Merge pull request #6706 from taosdata/feature/TD-4666

Feature/td 4666

src/client/inc/tscUtil.h

@@ -94,11 +94,21 @@ typedef struct SVgroupTableInfo {
   SArray     *itemList;   // SArray<STableIdInfo>
 } SVgroupTableInfo;
 
+typedef struct SBlockKeyTuple {
+  TSKEY skey;
+  void* payloadAddr;
+} SBlockKeyTuple;
+
+typedef struct SBlockKeyInfo {
+  int32_t         maxBytesAlloc;
+  SBlockKeyTuple* pKeyTuple;
+} SBlockKeyInfo;
+
 int32_t converToStr(char *str, int type, void *buf, int32_t bufSize, int32_t *len);
 
 int32_t tscCreateDataBlock(size_t initialSize, int32_t rowSize, int32_t startOffset, SName* name, STableMeta* pTableMeta, STableDataBlocks** dataBlocks);
 void tscDestroyDataBlock(STableDataBlocks* pDataBlock, bool removeMeta);
-void tscSortRemoveDataBlockDupRows(STableDataBlocks* dataBuf);
+int     tscSortRemoveDataBlockDupRows(STableDataBlocks* dataBuf, SBlockKeyInfo* pBlkKeyInfo);
 
 void tscDestroyBoundColumnInfo(SParsedDataColInfo* pColInfo);
 void doRetrieveSubqueryData(SSchedMsg *pMsg);

src/client/inc/tsclient.h

@@ -88,13 +88,43 @@ typedef struct SBoundColumn {
   int32_t offset;  // all column offset value
 } SBoundColumn;
 
+typedef struct {
+  uint16_t schemaColIdx;
+  uint16_t boundIdx;
+  uint16_t finalIdx;
+} SBoundIdxInfo;
+
+typedef enum _COL_ORDER_STATUS {
+  ORDER_STATUS_UNKNOWN = 0,
+  ORDER_STATUS_ORDERED = 1,
+  ORDER_STATUS_DISORDERED = 2,
+} EOrderStatus;
+
 typedef struct SParsedDataColInfo {
-  int16_t         numOfCols;
+  int16_t        numOfCols;
-  int16_t         numOfBound;
+  int16_t        numOfBound;
-  int32_t        *boundedColumns;
+  int32_t *      boundedColumns;  // bounded column idx according to schema
-  SBoundColumn   *cols;
+  SBoundColumn * cols;
+  SBoundIdxInfo *colIdxInfo;
+  int8_t         orderStatus;  // bounded columns: 
 } SParsedDataColInfo;
 
+#define IS_DATA_COL_ORDERED(s) ((s) == (int8_t)ORDER_STATUS_ORDERED)
+
+typedef struct {
+  SSchema *   pSchema;
+  int16_t     sversion;
+  int32_t     flen;
+  uint16_t    nCols;
+  void *      buf;
+  void *      pDataBlock;
+  SSubmitBlk *pSubmitBlk;
+} SMemRowBuilder;
+
+typedef struct {
+  TDRowLenT allNullLen;
+} SMemRowHelper;
+
 typedef struct STableDataBlocks {
   SName       tableName;
   int8_t      tsSource;     // where does the UNIX timestamp come from, server or client
@@ -109,12 +139,13 @@ typedef struct STableDataBlocks {
   STableMeta *pTableMeta;   // the tableMeta of current table, the table meta will be used during submit, keep a ref to avoid to be removed from cache
   char       *pData;
 
-  SParsedDataColInfo  boundColumnInfo;
+  SParsedDataColInfo boundColumnInfo;
 
   // for parameter ('?') binding
-  uint32_t    numOfAllocedParams;
+  uint32_t       numOfAllocedParams;
-  uint32_t    numOfParams;
+  uint32_t       numOfParams;
-  SParamInfo *params;
+  SParamInfo *   params;
+  SMemRowHelper  rowHelper;
 } STableDataBlocks;
 
 typedef struct {
@@ -243,6 +274,7 @@ typedef struct SSqlObj {
   void *           pStream;
   void *           pSubscription;
   char *           sqlstr;
+  void *           pBuf;  // table meta buffer
   char             parseRetry;
   char             retry;
   char             maxRetry;
@@ -382,12 +414,17 @@ extern int   tscRefId;
 extern int   tscNumOfObj;     // number of existed sqlObj in current process.
 
 extern int (*tscBuildMsg[TSDB_SQL_MAX])(SSqlObj *pSql, SSqlInfo *pInfo);
-
+ 
 void tscBuildVgroupTableInfo(SSqlObj* pSql, STableMetaInfo* pTableMetaInfo, SArray* tables);
 int16_t getNewResColId(SSqlCmd* pCmd);
 
+int32_t schemaIdxCompar(const void *lhs, const void *rhs);
+int32_t boundIdxCompar(const void *lhs, const void *rhs);
+int     initSMemRowHelper(SMemRowHelper *pHelper, SSchema *pSSchema, uint16_t nCols, uint16_t allNullColsLen);
+int32_t getExtendedRowSize(STableComInfo *tinfo);
+
 #ifdef __cplusplus
 }
 #endif
 
-#endif
+#endif

src/client/src/tscParseInsert.c

@@ -38,10 +38,33 @@ enum {
   TSDB_USE_CLI_TS = 1,
 };
 
+static uint8_t TRUE_VALUE = (uint8_t)TSDB_TRUE;
+static uint8_t FALSE_VALUE = (uint8_t)TSDB_FALSE;
+
 static int32_t tscAllocateMemIfNeed(STableDataBlocks *pDataBlock, int32_t rowSize, int32_t *numOfRows);
 static int32_t parseBoundColumns(SInsertStatementParam *pInsertParam, SParsedDataColInfo *pColInfo, SSchema *pSchema,
                                  char *str, char **end);
 
+int32_t getExtendedRowSize(STableComInfo *tinfo) {
+  return tinfo->rowSize + PAYLOAD_HEADER_LEN + PAYLOAD_COL_HEAD_LEN * tinfo->numOfColumns;
+}
+int initSMemRowHelper(SMemRowHelper *pHelper, SSchema *pSSchema, uint16_t nCols, uint16_t allNullColsLen) {
+  pHelper->allNullLen = allNullColsLen;  //  TODO: get allNullColsLen when creating or altering table meta
+  if (pHelper->allNullLen == 0) {
+    for (uint16_t i = 0; i < nCols; ++i) {
+      uint8_t type = pSSchema[i].type;
+      int32_t typeLen = TYPE_BYTES[type];
+      pHelper->allNullLen += typeLen;
+      if (TSDB_DATA_TYPE_BINARY == type) {
+        pHelper->allNullLen += (VARSTR_HEADER_SIZE + CHAR_BYTES);
+      } else if (TSDB_DATA_TYPE_NCHAR == type) {
+        int len = VARSTR_HEADER_SIZE + TSDB_NCHAR_SIZE;
+        pHelper->allNullLen += len;
+      }
+    }
+  }
+  return 0;
+}
 static int32_t tscToDouble(SStrToken *pToken, double *value, char **endPtr) {
   errno = 0;
   *value = strtold(pToken->z, endPtr);
@@ -378,6 +401,342 @@ int32_t tsParseOneColumn(SSchema *pSchema, SStrToken *pToken, char *payload, cha
   return TSDB_CODE_SUCCESS;
 }
 
