosArrayDestroy(sa);
   return pQueryHandle->realNumOfRows > 0;
 }
 
@@ -1250,13 +1246,19 @@ static int32_t dataBlockOrderCompar(const void* pLeft, const void* pRight, void*
 }
 
 static int32_t createDataBlocksInfo(STsdbQueryHandle* pQueryHandle, int32_t numOfBlocks, int32_t* numOfAllocBlocks) {
-  char* tmp = realloc(pQueryHandle->pDataBlockInfo, sizeof(STableBlockInfo) * numOfBlocks);
-  if (tmp == NULL) {
-    return TSDB_CODE_TDB_OUT_OF_MEMORY;
+  size_t size = sizeof(STableBlockInfo) * numOfBlocks;
+
+  if (pQueryHandle->allocSize < size) {
+    pQueryHandle->allocSize = size;
+    char* tmp = realloc(pQueryHandle->pDataBlockInfo, pQueryHandle->allocSize);
+    if (tmp == NULL) {
+      return TSDB_CODE_TDB_OUT_OF_MEMORY;
+    }
+
+    pQueryHandle->pDataBlockInfo = (STableBlockInfo*) tmp;
   }
 
-  pQueryHandle->pDataBlockInfo = (STableBlockInfo*) tmp;
-  memset(pQueryHandle->pDataBlockInfo, 0, sizeof(STableBlockInfo) * numOfBlocks);
+  memset(pQueryHandle->pDataBlockInfo, 0, size);
   *numOfAllocBlocks = numOfBlocks;
 
   int32_t numOfTables = taosArrayGetSize(pQueryHandle->pTableCheckInfo);
@@ -1492,9 +1494,8 @@ bool tsdbNextDataBlock(TsdbQueryHandleT* pHandle) {
       return false;
     }
     
-    SArray* sa = getDefaultLoadColumns(pQueryHandle, true);
     /*SDataBlockInfo* pBlockInfo =*/ tsdbRetrieveDataBlockInfo(pHandle, &blockInfo);
-    /*SArray *pDataBlock = */tsdbRetrieveDataBlock(pHandle, sa);
+    /*SArray *pDataBlock = */tsdbRetrieveDataBlock(pHandle, pQueryHandle->defaultLoadColumn);
   
     if (pQueryHandle->cur.win.ekey == pQueryHandle->window.skey) {
       // data already retrieve, discard other data rows and return
@@ -1508,7 +1509,6 @@ bool tsdbNextDataBlock(TsdbQueryHandleT* pHandle) {
       pQueryHandle->window   = pQueryHandle->cur.win;
       pQueryHandle->cur.rows = 1;
       pQueryHandle->type = TSDB_QUERY_TYPE_EXTERNAL;
-      taosArrayDestroy(sa);
       return true;
     } else {
       STsdbQueryHandle* pSecQueryHandle = calloc(1, sizeof(STsdbQueryHandle));
@@ -1565,7 +1565,7 @@ bool tsdbNextDataBlock(TsdbQueryHandleT* pHandle) {
       assert(ret);
 
       /*SDataBlockInfo* pBlockInfo =*/ tsdbRetrieveDataBlockInfo((void*) pSecQueryHandle, &blockInfo);
-      /*SArray *pDataBlock = */tsdbRetrieveDataBlock((void*) pSecQueryHandle, sa);
+      /*SArray *pDataBlock = */tsdbRetrieveDataBlock((void*) pSecQueryHandle, pSecQueryHandle->defaultLoadColumn);
   
       for (int32_t i = 0; i < numOfCols; ++i) {
         SColumnInfoData* pCol = taosArrayGet(pQueryHandle->pColumns, i);
@@ -2333,6 +2333,7 @@ void tsdbCleanupQueryHandle(TsdbQueryHandleT queryHandle) {
    }
 
   taosArrayDestroy(pQueryHandle->pColumns);
+  taosArrayDestroy(pQueryHandle->defaultLoadColumn);
   tfree(pQueryHandle->pDataBlockInfo);
   tfree(pQueryHandle->statis);
 
diff --git a/src/util/inc/tarray.h b/src/util/inc/tarray.h
index 4d44e82b1b174bf06a97bf826a029f67602f2592..c05e2757d60b08cc12d1a470480f7daf94fa61d1 100644
--- a/src/util/inc/tarray.h
+++ b/src/util/inc/tarray.h
@@ -23,7 +23,7 @@ extern "C" {
 #include "os.h"
 
 #define TARRAY_MIN_SIZE 8
-#define TARRAY_GET_ELEM(array, index) ((array)->pData + (index) * (array)->elemSize)
+#define TARRAY_GET_ELEM(array, index) ((void*)((array)->pData + (index) * (array)->elemSize))
 
 typedef struct SArray {
   size_t size;