+static FORCE_INLINE TDRowLenT tsSetPayloadColValue(char *payloadStart, char *payload, int16_t columnId,
+                                                   uint8_t columnType, const void *value, uint16_t valueLen, TDRowTLenT tOffset) {
+  payloadColSetId(payload, columnId);
+  payloadColSetType(payload, columnType);
+  memcpy(POINTER_SHIFT(payloadStart,tOffset), value, valueLen);
+  return valueLen;
+}
+
+static int32_t tsParseOneColumnKV(SSchema *pSchema, SStrToken *pToken, char *payloadStart, char *primaryKeyStart,
+                                  char *payload, char *msg, char **str, bool primaryKey, int16_t timePrec,
+                                  TDRowTLenT tOffset, TDRowLenT *sizeAppend, TDRowLenT *dataRowColDeltaLen,
+                                  TDRowLenT *kvRowColLen) {
+  int64_t iv;
+  int32_t ret;
+  char *  endptr = NULL;
+
+  if (IS_NUMERIC_TYPE(pSchema->type) && pToken->n == 0) {
+    return tscInvalidOperationMsg(msg, "invalid numeric data", pToken->z);
+  }
+
+  switch (pSchema->type) {
+    case TSDB_DATA_TYPE_BOOL: {  // bool
+      if (isNullStr(pToken)) {
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                           tdGetNullVal(TSDB_DATA_TYPE_BOOL), TYPE_BYTES[TSDB_DATA_TYPE_BOOL], tOffset);
+      } else {
+        if ((pToken->type == TK_BOOL || pToken->type == TK_STRING) && (pToken->n != 0)) {
+          if (strncmp(pToken->z, "true", pToken->n) == 0) {
+            *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &TRUE_VALUE,
+                                               TYPE_BYTES[TSDB_DATA_TYPE_BOOL], tOffset);
+            *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_BOOL]);
+          } else if (strncmp(pToken->z, "false", pToken->n) == 0) {
+            *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &FALSE_VALUE,
+                                               TYPE_BYTES[TSDB_DATA_TYPE_BOOL], tOffset);
+            *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_BOOL]);
+          } else {
+            return tscSQLSyntaxErrMsg(msg, "invalid bool data", pToken->z);
+          }
+        } else if (pToken->type == TK_INTEGER) {
+          iv = strtoll(pToken->z, NULL, 10);
+          *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                             ((iv == 0) ? &FALSE_VALUE : &TRUE_VALUE), TYPE_BYTES[TSDB_DATA_TYPE_BOOL], tOffset);
+          *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_BOOL]);
+        } else if (pToken->type == TK_FLOAT) {
+          double dv = strtod(pToken->z, NULL);
+          *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                             ((dv == 0) ? &FALSE_VALUE : &TRUE_VALUE), TYPE_BYTES[TSDB_DATA_TYPE_BOOL], tOffset);
+          *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_BOOL]);
+        } else {
+          return tscInvalidOperationMsg(msg, "invalid bool data", pToken->z);
+        }
+      }
+      break;
+    }
+
+    case TSDB_DATA_TYPE_TINYINT:
+      if (isNullStr(pToken)) {
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                           tdGetNullVal(TSDB_DATA_TYPE_TINYINT), TYPE_BYTES[TSDB_DATA_TYPE_TINYINT], tOffset);
+      } else {
+        ret = tStrToInteger(pToken->z, pToken->type, pToken->n, &iv, true);
+        if (ret != TSDB_CODE_SUCCESS) {
+          return tscInvalidOperationMsg(msg, "invalid tinyint data", pToken->z);
+        } else if (!IS_VALID_TINYINT(iv)) {
+          return tscInvalidOperationMsg(msg, "data overflow", pToken->z);
+        }
+
+        uint8_t tmpVal = (uint8_t)iv;
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &tmpVal,
+                                           TYPE_BYTES[TSDB_DATA_TYPE_TINYINT], tOffset);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_TINYINT]);
+      }
+
+      break;
+
+    case TSDB_DATA_TYPE_UTINYINT:
+      if (isNullStr(pToken)) {
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                           tdGetNullVal(TSDB_DATA_TYPE_UTINYINT), TYPE_BYTES[TSDB_DATA_TYPE_UTINYINT], tOffset);
+      } else {
+        ret = tStrToInteger(pToken->z, pToken->type, pToken->n, &iv, false);
+        if (ret != TSDB_CODE_SUCCESS) {
+          return tscInvalidOperationMsg(msg, "invalid unsigned tinyint data", pToken->z);
+        } else if (!IS_VALID_UTINYINT(iv)) {
+          return tscInvalidOperationMsg(msg, "unsigned tinyint data overflow", pToken->z);
+        }
+
+        uint8_t tmpVal = (uint8_t)iv;
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &tmpVal,
+                                           TYPE_BYTES[TSDB_DATA_TYPE_UTINYINT], tOffset);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_UTINYINT]);
+      }
+
+      break;
+
+    case TSDB_DATA_TYPE_SMALLINT:
+      if (isNullStr(pToken)) {
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                           tdGetNullVal(TSDB_DATA_TYPE_SMALLINT), TYPE_BYTES[TSDB_DATA_TYPE_SMALLINT], tOffset);
+      } else {
+        ret = tStrToInteger(pToken->z, pToken->type, pToken->n, &iv, true);
+        if (ret != TSDB_CODE_SUCCESS) {
+          return tscInvalidOperationMsg(msg, "invalid smallint data", pToken->z);
+        } else if (!IS_VALID_SMALLINT(iv)) {
+          return tscInvalidOperationMsg(msg, "smallint data overflow", pToken->z);
+        }
+
+        int16_t tmpVal = (int16_t)iv;
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &tmpVal,
+                                           TYPE_BYTES[TSDB_DATA_TYPE_SMALLINT], tOffset);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_SMALLINT]);
+      }
+
+      break;
+
+    case TSDB_DATA_TYPE_USMALLINT:
+      if (isNullStr(pToken)) {
+        *sizeAppend =
+            tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                 tdGetNullVal(TSDB_DATA_TYPE_USMALLINT), TYPE_BYTES[TSDB_DATA_TYPE_USMALLINT], tOffset);
+      } else {
+        ret = tStrToInteger(pToken->z, pToken->type, pToken->n, &iv, false);
+        if (ret != TSDB_CODE_SUCCESS) {
+          return tscInvalidOperationMsg(msg, "invalid unsigned smallint data", pToken->z);
+        } else if (!IS_VALID_USMALLINT(iv)) {
+          return tscInvalidOperationMsg(msg, "unsigned smallint data overflow", pToken->z);
+        }
+
+        uint16_t tmpVal = (uint16_t)iv;
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &tmpVal,
+                                           TYPE_BYTES[TSDB_DATA_TYPE_USMALLINT], tOffset);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_USMALLINT]);
+      }
+
+      break;
+
+    case TSDB_DATA_TYPE_INT:
+      if (isNullStr(pToken)) {
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                           tdGetNullVal(TSDB_DATA_TYPE_INT), TYPE_BYTES[TSDB_DATA_TYPE_INT], tOffset);
+      } else {
+        ret = tStrToInteger(pToken->z, pToken->type, pToken->n, &iv, true);
+        if (ret != TSDB_CODE_SUCCESS) {
+          return tscInvalidOperationMsg(msg, "invalid int data", pToken->z);
+        } else if (!IS_VALID_INT(iv)) {
+          return tscInvalidOperationMsg(msg, "int data overflow", pToken->z);
+        }
+
+        int32_t tmpVal = (int32_t)iv;
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &tmpVal,
+                                           TYPE_BYTES[TSDB_DATA_TYPE_INT], tOffset);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_INT]);
+      }
+
+      break;
+
+    case TSDB_DATA_TYPE_UINT:
+      if (isNullStr(pToken)) {
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                           tdGetNullVal(TSDB_DATA_TYPE_UINT), TYPE_BYTES[TSDB_DATA_TYPE_UINT], tOffset);
+      } else {
+        ret = tStrToInteger(pToken->z, pToken->type, pToken->n, &iv, false);
+        if (ret != TSDB_CODE_SUCCESS) {
+          return tscInvalidOperationMsg(msg, "invalid unsigned int data", pToken->z);
+        } else if (!IS_VALID_UINT(iv)) {
+          return tscInvalidOperationMsg(msg, "unsigned int data overflow", pToken->z);
+        }
+
+        uint32_t tmpVal = (uint32_t)iv;
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &tmpVal,
+                                           TYPE_BYTES[TSDB_DATA_TYPE_UINT], tOffset);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_UINT]);
+      }
+
+      break;
+
+    case TSDB_DATA_TYPE_BIGINT:
+      if (isNullStr(pToken)) {
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                           tdGetNullVal(TSDB_DATA_TYPE_BIGINT), TYPE_BYTES[TSDB_DATA_TYPE_BIGINT], tOffset);
+      } else {
+        ret = tStrToInteger(pToken->z, pToken->type, pToken->n, &iv, true);
+        if (ret != TSDB_CODE_SUCCESS) {
+          return tscInvalidOperationMsg(msg, "invalid bigint data", pToken->z);
+        } else if (!IS_VALID_BIGINT(iv)) {
+          return tscInvalidOperationMsg(msg, "bigint data overflow", pToken->z);
+        }
+
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &iv,
+                                           TYPE_BYTES[TSDB_DATA_TYPE_BIGINT], tOffset);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_BIGINT]);
+      }
+      break;
+
+    case TSDB_DATA_TYPE_UBIGINT:
+      if (isNullStr(pToken)) {
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                           tdGetNullVal(TSDB_DATA_TYPE_UBIGINT), TYPE_BYTES[TSDB_DATA_TYPE_UBIGINT], tOffset);
+      } else {
+        ret = tStrToInteger(pToken->z, pToken->type, pToken->n, &iv, false);
+        if (ret != TSDB_CODE_SUCCESS) {
+          return tscInvalidOperationMsg(msg, "invalid unsigned bigint data", pToken->z);
+        } else if (!IS_VALID_UBIGINT((uint64_t)iv)) {
+          return tscInvalidOperationMsg(msg, "unsigned bigint data overflow", pToken->z);
+        }
+
+        uint64_t tmpVal = (uint64_t)iv;
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &tmpVal,
+                                           TYPE_BYTES[TSDB_DATA_TYPE_UBIGINT], tOffset);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_UBIGINT]);
+      }
+      break;
+
+    case TSDB_DATA_TYPE_FLOAT:
+      if (isNullStr(pToken)) {
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                           getNullValue(TSDB_DATA_TYPE_FLOAT), TYPE_BYTES[TSDB_DATA_TYPE_FLOAT], tOffset);
+      } else {
+        double dv;
+        if (TK_ILLEGAL == tscToDouble(pToken, &dv, &endptr)) {
+          return tscInvalidOperationMsg(msg, "illegal float data", pToken->z);
+        }
+
+        if (((dv == HUGE_VAL || dv == -HUGE_VAL) && errno == ERANGE) || dv > FLT_MAX || dv < -FLT_MAX || isinf(dv) ||
+            isnan(dv)) {
+          return tscInvalidOperationMsg(msg, "illegal float data", pToken->z);
+        }
+
+        float tmpVal = (float)dv;
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &tmpVal,
+                                           TYPE_BYTES[TSDB_DATA_TYPE_FLOAT], tOffset);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_FLOAT]);
+      }
+      break;
+
+    case TSDB_DATA_TYPE_DOUBLE:
+      if (isNullStr(pToken)) {
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                           tdGetNullVal(TSDB_DATA_TYPE_DOUBLE), TYPE_BYTES[TSDB_DATA_TYPE_DOUBLE], tOffset);
+      } else {
+        double dv;
+        if (TK_ILLEGAL == tscToDouble(pToken, &dv, &endptr)) {
+          return tscInvalidOperationMsg(msg, "illegal double data", pToken->z);
+        }
+
+        if (((dv == HUGE_VAL || dv == -HUGE_VAL) && errno == ERANGE) || isinf(dv) || isnan(dv)) {
+          return tscInvalidOperationMsg(msg, "illegal double data", pToken->z);
+        }
+
+        *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type, &dv,
+                                           TYPE_BYTES[TSDB_DATA_TYPE_DOUBLE], tOffset);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_DOUBLE]);
+      }
+      break;
+
+    case TSDB_DATA_TYPE_BINARY:
+      // binary data cannot be null-terminated char string, otherwise the last char of the string is lost
+      if (pToken->type == TK_NULL) {
+        payloadColSetId(payload, pSchema->colId);
+        payloadColSetType(payload, pSchema->type);
+        memcpy(POINTER_SHIFT(payloadStart, tOffset), tdGetNullVal(TSDB_DATA_TYPE_BINARY), VARSTR_HEADER_SIZE + CHAR_BYTES);
+        *sizeAppend = (TDRowLenT)(VARSTR_HEADER_SIZE + CHAR_BYTES);
+      } else {  // too long values will return invalid sql, not be truncated automatically
+        if (pToken->n + VARSTR_HEADER_SIZE > pSchema->bytes) {  // todo refactor
+          return tscInvalidOperationMsg(msg, "string data overflow", pToken->z);
+        }
+        // STR_WITH_SIZE_TO_VARSTR(payload, pToken->z, pToken->n);
+
+        payloadColSetId(payload, pSchema->colId);
+        payloadColSetType(payload, pSchema->type);
+        varDataSetLen(POINTER_SHIFT(payloadStart,tOffset), pToken->n);
+        memcpy(varDataVal(POINTER_SHIFT(payloadStart,tOffset)), pToken->z, pToken->n);
+        *sizeAppend = (TDRowLenT)(VARSTR_HEADER_SIZE + pToken->n);
+        *dataRowColDeltaLen += (TDRowLenT)(pToken->n - CHAR_BYTES);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + VARSTR_HEADER_SIZE + pToken->n);
+      }
+
+      break;
+
+    case TSDB_DATA_TYPE_NCHAR:
+      if (pToken->type == TK_NULL) {
+        payloadColSetId(payload, pSchema->colId);
+        payloadColSetType(payload, pSchema->type);
+        memcpy(POINTER_SHIFT(payloadStart,tOffset), tdGetNullVal(TSDB_DATA_TYPE_NCHAR), VARSTR_HEADER_SIZE + TSDB_NCHAR_SIZE);
+        *sizeAppend = (TDRowLenT)(VARSTR_HEADER_SIZE + TSDB_NCHAR_SIZE);
+      } else {
+        // if the converted output len is over than pColumnModel->bytes, return error: 'Argument list too long'
+        int32_t output = 0;
+        payloadColSetId(payload, pSchema->colId);
+        payloadColSetType(payload, pSchema->type);
+        if (!taosMbsToUcs4(pToken->z, pToken->n, varDataVal(POINTER_SHIFT(payloadStart,tOffset)),
+                           pSchema->bytes - VARSTR_HEADER_SIZE, &output)) {
+          char buf[512] = {0};
+          snprintf(buf, tListLen(buf), "%s", strerror(errno));
+          return tscInvalidOperationMsg(msg, buf, pToken->z);
+        }
+
+        varDataSetLen(POINTER_SHIFT(payloadStart,tOffset), output);
+
+        *sizeAppend = (TDRowLenT)(VARSTR_HEADER_SIZE + output);
+        *dataRowColDeltaLen += (TDRowLenT)(output - sizeof(uint32_t));
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + VARSTR_HEADER_SIZE + output);
+      }
+      break;
+
+    case TSDB_DATA_TYPE_TIMESTAMP: {
+      if (pToken->type == TK_NULL) {
+        if (primaryKey) {
+          // When building SKVRow primaryKey, we should not skip even with NULL value.
+          int64_t tmpVal = 0;
+          *sizeAppend = tsSetPayloadColValue(payloadStart, primaryKeyStart, pSchema->colId, pSchema->type, &tmpVal,
+                                             TYPE_BYTES[TSDB_DATA_TYPE_TIMESTAMP], tOffset);
+          *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_TIMESTAMP]);
+        } else {
+          *sizeAppend = tsSetPayloadColValue(payloadStart, payload, pSchema->colId, pSchema->type,
+                                             tdGetNullVal(TSDB_DATA_TYPE_TIMESTAMP),
+                                             TYPE_BYTES[TSDB_DATA_TYPE_TIMESTAMP], tOffset);
+        }
+      } else {
+        int64_t tmpVal;
+        if (tsParseTime(pToken, &tmpVal, str, msg, timePrec) != TSDB_CODE_SUCCESS) {
+          return tscInvalidOperationMsg(msg, "invalid timestamp", pToken->z);
+        }
+
+        *sizeAppend = tsSetPayloadColValue(payloadStart, primaryKey ? primaryKeyStart : payload, pSchema->colId,
+                                           pSchema->type, &tmpVal, TYPE_BYTES[TSDB_DATA_TYPE_TIMESTAMP], tOffset);
+        *kvRowColLen += (TDRowLenT)(sizeof(SColIdx) + TYPE_BYTES[TSDB_DATA_TYPE_TIMESTAMP]);
+      }
+
+      break;
+    }
+  }
+
+  return TSDB_CODE_SUCCESS;
+}
+
 /*
  * The server time/client time should not be mixed up in one sql string
  * Do not employ sort operation is not involved if server time is used.
@@ -414,24 +773,38 @@ int32_t tsCheckTimestamp(STableDataBlocks *pDataBlocks, const char *start) {
   return TSDB_CODE_SUCCESS;
 }
 
-int tsParseOneRow(char **str, STableDataBlocks *pDataBlocks, int16_t timePrec, int32_t *len,
+int tsParseOneRow(char **str, STableDataBlocks *pDataBlocks, int16_t timePrec, int32_t *len, char *tmpTokenBuf,
-                  char *tmpTokenBuf, SInsertStatementParam* pInsertParam) {
+                  SInsertStatementParam *pInsertParam) {
-  int32_t    index = 0;
+  int32_t   index = 0;
-  SStrToken  sToken = {0};
+  SStrToken sToken = {0};
-  char      *payload = pDataBlocks->pData + pDataBlocks->size;
+
+  SMemRowHelper *pHelper = &pDataBlocks->rowHelper;
+  char *         payload = pDataBlocks->pData + pDataBlocks->size;
 
   SParsedDataColInfo *spd = &pDataBlocks->boundColumnInfo;
-  SSchema *schema = tscGetTableSchema(pDataBlocks->pTableMeta);
+  SSchema *           schema = tscGetTableSchema(pDataBlocks->pTableMeta);
+
+  TDRowTLenT dataRowLen = pHelper->allNullLen;
+  TDRowTLenT kvRowLen = TD_MEM_ROW_KV_VER_SIZE;
+  TDRowTLenT payloadValOffset = 0;
+  TDRowLenT  colValOffset = 0;
+  ASSERT(dataRowLen > 0);
+
+  payloadSetNCols(payload, spd->numOfBound);
+  payloadValOffset = payloadValuesOffset(payload); // rely on payloadNCols
+  // payloadSetTLen(payload, payloadValOffset);
+
+  char *kvPrimaryKeyStart = payload + PAYLOAD_HEADER_LEN;    // primaryKey in 1st column tuple
+  char *kvStart = kvPrimaryKeyStart + PAYLOAD_COL_HEAD_LEN;  // the column tuple behind the primaryKey
 
   // 1. set the parsed value from sql string
-  int32_t rowSize = 0;
   for (int i = 0; i < spd->numOfBound; ++i) {
     // the start position in data block buffer of current value in sql
     int32_t colIndex = spd->boundedColumns[i];
 
-    char    *start = payload + spd->cols[colIndex].offset;
+    char *start = payload + spd->cols[colIndex].offset;
-    SSchema *pSchema = &schema[colIndex];
+
-    rowSize += pSchema->bytes;
+    SSchema *pSchema = &schema[colIndex];  // get colId here
 
     index = 0;
     sToken = tStrGetToken(*str, &index, true);
@@ -453,7 +826,8 @@ int tsParseOneRow(char **str, STableDataBlocks *pDataBlocks, int16_t timePrec, i
 
     int16_t type = sToken.type;
     if ((type != TK_NOW && type != TK_INTEGER && type != TK_STRING && type != TK_FLOAT && type != TK_BOOL &&
-         type != TK_NULL && type != TK_HEX && type != TK_OCT && type != TK_BIN) || (sToken.n == 0) || (type == TK_RP)) {
+         type != TK_NULL && type != TK_HEX && type != TK_OCT && type != TK_BIN) ||
+        (sToken.n == 0) || (type == TK_RP)) {
       return tscSQLSyntaxErrMsg(pInsertParam->msg, "invalid data or symbol", sToken.z);
     }
 
@@ -467,10 +841,10 @@ int tsParseOneRow(char **str, STableDataBlocks *pDataBlocks, int16_t timePrec, i
       if (sToken.n >= TSDB_MAX_BYTES_PER_ROW) {
         return tscSQLSyntaxErrMsg(pInsertParam->msg, "too long string", sToken.z);
       }
-      
+
       for (uint32_t k = 1; k < sToken.n - 1; ++k) {
         if (sToken.z[k] == '\\' || (sToken.z[k] == delim && sToken.z[k + 1] == delim)) {
-            tmpTokenBuf[j] = sToken.z[k + 1];
+          tmpTokenBuf[j] = sToken.z[k + 1];
 
           cnt++;
           j++;
@@ -487,42 +861,54 @@ int tsParseOneRow(char **str, STableDataBlocks *pDataBlocks, int16_t timePrec, i
       sToken.n -= 2 + cnt;
     }
 
-    bool    isPrimaryKey = (colIndex == PRIMARYKEY_TIMESTAMP_COL_INDEX);
+    bool      isPrimaryKey = (colIndex == PRIMARYKEY_TIMESTAMP_COL_INDEX);
-    int32_t ret = tsParseOneColumn(pSchema, &sToken, start, pInsertParam->msg, str, isPrimaryKey, timePrec);
+    TDRowLenT dataRowDeltaColLen = 0;  // When combine the data as SDataRow, the delta len between all NULL columns.
+    TDRowLenT kvRowColLen = 0;
+    TDRowLenT colValAppended = 0;
+
+    if (!IS_DATA_COL_ORDERED(spd->orderStatus)) {
+      ASSERT(spd->colIdxInfo != NULL);
+      if(!isPrimaryKey) {
+        kvStart = POINTER_SHIFT(kvPrimaryKeyStart, spd->colIdxInfo[i].finalIdx * PAYLOAD_COL_HEAD_LEN);
+      } else {
+        ASSERT(spd->colIdxInfo[i].finalIdx == 0);
+      }
+    }
+    // the primary key locates in 1st column
+    int32_t ret = tsParseOneColumnKV(pSchema, &sToken, payload, kvPrimaryKeyStart, kvStart, pInsertParam->msg, str,
+                                     isPrimaryKey, timePrec, payloadValOffset + colValOffset, &colValAppended,
+                                     &dataRowDeltaColLen, &kvRowColLen);
     if (ret != TSDB_CODE_SUCCESS) {
       return ret;
     }
 
-    if (isPrimaryKey && tsCheckTimestamp(pDataBlocks, start) != TSDB_CODE_SUCCESS) {
+    if (isPrimaryKey) {
-      tscInvalidOperationMsg(pInsertParam->msg, "client time/server time can not be mixed up", sToken.z);
+      if (tsCheckTimestamp(pDataBlocks, payloadValues(payload)) != TSDB_CODE_SUCCESS) {
-      return TSDB_CODE_TSC_INVALID_TIME_STAMP;
+        tscInvalidOperationMsg(pInsertParam->msg, "client time/server time can not be mixed up", sToken.z);
-    }
+        return TSDB_CODE_TSC_INVALID_TIME_STAMP;
-  }
+      }
-
+      payloadColSetOffset(kvPrimaryKeyStart, colValOffset);
-  // 2. set the null value for the columns that do not assign values
+    } else {
-  if (spd->numOfBound < spd->numOfCols) {
+      payloadColSetOffset(kvStart, colValOffset);
-    char *ptr = payload;
+      if (IS_DATA_COL_ORDERED(spd->orderStatus)) {
-
+        kvStart += PAYLOAD_COL_HEAD_LEN;  // move to next column
-    for (int32_t i = 0; i < spd->numOfCols; ++i) {
-      if (!spd->cols[i].hasVal) {  // current column do not have any value to insert, set it to null
-        if (schema[i].type == TSDB_DATA_TYPE_BINARY) {
-          varDataSetLen(ptr, sizeof(int8_t));
-          *(uint8_t*) varDataVal(ptr) = TSDB_DATA_BINARY_NULL;
-        } else if (schema[i].type == TSDB_DATA_TYPE_NCHAR) {
-          varDataSetLen(ptr, sizeof(int32_t));
-          *(uint32_t*) varDataVal(ptr) = TSDB_DATA_NCHAR_NULL;
-        } else {
-          setNull(ptr, schema[i].type, schema[i].bytes);
-        }
       }
-      
-      ptr += schema[i].bytes;
     }
-
+    
-    rowSize = (int32_t)(ptr - payload);
+    colValOffset += colValAppended;
+    kvRowLen += kvRowColLen;
+    dataRowLen += dataRowDeltaColLen;
   }
 
-  *len = rowSize;
+  if (kvRowLen < dataRowLen) {
+    payloadSetType(payload, SMEM_ROW_KV);
+  } else {
+    payloadSetType(payload, SMEM_ROW_DATA);
+  }
+
+  *len = (int32_t)(payloadValOffset + colValOffset);
+  payloadSetTLen(payload, *len);
+  
   return TSDB_CODE_SUCCESS;
 }
 
@@ -536,6 +922,27 @@ static int32_t rowDataCompar(const void *lhs, const void *rhs) {
     return left > right ? 1 : -1;
   }
 }
+int32_t schemaIdxCompar(const void *lhs, const void *rhs) {
+  uint16_t left = *(uint16_t *)lhs;
+  uint16_t right = *(uint16_t *)rhs;
+
+  if (left == right) {
+    return 0;
+  } else {
+    return left > right ? 1 : -1;
+  }
+}
+
+int32_t boundIdxCompar(const void *lhs, const void *rhs) {
+  uint16_t left = *(uint16_t *)POINTER_SHIFT(lhs, sizeof(uint16_t));
+  uint16_t right = *(uint16_t *)POINTER_SHIFT(rhs, sizeof(uint16_t));
+
+  if (left == right) {
+    return 0;
+  } else {
+    return left > right ? 1 : -1;
+  }
+}
 
 int32_t tsParseValues(char **str, STableDataBlocks *pDataBlock, int maxRows, SInsertStatementParam *pInsertParam,
     int32_t* numOfRows, char *tmpTokenBuf) {
@@ -551,21 +958,26 @@ int32_t tsParseValues(char **str, STableDataBlocks *pDataBlock, int maxRows, SIn
   
   int32_t  precision = tinfo.precision;
 
+  int32_t extendedRowSize = getExtendedRowSize(&tinfo);
+
+  initSMemRowHelper(&pDataBlock->rowHelper, tscGetTableSchema(pDataBlock->pTableMeta),
+                    tscGetNumOfColumns(pDataBlock->pTableMeta), 0);
+
   while (1) {
     index = 0;
     sToken = tStrGetToken(*str, &index, false);
     if (sToken.n == 0 || sToken.type != TK_LP) break;
 
     *str += index;
-    if ((*numOfRows) >= maxRows || pDataBlock->size + tinfo.rowSize >= pDataBlock->nAllocSize) {
+    if ((*numOfRows) >= maxRows || pDataBlock->size + extendedRowSize >= pDataBlock->nAllocSize) {
       int32_t tSize;
-      code = tscAllocateMemIfNeed(pDataBlock, tinfo.rowSize, &tSize);
+      code = tscAllocateMemIfNeed(pDataBlock, extendedRowSize, &tSize);
       if (code != TSDB_CODE_SUCCESS) {  //TODO pass the correct error code to client
         strcpy(pInsertParam->msg, "client out of memory");
         return TSDB_CODE_TSC_OUT_OF_MEMORY;
       }
 
-      ASSERT(tSize > maxRows);
+      ASSERT(tSize >= maxRows);
       maxRows = tSize;
     }
 
@@ -601,9 +1013,10 @@ int32_t tsParseValues(char **str, STableDataBlocks *pDataBlock, int maxRows, SIn
 void tscSetBoundColumnInfo(SParsedDataColInfo *pColInfo, SSchema *pSchema, int32_t numOfCols) {
   pColInfo->numOfCols = numOfCols;
   pColInfo->numOfBound = numOfCols;
-
+  pColInfo->orderStatus = ORDER_STATUS_ORDERED;
   pColInfo->boundedColumns = calloc(pColInfo->numOfCols, sizeof(int32_t));
   pColInfo->cols = calloc(pColInfo->numOfCols, sizeof(SBoundColumn));
+  pColInfo->colIdxInfo = NULL;
 
   for (int32_t i = 0; i < pColInfo->numOfCols; ++i) {
     if (i > 0) {
@@ -656,8 +1069,9 @@ static int32_t tsSetBlockInfo(SSubmitBlk *pBlocks, const STableMeta *pTableMeta,
   }
 }
 
+#if 0
 // data block is disordered, sort it in ascending order
-void tscSortRemoveDataBlockDupRows(STableDataBlocks *dataBuf) {
+static void tscSortRemoveDataBlockDupRowsOld(STableDataBlocks *dataBuf) {
   SSubmitBlk *pBlocks = (SSubmitBlk *)dataBuf->pData;
 
   // size is less than the total size, since duplicated rows may be removed yet.
@@ -700,12 +1114,91 @@ void tscSortRemoveDataBlockDupRows(STableDataBlocks *dataBuf) {
 
   dataBuf->prevTS = INT64_MIN;
 }
+#endif
+
+// data block is disordered, sort it in ascending order
+int tscSortRemoveDataBlockDupRows(STableDataBlocks *dataBuf, SBlockKeyInfo *pBlkKeyInfo) {
+  SSubmitBlk *pBlocks = (SSubmitBlk *)dataBuf->pData;
+  int16_t     nRows = pBlocks->numOfRows;
+
+  // size is less than the total size, since duplicated rows may be removed yet.
+
+  // if use server time, this block must be ordered
+  if (dataBuf->tsSource == TSDB_USE_SERVER_TS) {
+    assert(dataBuf->ordered);
+  }
+      // allocate memory
+  size_t nAlloc = nRows * sizeof(SBlockKeyTuple);
+  if (pBlkKeyInfo->pKeyTuple == NULL || pBlkKeyInfo->maxBytesAlloc < nAlloc) {
+    size_t nRealAlloc = nAlloc + 10 * sizeof(SBlockKeyTuple);
+    char * tmp = trealloc(pBlkKeyInfo->pKeyTuple, nRealAlloc);
+    if (tmp == NULL) {
+        return TSDB_CODE_TSC_OUT_OF_MEMORY;
+    }
+    pBlkKeyInfo->pKeyTuple = (SBlockKeyTuple *)tmp;
+    pBlkKeyInfo->maxBytesAlloc = (int32_t)nRealAlloc;
+  }
+  memset(pBlkKeyInfo->pKeyTuple, 0, nAlloc);
+
+  SBlockKeyTuple *pBlkKeyTuple = pBlkKeyInfo->pKeyTuple;
+  char *          pBlockData = pBlocks->data;
+  TDRowTLenT      totolPayloadTLen = 0;
+  TDRowTLenT      payloadTLen = 0;
+  int             n = 0;
+  while (n < nRows) {
+    pBlkKeyTuple->skey = payloadTSKey(pBlockData);
+    pBlkKeyTuple->payloadAddr = pBlockData;
+    payloadTLen = payloadTLen(pBlockData);
+#if 0
+    ASSERT(payloadNCols(pBlockData) <= 4096);
+    ASSERT(payloadTLen(pBlockData) < 65536);
+#endif
+    totolPayloadTLen += payloadTLen;
+    // next loop
+    pBlockData += payloadTLen;
+    ++pBlkKeyTuple;
+    ++n;
+  }
+
+  if (!dataBuf->ordered) {
+    pBlkKeyTuple = pBlkKeyInfo->pKeyTuple;
+    qsort(pBlkKeyTuple, nRows, sizeof(SBlockKeyTuple), rowDataCompar);
+
+    pBlkKeyTuple = pBlkKeyInfo->pKeyTuple;
+    int32_t i = 0;
+    int32_t j = 1;
+    while (j < nRows) {
+      TSKEY ti = (pBlkKeyTuple + i)->skey;
+      TSKEY tj = (pBlkKeyTuple + j)->skey;
+
+      if (ti == tj) {
+        totolPayloadTLen -= payloadTLen(pBlkKeyTuple + j);
+        ++j;
+        continue;
+      }
+
+      int32_t nextPos = (++i);
+      if (nextPos != j) {
+        memmove(pBlkKeyTuple + nextPos, pBlkKeyTuple + j, sizeof(SBlockKeyTuple));
+      }
+      ++j;
+    }
+
+    dataBuf->ordered = true;
+    pBlocks->numOfRows = i + 1;
+  }
+
+  dataBuf->size = sizeof(SSubmitBlk) + totolPayloadTLen;
+  dataBuf->prevTS = INT64_MIN;
+
+  return 0;
+}
 
 static int32_t doParseInsertStatement(SInsertStatementParam *pInsertParam, char **str, STableDataBlocks* dataBuf, int32_t *totalNum) {
   STableComInfo tinfo = tscGetTableInfo(dataBuf->pTableMeta);
   
   int32_t maxNumOfRows;
-  int32_t code = tscAllocateMemIfNeed(dataBuf, tinfo.rowSize, &maxNumOfRows);
+  int32_t code = tscAllocateMemIfNeed(dataBuf, getExtendedRowSize(&tinfo), &maxNumOfRows);
   if (TSDB_CODE_SUCCESS != code) {
     return TSDB_CODE_TSC_OUT_OF_MEMORY;
   }
@@ -1038,10 +1531,11 @@ static int32_t validateDataSource(SInsertStatementParam *pInsertParam, int32_t t
 
 static int32_t parseBoundColumns(SInsertStatementParam *pInsertParam, SParsedDataColInfo* pColInfo, SSchema* pSchema,
     char* str, char **end) {
-  pColInfo->numOfBound = 0;
+  int32_t nCols = pColInfo->numOfCols;
 
-  memset(pColInfo->boundedColumns, 0, sizeof(int32_t) * pColInfo->numOfCols);
+  pColInfo->numOfBound = 0;
-  for(int32_t i = 0; i < pColInfo->numOfCols; ++i) {
+  memset(pColInfo->boundedColumns, 0, sizeof(int32_t) * nCols);
+  for (int32_t i = 0; i < nCols; ++i) {
     pColInfo->cols[i].hasVal = false;
   }
 
@@ -1056,6 +1550,8 @@ static int32_t parseBoundColumns(SInsertStatementParam *pInsertParam, SParsedDat
     goto _clean;
   }
 
+  bool    isOrdered = true;
+  int32_t lastColIdx = -1;  // last column found
   while (1) {
     index = 0;
     sToken = tStrGetToken(str, &index, false);
@@ -1076,7 +1572,8 @@ static int32_t parseBoundColumns(SInsertStatementParam *pInsertParam, SParsedDat
     bool findColumnIndex = false;
 
     // todo speedup by using hash list
-    for (int32_t t = 0; t < pColInfo->numOfCols; ++t) {
+    int32_t nScanned = 0, t = lastColIdx + 1;
+    while (t < nCols) {
       if (strncmp(sToken.z, pSchema[t].name, sToken.n) == 0 && strlen(pSchema[t].name) == sToken.n) {
         if (pColInfo->cols[t].hasVal == true) {
           code = tscInvalidOperationMsg(pInsertParam->msg, "duplicated column name", sToken.z);
@@ -1085,10 +1582,39 @@ static int32_t parseBoundColumns(SInsertStatementParam *pInsertParam, SParsedDat
 
         pColInfo->cols[t].hasVal = true;
         pColInfo->boundedColumns[pColInfo->numOfBound] = t;
-        pColInfo->numOfBound += 1;
+        ++pColInfo->numOfBound;
         findColumnIndex = true;
+        if (isOrdered && (lastColIdx > t)) {
+          isOrdered = false;
+        }
+        lastColIdx = t;
         break;
       }
+      ++t;
+      ++nScanned;
+    }
+    if (!findColumnIndex) {
+      t = 0;
+      int32_t nRemain = nCols - nScanned;
+      while (t < nRemain) {
+        if (strncmp(sToken.z, pSchema[t].name, sToken.n) == 0 && strlen(pSchema[t].name) == sToken.n) {
+          if (pColInfo->cols[t].hasVal == true) {
+            code = tscInvalidOperationMsg(pInsertParam->msg, "duplicated column name", sToken.z);
+            goto _clean;
+          }
+
+          pColInfo->cols[t].hasVal = true;
+          pColInfo->boundedColumns[pColInfo->numOfBound] = t;
+          ++pColInfo->numOfBound;
+          findColumnIndex = true;
+          if (isOrdered && (lastColIdx > t)) {
+            isOrdered = false;
+          }
+          lastColIdx = t;
+          break;
+        }
+        ++t;
+      }
     }
 
     if (!findColumnIndex) {
@@ -1097,10 +1623,32 @@ static int32_t parseBoundColumns(SInsertStatementParam *pInsertParam, SParsedDat
     }
   }
 
-  memset(&pColInfo->boundedColumns[pColInfo->numOfBound], 0 , sizeof(int32_t) * (pColInfo->numOfCols - pColInfo->numOfBound));
+  pColInfo->orderStatus = isOrdered ? ORDER_STATUS_ORDERED : ORDER_STATUS_DISORDERED;
+
+  if (!isOrdered) {
+    pColInfo->colIdxInfo = tcalloc(pColInfo->numOfBound, sizeof(SBoundIdxInfo));
+    if (pColInfo->colIdxInfo == NULL) {
+      code = TSDB_CODE_TSC_OUT_OF_MEMORY;
+      goto _clean;
+    }
+    SBoundIdxInfo *pColIdx = pColInfo->colIdxInfo;
+    for (uint16_t i = 0; i < pColInfo->numOfBound; ++i) {
+      pColIdx[i].schemaColIdx = (uint16_t)pColInfo->boundedColumns[i];
+      pColIdx[i].boundIdx = i;
+    }
+    qsort(pColIdx, pColInfo->numOfBound, sizeof(SBoundIdxInfo), schemaIdxCompar);
+    for (uint16_t i = 0; i < pColInfo->numOfBound; ++i) {
+      pColIdx[i].finalIdx = i;
+    }
+    qsort(pColIdx, pColInfo->numOfBound, sizeof(SBoundIdxInfo), boundIdxCompar);
+  }
+
+  memset(&pColInfo->boundedColumns[pColInfo->numOfBound], 0,
+         sizeof(int32_t) * (pColInfo->numOfCols - pColInfo->numOfBound));
+
   return TSDB_CODE_SUCCESS;
 
-  _clean:
+_clean:
   pInsertParam->sql = NULL;
   return code;
 }
@@ -1494,21 +2042,24 @@ static void parseFileSendDataBlock(void *param, TAOS_RES *tres, int32_t numOfRow
   }
 
   STableDataBlocks *pTableDataBlock = NULL;
-  int32_t           ret =
+  int32_t ret = tscGetDataBlockFromList(pInsertParam->pTableBlockHashList, pTableMeta->id.uid, TSDB_PAYLOAD_SIZE,
-      tscGetDataBlockFromList(pInsertParam->pTableBlockHashList, pTableMeta->id.uid, TSDB_PAYLOAD_SIZE, sizeof(SSubmitBlk),
+                                        sizeof(SSubmitBlk), tinfo.rowSize, &pTableMetaInfo->name, pTableMeta,
-                              tinfo.rowSize, &pTableMetaInfo->name, pTableMeta, &pTableDataBlock, NULL);
+                                        &pTableDataBlock, NULL);
   if (ret != TSDB_CODE_SUCCESS) {
     pParentSql->res.code = TSDB_CODE_TSC_OUT_OF_MEMORY;
     goto _error;
   }
 
-  tscAllocateMemIfNeed(pTableDataBlock, tinfo.rowSize, &maxRows);
+  tscAllocateMemIfNeed(pTableDataBlock, getExtendedRowSize(&tinfo), &maxRows);
   tokenBuf = calloc(1, TSDB_MAX_BYTES_PER_ROW);
   if (tokenBuf == NULL) {
     code = TSDB_CODE_TSC_OUT_OF_MEMORY;
     goto _error;
   }
 
+  initSMemRowHelper(&pTableDataBlock->rowHelper, tscGetTableSchema(pTableDataBlock->pTableMeta),
+                    tscGetNumOfColumns(pTableDataBlock->pTableMeta), 0);
+
   while ((readLen = tgetline(&line, &n, fp)) != -1) {
     if (('\r' == line[readLen - 1]) || ('\n' == line[readLen - 1])) {
       line[--readLen] = 0;

src/client/src/tscPrepare.c

@@ -291,6 +291,7 @@ static char* normalStmtBuildSql(STscStmt* stmt) {
   return taosStringBuilderGetResult(&sb, NULL);
 }
 
+#if 0
 static int fillColumnsNull(STableDataBlocks* pBlock, int32_t rowNum) {
   SParsedDataColInfo* spd = &pBlock->boundColumnInfo;
   int32_t offset = 0;
@@ -318,8 +319,129 @@ static int fillColumnsNull(STableDataBlocks* pBlock, int32_t rowNum) {
 
   return TSDB_CODE_SUCCESS;
 }
+#endif
+
+/**
+ * input:
+ * - schema:
+ * - payload:
+ * - spd:
+ * output:
+ * - pBlock with data block replaced by K-V format
+ */
+static int refactorPayload(STableDataBlocks* pBlock, int32_t rowNum) {
+  SParsedDataColInfo* spd = &pBlock->boundColumnInfo;
+  SSchema*            schema = (SSchema*)pBlock->pTableMeta->schema;
+  SMemRowHelper*      pHelper = &pBlock->rowHelper;
+  STableMeta*         pTableMeta = pBlock->pTableMeta;
+  STableComInfo       tinfo = tscGetTableInfo(pTableMeta);
+  int                 code = TSDB_CODE_SUCCESS;
+  int32_t             extendedRowSize = getExtendedRowSize(&tinfo);
+  TDRowTLenT          destPayloadSize = sizeof(SSubmitBlk);
+
+  ASSERT(pHelper->allNullLen >= 8);
+
+  TDRowTLenT  destAllocSize = sizeof(SSubmitBlk) + rowNum * extendedRowSize;
+  SSubmitBlk* pDestBlock = tcalloc(destAllocSize, 1);
+  if (pDestBlock == NULL) {
+    return TSDB_CODE_TSC_OUT_OF_MEMORY;
+  }
+  memcpy(pDestBlock, pBlock->pData, sizeof(SSubmitBlk));
+  char* destPayload = (char*)pDestBlock + sizeof(SSubmitBlk);
+
+  char* srcPayload = (char*)pBlock->pData + sizeof(SSubmitBlk);
+
+  for (int n = 0; n < rowNum; ++n) {
+    payloadSetNCols(destPayload, spd->numOfBound);
+
+    TDRowTLenT dataRowLen = pHelper->allNullLen;
+    TDRowTLenT kvRowLen = TD_MEM_ROW_KV_VER_SIZE + sizeof(SColIdx) * spd->numOfBound;
+    TDRowTLenT payloadValOffset = payloadValuesOffset(destPayload);  // rely on payloadNCols
+    TDRowLenT  colValOffset = 0;
+
+    char* kvPrimaryKeyStart = destPayload + PAYLOAD_HEADER_LEN;  // primaryKey in 1st column tuple
+    char* kvStart = kvPrimaryKeyStart + PAYLOAD_COL_HEAD_LEN;    // the column tuple behind the primaryKey
+
+    for (int32_t i = 0; i < spd->numOfBound; ++i) {
+      int32_t colIndex = spd->boundedColumns[i];
+      ASSERT(spd->cols[colIndex].hasVal);
+      char*    start = srcPayload + spd->cols[colIndex].offset;
+      SSchema* pSchema = &schema[colIndex];  // get colId here
+      bool     isPrimaryKey = (colIndex == PRIMARYKEY_TIMESTAMP_COL_INDEX);
+
+      // the primary key locates in 1st column
+      if (!IS_DATA_COL_ORDERED(spd->orderStatus)) {
+        ASSERT(spd->colIdxInfo != NULL);
+        if (!isPrimaryKey) {
+          kvStart = POINTER_SHIFT(kvPrimaryKeyStart, spd->colIdxInfo[i].finalIdx * PAYLOAD_COL_HEAD_LEN);
+        } else {
+          ASSERT(spd->colIdxInfo[i].finalIdx == 0);
+        }
+      }
+      if (isPrimaryKey) {
+        payloadColSetId(kvPrimaryKeyStart, pSchema->colId);
+        payloadColSetType(kvPrimaryKeyStart, pSchema->type);
+        payloadColSetOffset(kvPrimaryKeyStart, colValOffset);
+        memcpy(POINTER_SHIFT(destPayload, payloadValOffset + colValOffset), start, TYPE_BYTES[pSchema->type]);
+        colValOffset += TYPE_BYTES[pSchema->type];
+        kvRowLen += TYPE_BYTES[pSchema->type];
+      } else {
+        payloadColSetId(kvStart, pSchema->colId);
+        payloadColSetType(kvStart, pSchema->type);
+        payloadColSetOffset(kvStart, colValOffset);
+        if (IS_VAR_DATA_TYPE(pSchema->type)) {
+          varDataCopy(POINTER_SHIFT(destPayload, payloadValOffset + colValOffset), start);
+          colValOffset += varDataTLen(start);
+          kvRowLen += varDataTLen(start);
+          if (pSchema->type == TSDB_DATA_TYPE_BINARY) {
+            dataRowLen += (varDataLen(start) - CHAR_BYTES);
+          } else if (pSchema->type == TSDB_DATA_TYPE_NCHAR) {
+            dataRowLen += (varDataLen(start) - TSDB_NCHAR_SIZE);
+          } else {
+            ASSERT(0);
+          }
+        } else {
+          memcpy(POINTER_SHIFT(destPayload, payloadValOffset + colValOffset), start, TYPE_BYTES[pSchema->type]);
+          colValOffset += TYPE_BYTES[pSchema->type];
+          kvRowLen += TYPE_BYTES[pSchema->type];
+        }
+
+        if (IS_DATA_COL_ORDERED(spd->orderStatus)) {
+          kvStart += PAYLOAD_COL_HEAD_LEN;  // move to next column
+        }
+      }
 
 
+    }  // end of column
+
+    if (kvRowLen < dataRowLen) {
+      payloadSetType(destPayload, SMEM_ROW_KV);
+    } else {
+      payloadSetType(destPayload, SMEM_ROW_DATA);
+    }
+
+    ASSERT(colValOffset <= TSDB_MAX_BYTES_PER_ROW);
+
+    TDRowTLenT len = payloadValOffset + colValOffset;
+    payloadSetTLen(destPayload, len);
+
+    // next loop
+    srcPayload += pBlock->rowSize;
+    destPayload += len;
+
+    destPayloadSize += len;
+  }  // end of row
+  
+  ASSERT(destPayloadSize <= destAllocSize);
+
+  tfree(pBlock->pData);
+  pBlock->pData = (char*)pDestBlock;
+  pBlock->nAllocSize = destAllocSize;
+  pBlock->size = destPayloadSize;
+
+  return code;
+}
+#if 0
 int32_t fillTablesColumnsNull(SSqlObj* pSql) {
   SSqlCmd* pCmd = &pSql->cmd;
 
@@ -342,17 +464,98 @@ int32_t fillTablesColumnsNull(SSqlObj* pSql) {
 
   return TSDB_CODE_SUCCESS;
 }
+#endif
 
-
+/**
-
+ * check and sort
-////////////////////////////////////////////////////////////////////////////////
+ */
-// functions for insertion statement preparation
+static int initPayloadEnv(STableDataBlocks* pBlock, int32_t rowNum) {
-static int doBindParam(STableDataBlocks* pBlock, char* data, SParamInfo* param, TAOS_BIND* bind, int32_t colNum) {
+  SParsedDataColInfo* spd = &pBlock->boundColumnInfo;
-  if (bind->is_null != NULL && *(bind->is_null)) {
+  if (spd->orderStatus != ORDER_STATUS_UNKNOWN) {
-    setNull(data + param->offset, param->type, param->bytes);
     return TSDB_CODE_SUCCESS;
   }
 
+  bool isOrdered = true;
+  int32_t lastColIdx = -1;
+  for (int32_t i = 0; i < spd->numOfBound; ++i) {
+    ASSERT(spd->cols[i].hasVal);
+    int32_t colIdx = spd->boundedColumns[i];
+    if (isOrdered) {
+      if (lastColIdx > colIdx) {
+        isOrdered = false;
+        break;
+      } else {
+        lastColIdx = colIdx;
+      }
+    }
+  }
+
+  spd->orderStatus = isOrdered ? ORDER_STATUS_ORDERED : ORDER_STATUS_DISORDERED;
+
+  if (isOrdered) {
+    spd->colIdxInfo = NULL;
+  } else {
+    spd->colIdxInfo = calloc(spd->numOfBound, sizeof(SBoundIdxInfo));
+    if (spd->colIdxInfo == NULL) {
+      return TSDB_CODE_TSC_OUT_OF_MEMORY;
+    }
+    SBoundIdxInfo* pColIdx = spd->colIdxInfo;
+    for (uint16_t i = 0; i < spd->numOfBound; ++i) {
+      pColIdx[i].schemaColIdx = (uint16_t)spd->boundedColumns[i];
+      pColIdx[i].boundIdx = i;
+    }
+    qsort(pColIdx, spd->numOfBound, sizeof(SBoundIdxInfo), schemaIdxCompar);
+    for (uint16_t i = 0; i < spd->numOfBound; ++i) {
+      pColIdx[i].finalIdx = i;
+    }
+    qsort(pColIdx, spd->numOfBound, sizeof(SBoundIdxInfo), boundIdxCompar);
+  }
+
+  return TSDB_CODE_SUCCESS;
+}
+
+/**
+ *  Refactor the raw payload structure to K-V format as the in tsParseOneRow()
+ */
+int32_t fillTablesPayload(SSqlObj* pSql) {
+  SSqlCmd* pCmd = &pSql->cmd;
+  int      code = TSDB_CODE_SUCCESS;
+
+  STableDataBlocks** p = taosHashIterate(pCmd->insertParam.pTableBlockHashList, NULL);
+
+  STableDataBlocks* pOneTableBlock = *p;
+  while (pOneTableBlock) {
+    SSubmitBlk* pBlocks = (SSubmitBlk*)pOneTableBlock->pData;
+
+    if (pBlocks->numOfRows > 0) {
+      initSMemRowHelper(&pOneTableBlock->rowHelper, tscGetTableSchema(pOneTableBlock->pTableMeta),
+                        tscGetNumOfColumns(pOneTableBlock->pTableMeta), 0);
+      if ((code = initPayloadEnv(pOneTableBlock, pBlocks->numOfRows)) != TSDB_CODE_SUCCESS) {
+        return code;
+      }
+      if ((code = refactorPayload(pOneTableBlock, pBlocks->numOfRows)) != TSDB_CODE_SUCCESS) {
+        return code;
+      };
+    }
+
+    p = taosHashIterate(pCmd->insertParam.pTableBlockHashList, p);
+    if (p == NULL) {
+      break;
+    }
+
+    pOneTableBlock = *p;
+  }
+
+  return code;
+}
+  ////////////////////////////////////////////////////////////////////////////////
+  // functions for insertion statement preparation
+  static int doBindParam(STableDataBlocks* pBlock, char* data, SParamInfo* param, TAOS_BIND* bind, int32_t colNum) {
+    if (bind->is_null != NULL && *(bind->is_null)) {
+      setNull(data + param->offset, param->type, param->bytes);
+      return TSDB_CODE_SUCCESS;
+    }
+
 #if 0
   if (0) {
     // allow user bind param data with different type
@@ -1106,9 +1309,12 @@ static int insertStmtExecute(STscStmt* stmt) {
   pBlk->uid = pTableMeta->id.uid;
   pBlk->tid = pTableMeta->id.tid;
 
-  fillTablesColumnsNull(stmt->pSql);
+  int code = fillTablesPayload(stmt->pSql);
+  if (code != TSDB_CODE_SUCCESS) {
+    return code;
+  }
 
-  int code = tscMergeTableDataBlocks(&stmt->pSql->cmd.insertParam, false);
+  code = tscMergeTableDataBlocks(&stmt->pSql->cmd.insertParam, false);
   if (code != TSDB_CODE_SUCCESS) {
     return code;
   }
@@ -1185,7 +1391,7 @@ static int insertBatchStmtExecute(STscStmt* pStmt) {
     return TSDB_CODE_TSC_APP_ERROR;
   }
 
-  fillTablesColumnsNull(pStmt->pSql);
+  fillTablesPayload(pStmt->pSql);
 
   if ((code = tscMergeTableDataBlocks(&pStmt->pSql->cmd.insertParam, false)) != TSDB_CODE_SUCCESS) {
     return code;

src/client/src/tscSQLParser.c

@@ -1396,12 +1396,16 @@ static bool validateTableColumnInfo(SArray* pFieldList, SSqlCmd* pCmd) {
   const char* msg4 = "invalid data type";
   const char* msg5 = "invalid binary/nchar column length";
   const char* msg6 = "invalid column name";
+  const char* msg7 = "too many columns";
 
   // number of fields no less than 2
   size_t numOfCols = taosArrayGetSize(pFieldList);
-  if (numOfCols <= 1 || numOfCols > TSDB_MAX_COLUMNS) {
+  if (numOfCols <= 1 ) {
     invalidOperationMsg(tscGetErrorMsgPayload(pCmd), msg);
     return false;
+  } else if (numOfCols > TSDB_MAX_COLUMNS) {
+    invalidOperationMsg(tscGetErrorMsgPayload(pCmd), msg7);
+    return false;
   }
 
   // first column must be timestamp
@@ -8055,8 +8059,12 @@ int32_t loadAllTableMeta(SSqlObj* pSql, struct SSqlInfo* pInfo) {
   uint32_t maxSize = tscGetTableMetaMaxSize();
   char     name[TSDB_TABLE_FNAME_LEN] = {0};
 
-  char buf[80 * 1024] = {0};
+  assert(maxSize < 80 * TSDB_MAX_COLUMNS);
-  assert(maxSize < 80 * 1024);
+  if (!pSql->pBuf) {
+    if (NULL == (pSql->pBuf = tcalloc(1, 80 * TSDB_MAX_COLUMNS))) {
+      return TSDB_CODE_TSC_OUT_OF_MEMORY;
+    }
+  }
   pTableMeta = calloc(1, maxSize);
 
   plist = taosArrayInit(4, POINTER_BYTES);
@@ -8073,7 +8081,7 @@ int32_t loadAllTableMeta(SSqlObj* pSql, struct SSqlInfo* pInfo) {
 
     if (pTableMeta->id.uid > 0) {
       if (pTableMeta->tableType == TSDB_CHILD_TABLE) {
-        code = tscCreateTableMetaFromSTableMeta(pTableMeta, name, buf);
+        code = tscCreateTableMetaFromSTableMeta(pTableMeta, name, pSql->pBuf);
 
         // create the child table meta from super table failed, try load it from mnode
         if (code != TSDB_CODE_SUCCESS) {

src/client/src/tscServer.c

@@ -2804,22 +2804,24 @@ int32_t tscGetTableMetaImpl(SSqlObj* pSql, STableMetaInfo *pTableMetaInfo, bool
   taosHashGetClone(tscTableMetaInfo, name, len, NULL, pTableMetaInfo->pTableMeta, -1);
 
   // TODO resize the tableMeta
-  char buf[80*1024] = {0};
+  assert(size < 80 * TSDB_MAX_COLUMNS);
-  assert(size < 80*1024);
+  if (!pSql->pBuf) {
+    if (NULL == (pSql->pBuf = tcalloc(1, 80 * TSDB_MAX_COLUMNS))) {
+      return TSDB_CODE_TSC_OUT_OF_MEMORY;
+    }
+  }
 
   STableMeta* pMeta = pTableMetaInfo->pTableMeta;
   if (pMeta->id.uid > 0) {
     // in case of child table, here only get the
     if (pMeta->tableType == TSDB_CHILD_TABLE) {
-      int32_t code = tscCreateTableMetaFromSTableMeta(pTableMetaInfo->pTableMeta, name, buf);
+      int32_t code = tscCreateTableMetaFromSTableMeta(pTableMetaInfo->pTableMeta, name, pSql->pBuf);
       if (code != TSDB_CODE_SUCCESS) {
         return getTableMetaFromMnode(pSql, pTableMetaInfo, autocreate);
       }
     }
-
     return TSDB_CODE_SUCCESS;
   }
-
   return getTableMetaFromMnode(pSql, pTableMetaInfo, autocreate);
 }
 
@@ -2934,7 +2936,6 @@ int tscGetSTableVgroupInfo(SSqlObj *pSql, SQueryInfo* pQueryInfo) {
   if (allVgroupInfoRetrieved(pQueryInfo)) {
     return TSDB_CODE_SUCCESS;
   }
-
   SSqlObj *pNew = calloc(1, sizeof(SSqlObj));
   pNew->pTscObj = pSql->pTscObj;
   pNew->signature = pNew;

src/client/src/tscSubquery.c

@@ -2436,9 +2436,9 @@ int32_t tscHandleMasterSTableQuery(SSqlObj *pSql) {
   tOrderDescriptor *pDesc  = NULL;
 
   pRes->qId = 0x1;  // hack the qhandle check
-  
+
-  const uint32_t nBufferSize = (1u << 16u);  // 64KB
+  const uint32_t nBufferSize = (1u << 18u);  // 256KB
-  
+
   SQueryInfo     *pQueryInfo = tscGetQueryInfo(pCmd);
   STableMetaInfo *pTableMetaInfo = tscGetMetaInfo(pQueryInfo, 0);
   SSubqueryState *pState = &pSql->subState;

src/client/src/tscUtil.c

@@ -1488,6 +1488,7 @@ void tscFreeSqlObj(SSqlObj* pSql) {
   pSql->signature = NULL;
   pSql->fp = NULL;
   tfree(pSql->sqlstr);
+  tfree(pSql->pBuf);
 
   tfree(pSql->pSubs);
   pSql->subState.numOfSub = 0;
@@ -1499,7 +1500,7 @@ void tscFreeSqlObj(SSqlObj* pSql) {
   memset(pCmd->payload, 0, (size_t)pCmd->allocSize);
   tfree(pCmd->payload);
   pCmd->allocSize = 0;
-  
+
   tsem_destroy(&pSql->rspSem);
   memset(pSql, 0, sizeof(*pSql));
   free(pSql);
@@ -1508,6 +1509,7 @@ void tscFreeSqlObj(SSqlObj* pSql) {
 void tscDestroyBoundColumnInfo(SParsedDataColInfo* pColInfo) {
   tfree(pColInfo->boundedColumns);
   tfree(pColInfo->cols);
+  tfree(pColInfo->colIdxInfo);
 }
 
 void tscDestroyDataBlock(STableDataBlocks* pDataBlock, bool removeMeta) {
@@ -1762,11 +1764,107 @@ int32_t tscGetDataBlockFromList(SHashObj* pHashList, int64_t id, int32_t size, i
   return TSDB_CODE_SUCCESS;
 }
 
-static int trimDataBlock(void* pDataBlock, STableDataBlocks* pTableDataBlock, bool includeSchema) {
+static SMemRow tdGenMemRowFromBuilder(SMemRowBuilder* pBuilder) {
+  SSchema* pSchema = pBuilder->pSchema;
+  char*    p = (char*)pBuilder->buf;
+  int      toffset = 0;
+  uint16_t nCols = pBuilder->nCols;
+
+  uint8_t  memRowType = payloadType(p);
+  uint16_t nColsBound = payloadNCols(p);
+  if (pBuilder->nCols <= 0 || nColsBound <= 0) {
+    return NULL;
+  }
+  char*    pVals = POINTER_SHIFT(p, payloadValuesOffset(p));
+  SMemRow* memRow = (SMemRow)pBuilder->pDataBlock;
+  memRowSetType(memRow, memRowType);
+
+  // ----------------- Raw payload structure for row:
+  /* |<------------ Head ------------->|<----------- body of column data tuple ------------------->|
+   * |                                 |<----------------- flen ------------->|<--- value part --->|
+   * |SMemRowType| dataTLen |  nCols   |  colId  | colType | offset   |  ...  | value |...|...|... |
+   * +-----------+----------+----------+--------------------------------------|--------------------|
+   * | uint8_t   | uint32_t | uint16_t | int16_t | uint8_t | uint16_t |  ...  |.......|...|...|... |
+   * +-----------+----------+----------+--------------------------------------+--------------------|
+   *  1. offset in column data tuple starts from the value part in case of uint16_t overflow.
+   *  2. dataTLen: total length including the header and body.
+   */
+
+  if (memRowType == SMEM_ROW_DATA) {
+    SDataRow trow = (SDataRow)memRowDataBody(memRow);
+    dataRowSetLen(trow, (TDRowLenT)(TD_DATA_ROW_HEAD_SIZE + pBuilder->flen));
+    dataRowSetVersion(trow, pBuilder->sversion);
+
+    p = (char*)payloadBody(pBuilder->buf);
+    uint16_t i = 0, j = 0;
+    while (j < nCols) {
+      if (i >= nColsBound) {
+        break;
+      }
+      int16_t colId = payloadColId(p);
+      if (colId == pSchema[j].colId) {
+        // ASSERT(payloadColType(p) == pSchema[j].type);
+        tdAppendColVal(trow, POINTER_SHIFT(pVals, payloadColOffset(p)), pSchema[j].type, toffset);
+        toffset += TYPE_BYTES[pSchema[j].type];
+        p = payloadNextCol(p);
+        ++i;
+        ++j;
+      } else if (colId < pSchema[j].colId) {
+        p = payloadNextCol(p);
+        ++i;
+      } else {
+        tdAppendColVal(trow, tdGetNullVal(pSchema[j].type), pSchema[j].type, toffset);
+        toffset += TYPE_BYTES[pSchema[j].type];
+        ++j;
+      }
+    }
+
+    while (j < nCols) {
+      tdAppendColVal(trow, tdGetNullVal(pSchema[j].type), pSchema[j].type, toffset);
+      toffset += TYPE_BYTES[pSchema[j].type];
+      ++j;
+    }
+    
+    #if 0 // no need anymore
+    while (i < nColsBound) {
+      p = payloadNextCol(p);
+      ++i;
+    }
+    #endif
+
+  } else if (memRowType == SMEM_ROW_KV) {
+    SKVRow   kvRow = (SKVRow)memRowKvBody(memRow);
+    kvRowSetLen(kvRow, (TDRowLenT)(TD_KV_ROW_HEAD_SIZE + sizeof(SColIdx) * nColsBound));
+    kvRowSetNCols(kvRow, nColsBound);
+    memRowSetKvVersion(memRow, pBuilder->sversion);
+
+    p = (char*)payloadBody(pBuilder->buf);
+    int i = 0;
+    while (i < nColsBound) {
+      int16_t colId = payloadColId(p);
+      uint8_t colType = payloadColType(p);
+      tdAppendKvColVal(kvRow, POINTER_SHIFT(pVals,payloadColOffset(p)), colId, colType, toffset);
+      toffset += sizeof(SColIdx);
+      p = payloadNextCol(p);
+      ++i;
+    }
+
+  } else {
+    ASSERT(0);
+  }
+  int32_t rowTLen = memRowTLen(memRow);
+  pBuilder->pDataBlock = (char*)pBuilder->pDataBlock + rowTLen;  // next row
+  pBuilder->pSubmitBlk->dataLen += rowTLen;
+
+  return memRow;
+}
+
+// Erase the empty space reserved for binary data
+static int trimDataBlock(void* pDataBlock, STableDataBlocks* pTableDataBlock, bool includeSchema, SBlockKeyTuple *blkKeyTuple) {
   // TODO: optimize this function, handle the case while binary is not presented
-  STableMeta*   pTableMeta = pTableDataBlock->pTableMeta;
+  STableMeta*     pTableMeta = pTableDataBlock->pTableMeta;
-  STableComInfo tinfo = tscGetTableInfo(pTableMeta);
+  STableComInfo   tinfo = tscGetTableInfo(pTableMeta);
-  SSchema*      pSchema = tscGetTableSchema(pTableMeta);
+  SSchema*        pSchema = tscGetTableSchema(pTableMeta);
 
   SSubmitBlk* pBlock = pDataBlock;
   memcpy(pDataBlock, pTableDataBlock->pData, sizeof(SSubmitBlk));
@@ -1801,21 +1899,19 @@ static int trimDataBlock(void* pDataBlock, STableDataBlocks* pTableDataBlock, bo
   char* p = pTableDataBlock->pData + sizeof(SSubmitBlk);
   pBlock->dataLen = 0;
   int32_t numOfRows = htons(pBlock->numOfRows);
-  
+
+  SMemRowBuilder rowBuilder;
+  rowBuilder.pSchema = pSchema;
+  rowBuilder.sversion = pTableMeta->sversion;
+  rowBuilder.flen = flen;
+  rowBuilder.nCols = tinfo.numOfColumns;
+  rowBuilder.pDataBlock = pDataBlock;
+  rowBuilder.pSubmitBlk = pBlock;
+  rowBuilder.buf = p;
+
   for (int32_t i = 0; i < numOfRows; ++i) {
-    SDataRow trow = (SDataRow) pDataBlock;
+    rowBuilder.buf = (blkKeyTuple + i)->payloadAddr;
-    dataRowSetLen(trow, (uint16_t)(TD_DATA_ROW_HEAD_SIZE + flen));
+    tdGenMemRowFromBuilder(&rowBuilder);
-    dataRowSetVersion(trow, pTableMeta->sversion);
-
-    int toffset = 0;
-    for (int32_t j = 0; j < tinfo.numOfColumns; j++) {
-      tdAppendColVal(trow, p, pSchema[j].type, pSchema[j].bytes, toffset);
-      toffset += TYPE_BYTES[pSchema[j].type];
-      p += pSchema[j].bytes;
-    }
-
-    pDataBlock = (char*)pDataBlock + dataRowLen(trow);
-    pBlock->dataLen += dataRowLen(trow);
   }
 
   int32_t len = pBlock->dataLen + pBlock->schemaLen;
@@ -1826,7 +1922,7 @@ static int trimDataBlock(void* pDataBlock, STableDataBlocks* pTableDataBlock, bo
 }
 
 static int32_t getRowExpandSize(STableMeta* pTableMeta) {
-  int32_t result = TD_DATA_ROW_HEAD_SIZE;
+  int32_t  result = TD_MEM_ROW_DATA_HEAD_SIZE;
   int32_t columns = tscGetNumOfColumns(pTableMeta);
   SSchema* pSchema = tscGetTableSchema(pTableMeta);
   for(int32_t i = 0; i < columns; i++) {
@@ -1862,13 +1958,16 @@ static void extractTableNameList(SInsertStatementParam *pInsertParam, bool freeB
 
 int32_t tscMergeTableDataBlocks(SInsertStatementParam *pInsertParam, bool freeBlockMap) {
   const int INSERT_HEAD_SIZE = sizeof(SMsgDesc) + sizeof(SSubmitMsg);
-
+  int       code = 0;
-  void* pVnodeDataBlockHashList = taosHashInit(128, taosGetDefaultHashFunction(TSDB_DATA_TYPE_BIGINT), true, false);
+  void*     pVnodeDataBlockHashList = taosHashInit(128, taosGetDefaultHashFunction(TSDB_DATA_TYPE_BIGINT), true, false);
-  SArray* pVnodeDataBlockList = taosArrayInit(8, POINTER_BYTES);
+  SArray*   pVnodeDataBlockList = taosArrayInit(8, POINTER_BYTES);
 
   STableDataBlocks** p = taosHashIterate(pInsertParam->pTableBlockHashList, NULL);
 
   STableDataBlocks* pOneTableBlock = *p;
+
+  SBlockKeyInfo blkKeyInfo = {0};  // share by pOneTableBlock
+  
   while(pOneTableBlock) {
     SSubmitBlk* pBlocks = (SSubmitBlk*) pOneTableBlock->pData;
     if (pBlocks->numOfRows > 0) {
@@ -1882,6 +1981,7 @@ int32_t tscMergeTableDataBlocks(SInsertStatementParam *pInsertParam, bool freeBl
         tscError("0x%"PRIx64" failed to prepare the data block buffer for merging table data, code:%d", pInsertParam->objectId, ret);
         taosHashCleanup(pVnodeDataBlockHashList);
         tscDestroyBlockArrayList(pVnodeDataBlockList);
+        tfree(blkKeyInfo.pKeyTuple);
         return ret;
       }
 
@@ -1902,16 +2002,26 @@ int32_t tscMergeTableDataBlocks(SInsertStatementParam *pInsertParam, bool freeBl
           taosHashCleanup(pVnodeDataBlockHashList);
           tscDestroyBlockArrayList(pVnodeDataBlockList);
           tfree(dataBuf->pData);
+          tfree(blkKeyInfo.pKeyTuple);
 
           return TSDB_CODE_TSC_OUT_OF_MEMORY;
         }
       }
 
-      tscSortRemoveDataBlockDupRows(pOneTableBlock);
+      if ((code = tscSortRemoveDataBlockDupRows(pOneTableBlock, &blkKeyInfo)) != 0) {
-      char* ekey = (char*)pBlocks->data + pOneTableBlock->rowSize*(pBlocks->numOfRows-1);
+        taosHashCleanup(pVnodeDataBlockHashList);
+        tscDestroyBlockArrayList(pVnodeDataBlockList);
+        tfree(dataBuf->pData);
+        tfree(blkKeyInfo.pKeyTuple);
+        return code;
+      }
 
-      tscDebug("0x%"PRIx64" name:%s, tid:%d rows:%d sversion:%d skey:%" PRId64 ", ekey:%" PRId64, pInsertParam->objectId, tNameGetTableName(&pOneTableBlock->tableName),
+      ASSERT(blkKeyInfo.pKeyTuple != NULL && pBlocks->numOfRows > 0);
-          pBlocks->tid, pBlocks->numOfRows, pBlocks->sversion, GET_INT64_VAL(pBlocks->data), GET_INT64_VAL(ekey));
+
+      SBlockKeyTuple* pLastKeyTuple = blkKeyInfo.pKeyTuple + pBlocks->numOfRows - 1;
+      tscDebug("0x%" PRIx64 " name:%s, tid:%d rows:%d sversion:%d skey:%" PRId64 ", ekey:%" PRId64,
+               pInsertParam->objectId, tNameGetTableName(&pOneTableBlock->tableName), pBlocks->tid, pBlocks->numOfRows,
+               pBlocks->sversion, blkKeyInfo.pKeyTuple->skey, pLastKeyTuple->skey);
 
       int32_t len = pBlocks->numOfRows * (pOneTableBlock->rowSize + expandSize) + sizeof(STColumn) * tscGetNumOfColumns(pOneTableBlock->pTableMeta);
 
@@ -1922,7 +2032,7 @@ int32_t tscMergeTableDataBlocks(SInsertStatementParam *pInsertParam, bool freeBl
       pBlocks->schemaLen = 0;
 
       // erase the empty space reserved for binary data
-      int32_t finalLen = trimDataBlock(dataBuf->pData + dataBuf->size, pOneTableBlock, pInsertParam->schemaAttached);
+      int32_t finalLen = trimDataBlock(dataBuf->pData + dataBuf->size, pOneTableBlock, pInsertParam->schemaAttached, blkKeyInfo.pKeyTuple);
       assert(finalLen <= len);
 
       dataBuf->size += (finalLen + sizeof(SSubmitBlk));
@@ -1950,6 +2060,7 @@ int32_t tscMergeTableDataBlocks(SInsertStatementParam *pInsertParam, bool freeBl
   // free the table data blocks;
   pInsertParam->pDataBlocks = pVnodeDataBlockList;
   taosHashCleanup(pVnodeDataBlockHashList);
+  tfree(blkKeyInfo.pKeyTuple);
 
   return TSDB_CODE_SUCCESS;
 }

src/common/inc/tdataformat.h

@@ -24,6 +24,35 @@
 extern "C" {
 #endif
 
+#pragma pack(push, 1)
+typedef struct {
+  VarDataLenT len;
+  uint8_t     data;
+} SBinaryNullT;
+
+typedef struct {
+  VarDataLenT len;
+  uint32_t    data;
+} SNCharNullT;
+#pragma pack(pop)
+
+extern const uint8_t      BoolNull;
+extern const uint8_t      TinyintNull;
+extern const uint16_t     SmallintNull;
+extern const uint32_t     IntNull;
+extern const uint64_t     BigintNull;
+extern const uint64_t     TimestampNull;
+extern const uint8_t      UTinyintNull;
+extern const uint16_t     USmallintNull;
+extern const uint32_t     UIntNull;
+extern const uint64_t     UBigintNull;
+extern const uint32_t     FloatNull;
+extern const uint64_t     DoubleNull;
+extern const SBinaryNullT BinaryNull;
+extern const SNCharNullT  NcharNull;
+
+const void *tdGetNullVal(int8_t type);
+
 #define STR_TO_VARSTR(x, str)                     \
   do {                                            \
     VarDataLenT __len = (VarDataLenT)strlen(str); \
@@ -53,10 +82,10 @@ extern "C" {
 
 // ----------------- TSDB COLUMN DEFINITION
 typedef struct {
-  int8_t  type;    // Column type
+  int8_t   type;    // Column type
-  int16_t colId;   // column ID
+  int16_t  colId;   // column ID
-  int16_t bytes;   // column bytes
+  uint16_t bytes;   // column bytes
-  int16_t offset;  // point offset in SDataRow after the header part
+  uint16_t offset;  // point offset in SDataRow after the header part.
 } STColumn;
 
 #define colType(col) ((col)->type)
@@ -167,10 +196,11 @@ static FORCE_INLINE int tkeyComparFn(const void *tkey1, const void *tkey2) {
     return 0;
   }
 }
+
 // ----------------- Data row structure
 
 /* A data row, the format is like below:
- * |<--------------------+--------------------------- len ---------------------------------->|
+ * |<------------------------------------------------ len ---------------------------------->|
  * |<--     Head      -->|<---------   flen -------------->|                                 |
  * +---------------------+---------------------------------+---------------------------------+
  * | uint16_t |  int16_t |                                 |                                 |
@@ -184,8 +214,8 @@ typedef void *SDataRow;
 
 #define TD_DATA_ROW_HEAD_SIZE (sizeof(uint16_t) + sizeof(int16_t))
 
-#define dataRowLen(r) (*(uint16_t *)(r))
+#define dataRowLen(r) (*(TDRowLenT *)(r))  // 0~65535
-#define dataRowVersion(r) *(int16_t *)POINTER_SHIFT(r, sizeof(int16_t))
+#define dataRowVersion(r) (*(int16_t *)POINTER_SHIFT(r, sizeof(int16_t)))
 #define dataRowTuple(r) POINTER_SHIFT(r, TD_DATA_ROW_HEAD_SIZE)
 #define dataRowTKey(r) (*(TKEY *)(dataRowTuple(r)))
 #define dataRowKey(r) tdGetKey(dataRowTKey(r))
@@ -201,20 +231,19 @@ void     tdInitDataRow(SDataRow row, STSchema *pSchema);
 SDataRow tdDataRowDup(SDataRow row);
 
 // offset here not include dataRow header length
-static FORCE_INLINE int tdAppendColVal(SDataRow row, void *value, int8_t type, int32_t bytes, int32_t offset) {
+static FORCE_INLINE int tdAppendColVal(SDataRow row, const void *value, int8_t type, int32_t offset) {
   ASSERT(value != NULL);
   int32_t toffset = offset + TD_DATA_ROW_HEAD_SIZE;
-  char *  ptr = (char *)POINTER_SHIFT(row, dataRowLen(row));
 
   if (IS_VAR_DATA_TYPE(type)) {
     *(VarDataOffsetT *)POINTER_SHIFT(row, toffset) = dataRowLen(row);
-    memcpy(ptr, value, varDataTLen(value));
+    memcpy(POINTER_SHIFT(row, dataRowLen(row)), value, varDataTLen(value));
     dataRowLen(row) += varDataTLen(value);
   } else {
     if (offset == 0) {
       ASSERT(type == TSDB_DATA_TYPE_TIMESTAMP);
       TKEY tvalue = tdGetTKEY(*(TSKEY *)value);
-      memcpy(POINTER_SHIFT(row, toffset), (void *)(&tvalue), TYPE_BYTES[type]);
+      memcpy(POINTER_SHIFT(row, toffset), (const void *)(&tvalue), TYPE_BYTES[type]);
     } else {
       memcpy(POINTER_SHIFT(row, toffset), value, TYPE_BYTES[type]);
     }
@@ -245,17 +274,21 @@ typedef struct SDataCol {
   TSKEY           ts;         // only used in last NULL column
 } SDataCol;
 
+#define isAllRowsNull(pCol) ((pCol)->len == 0)
 static FORCE_INLINE void dataColReset(SDataCol *pDataCol) { pDataCol->len = 0; }
 
 void dataColInit(SDataCol *pDataCol, STColumn *pCol, void **pBuf, int maxPoints);
-void dataColAppendVal(SDataCol *pCol, void *value, int numOfRows, int maxPoints);
+void dataColAppendVal(SDataCol *pCol, const void *value, int numOfRows, int maxPoints);
 void dataColSetOffset(SDataCol *pCol, int nEle);
 
 bool isNEleNull(SDataCol *pCol, int nEle);
 void dataColSetNEleNull(SDataCol *pCol, int nEle, int maxPoints);
 
 // Get the data pointer from a column-wised data
-static FORCE_INLINE void *tdGetColDataOfRow(SDataCol *pCol, int row) {
+static FORCE_INLINE const void *tdGetColDataOfRow(SDataCol *pCol, int row) {
+  if (isAllRowsNull(pCol)) {
+    return tdGetNullVal(pCol->type);
+  }
   if (IS_VAR_DATA_TYPE(pCol->type)) {
     return POINTER_SHIFT(pCol->pData, pCol->dataOff[row]);
   } else {
@@ -287,7 +320,7 @@ typedef struct {
 } SDataCols;
 
 #define keyCol(pCols) (&((pCols)->cols[0]))  // Key column
-#define dataColsTKeyAt(pCols, idx) ((TKEY *)(keyCol(pCols)->pData))[(idx)]
+#define dataColsTKeyAt(pCols, idx) ((TKEY *)(keyCol(pCols)->pData))[(idx)]  // the idx row of column-wised data
 #define dataColsKeyAt(pCols, idx) tdGetKey(dataColsTKeyAt(pCols, idx))
 static FORCE_INLINE TKEY dataColsTKeyFirst(SDataCols *pCols) {
   if (pCols->numOfRows) {
@@ -331,13 +364,13 @@ void       tdResetDataCols(SDataCols *pCols);
 int        tdInitDataCols(SDataCols *pCols, STSchema *pSchema);
 SDataCols *tdDupDataCols(SDataCols *pCols, bool keepData);
 SDataCols *tdFreeDataCols(SDataCols *pCols);
-void       tdAppendDataRowToDataCol(SDataRow row, STSchema *pSchema, SDataCols *pCols);
 int        tdMergeDataCols(SDataCols *target, SDataCols *source, int rowsToMerge, int *pOffset);
 
 // ----------------- K-V data row structure
-/*
+/* |<-------------------------------------- len -------------------------------------------->|
+ * |<----- header  ----->|<--------------------------- body -------------------------------->|
  * +----------+----------+---------------------------------+---------------------------------+
- * |  int16_t |  int16_t |                                 |                                 |
+ * | uint16_t |  int16_t |                                 |                                 |
  * +----------+----------+---------------------------------+---------------------------------+
  * |    len   |   ncols  |           cols index            |             data part           |
  * +----------+----------+---------------------------------+---------------------------------+
@@ -345,14 +378,14 @@ int        tdMergeDataCols(SDataCols *target, SDataCols *source, int rowsToMerge
 typedef void *SKVRow;
 
 typedef struct {
-  int16_t colId;
+  int16_t  colId;
-  int16_t offset;
+  uint16_t offset;
 } SColIdx;
 
-#define TD_KV_ROW_HEAD_SIZE (2 * sizeof(int16_t))
+#define TD_KV_ROW_HEAD_SIZE (sizeof(uint16_t) + sizeof(int16_t))
 
-#define kvRowLen(r) (*(int16_t *)(r))
+#define kvRowLen(r) (*(TDRowLenT *)(r))
-#define kvRowNCols(r) (*(int16_t *)POINTER_SHIFT(r, sizeof(int16_t)))
+#define kvRowNCols(r) (*(int16_t *)POINTER_SHIFT(r, sizeof(uint16_t)))
 #define kvRowSetLen(r, len) kvRowLen(r) = (len)
 #define kvRowSetNCols(r, n) kvRowNCols(r) = (n)
 #define kvRowColIdx(r) (SColIdx *)POINTER_SHIFT(r, TD_KV_ROW_HEAD_SIZE)
@@ -362,6 +395,9 @@ typedef struct {
 #define kvRowColIdxAt(r, i) (kvRowColIdx(r) + (i))
 #define kvRowFree(r) tfree(r)
 #define kvRowEnd(r) POINTER_SHIFT(r, kvRowLen(r))
+#define kvRowTKey(r) (*(TKEY *)(kvRowValues(r)))
+#define kvRowKey(r) tdGetKey(kvRowTKey(r))
+#define kvRowDeleted(r) TKEY_IS_DELETED(kvRowTKey(r))
 
 SKVRow tdKVRowDup(SKVRow row);
 int    tdSetKVRowDataOfCol(SKVRow *orow, int16_t colId, int8_t type, void *value);
@@ -385,13 +421,44 @@ static FORCE_INLINE void *tdGetKVRowValOfCol(SKVRow row, int16_t colId) {
   return kvRowColVal(row, (SColIdx *)ret);
 }
 
+static FORCE_INLINE void *tdGetKVRowIdxOfCol(SKVRow row, int16_t colId) {
+  return taosbsearch(&colId, kvRowColIdx(row), kvRowNCols(row), sizeof(SColIdx), comparTagId, TD_EQ);
+}
+
+// offset here not include kvRow header length
+static FORCE_INLINE int tdAppendKvColVal(SKVRow row, const void *value, int16_t colId, int8_t type, int32_t offset) {
+  ASSERT(value != NULL);
+  int32_t  toffset = offset + TD_KV_ROW_HEAD_SIZE;
+  SColIdx *pColIdx = (SColIdx *)POINTER_SHIFT(row, toffset);
+  char *   ptr = (char *)POINTER_SHIFT(row, kvRowLen(row));
+
+  pColIdx->colId = colId;
+  pColIdx->offset = kvRowLen(row);  // offset of pColIdx including the TD_KV_ROW_HEAD_SIZE
+
+  if (IS_VAR_DATA_TYPE(type)) {
+    memcpy(ptr, value, varDataTLen(value));
+    kvRowLen(row) += varDataTLen(value);
+  } else {
+    if (offset == 0) {
+      ASSERT(type == TSDB_DATA_TYPE_TIMESTAMP);
+      TKEY tvalue = tdGetTKEY(*(TSKEY *)value);
+      memcpy(ptr, (void *)(&tvalue), TYPE_BYTES[type]);
+    } else {
+      memcpy(ptr, value, TYPE_BYTES[type]);
+    }
+    kvRowLen(row) += TYPE_BYTES[type];
+  }
+
+  return 0;
+}
+
 // ----------------- K-V data row builder
 typedef struct {
   int16_t  tCols;
   int16_t  nCols;
   SColIdx *pColIdx;
-  int16_t  alloc;
+  uint16_t alloc;
-  int16_t  size;
+  uint16_t size;
   void *   buf;
 } SKVRowBuilder;
 
@@ -427,8 +494,146 @@ static FORCE_INLINE int tdAddColToKVRow(SKVRowBuilder *pBuilder, int16_t colId,
   return 0;
 }
 
+// ----------------- SMemRow appended with sequential data row structure
+/*
+ * |---------|------------------------------------------------- len ---------------------------------->|
+ * |<--------     Head      ------>|<---------   flen -------------->|                                 |
+ * |---------+---------------------+---------------------------------+---------------------------------+
+ * | uint8_t | uint16_t |  int16_t |                                 |                                 |
+ * |---------+----------+----------+---------------------------------+---------------------------------+
+ * |  flag   |   len    | sversion |           First part            |             Second part         |
+ * +---------+----------+----------+---------------------------------+---------------------------------+
+ *
+ * NOTE: timestamp in this row structure is TKEY instead of TSKEY
+ */
+
+// ----------------- SMemRow appended with extended K-V data row structure
+/* |--------------------|------------------------------------------------  len ---------------------------------->|
+ * |<-------------     Head      ------------>|<---------   flen -------------->|                                 |
+ * |--------------------+----------+--------------------------------------------+---------------------------------+
+ * | uint8_t | int16_t  | uint16_t |  int16_t |                                 |                                 |
+ * |---------+----------+----------+----------+---------------------------------+---------------------------------+
+ * |   flag  | sversion |   len    |   ncols  |           cols index            |             data part           |
+ * |---------+----------+----------+----------+---------------------------------+---------------------------------+
+ */
+
+typedef void *SMemRow;
+
+#define TD_MEM_ROW_TYPE_SIZE sizeof(uint8_t)
+#define TD_MEM_ROW_KV_VER_SIZE sizeof(int16_t)
+#define TD_MEM_ROW_KV_TYPE_VER_SIZE (TD_MEM_ROW_TYPE_SIZE + TD_MEM_ROW_KV_VER_SIZE)
+#define TD_MEM_ROW_DATA_HEAD_SIZE (TD_MEM_ROW_TYPE_SIZE + TD_DATA_ROW_HEAD_SIZE)
+// #define TD_MEM_ROW_KV_HEAD_SIZE (TD_MEM_ROW_TYPE_SIZE + TD_MEM_ROW_KV_VER_SIZE + TD_KV_ROW_HEAD_SIZE)
+
+#define SMEM_ROW_DATA 0U  // SDataRow
+#define SMEM_ROW_KV 1U    // SKVRow
+
+#define memRowType(r) (*(uint8_t *)(r))
+#define isDataRow(r) (SMEM_ROW_DATA == memRowType(r))
+#define isKvRow(r) (SMEM_ROW_KV == memRowType(r))
+
+#define memRowDataBody(r) POINTER_SHIFT(r, TD_MEM_ROW_TYPE_SIZE)  // section after flag
+#define memRowKvBody(r) \
+  POINTER_SHIFT(r, TD_MEM_ROW_KV_TYPE_VER_SIZE)  // section after flag + sversion as to reuse SKVRow
+
+#define memRowDataLen(r) (*(TDRowLenT *)memRowDataBody(r))  //  0~65535
+#define memRowKvLen(r) (*(TDRowLenT *)memRowKvBody(r))      //  0~65535
+
+#define memRowDataTLen(r) \
+  ((TDRowTLenT)(memRowDataLen(r) + TD_MEM_ROW_TYPE_SIZE))  // using uint32_t/int32_t to store the TLen
+
+#define memRowKvTLen(r) ((TDRowTLenT)(memRowKvLen(r) + TD_MEM_ROW_KV_TYPE_VER_SIZE))
+
+#define memRowLen(r) (isDataRow(r) ? memRowDataLen(r) : memRowKvLen(r))
+#define memRowTLen(r) (isDataRow(r) ? memRowDataTLen(r) : memRowKvTLen(r))  // using uint32_t/int32_t to store the TLen
+
+#define memRowDataVersion(r) dataRowVersion(memRowDataBody(r))
+#define memRowKvVersion(r) (*(int16_t *)POINTER_SHIFT(r, TD_MEM_ROW_TYPE_SIZE))
+#define memRowVersion(r) (isDataRow(r) ? memRowDataVersion(r) : memRowKvVersion(r))  // schema version
+#define memRowSetKvVersion(r, v) (memRowKvVersion(r) = (v))
+#define memRowTuple(r) (isDataRow(r) ? dataRowTuple(memRowDataBody(r)) : kvRowValues(memRowKvBody(r)))
+
+#define memRowTKey(r) (isDataRow(r) ? dataRowTKey(memRowDataBody(r)) : kvRowTKey(memRowKvBody(r)))
+#define memRowKey(r) (isDataRow(r) ? dataRowKey(memRowDataBody(r)) : kvRowKey(memRowKvBody(r)))
+#define memRowSetTKey(r, k)                 \
+  do {                                      \
+    if (isDataRow(r)) {                     \
+      dataRowTKey(memRowDataBody(r)) = (k); \
+    } else {                                \
+      kvRowTKey(memRowKvBody(r)) = (k);     \
+    }                                       \
+  } while (0)
+
+#define memRowSetType(r, t) (memRowType(r) = (t))
+#define memRowSetLen(r, l) (isDataRow(r) ? memRowDataLen(r) = (l) : memRowKvLen(r) = (l))
+#define memRowSetVersion(r, v) (isDataRow(r) ? dataRowSetVersion(memRowDataBody(r), v) : memRowKvSetVersion(r, v))
+#define memRowCpy(dst, r) memcpy((dst), (r), memRowTLen(r))
+#define memRowMaxBytesFromSchema(s) (schemaTLen(s) + TD_MEM_ROW_DATA_HEAD_SIZE)
+#define memRowDeleted(r) TKEY_IS_DELETED(memRowTKey(r))
+
+SMemRow tdMemRowDup(SMemRow row);
+void    tdAppendMemRowToDataCol(SMemRow row, STSchema *pSchema, SDataCols *pCols);
+// NOTE: offset here including the header size
+static FORCE_INLINE void *tdGetKvRowDataOfCol(void *row, int32_t offset) { return POINTER_SHIFT(row, offset); }
+// NOTE: offset here including the header size
+static FORCE_INLINE void *tdGetMemRowDataOfCol(void *row, int8_t type, int32_t offset) {
+  if (isDataRow(row)) {
+    return tdGetRowDataOfCol(row, type, offset);
+  } else if (isKvRow(row)) {
+    return tdGetKvRowDataOfCol(row, offset);
+  } else {
+    ASSERT(0);
+  }
+  return NULL;
+}
+
+// ----------------- Raw payload structure for row:
+/* |<------------ Head ------------->|<----------- body of column data tuple ------------------->|
+ * |                                 |<----------------- flen ------------->|<--- value part --->|
+ * |SMemRowType| dataTLen |  nCols   |  colId  | colType | offset   |  ...  | value |...|...|... |
+ * +-----------+----------+----------+--------------------------------------|--------------------|
+ * | uint8_t   | uint32_t | uint16_t | int16_t | uint8_t | uint16_t |  ...  |.......|...|...|... |
+ * +-----------+----------+----------+--------------------------------------+--------------------|
+ *  1. offset in column data tuple starts from the value part in case of uint16_t overflow.
+ *  2. dataTLen: total length including the header and body.
+ */
+
+#define PAYLOAD_NCOLS_LEN sizeof(uint16_t)
+#define PAYLOAD_NCOLS_OFFSET (sizeof(uint8_t) + sizeof(TDRowTLenT))
+#define PAYLOAD_HEADER_LEN (PAYLOAD_NCOLS_OFFSET + PAYLOAD_NCOLS_LEN)
+#define PAYLOAD_ID_LEN sizeof(int16_t)
+#define PAYLOAD_ID_TYPE_LEN (sizeof(int16_t) + sizeof(uint8_t))
+#define PAYLOAD_COL_HEAD_LEN (PAYLOAD_ID_TYPE_LEN + sizeof(uint16_t))
+#define PAYLOAD_PRIMARY_COL_LEN (PAYLOAD_ID_TYPE_LEN + sizeof(TSKEY))
+
+#define payloadBody(r) POINTER_SHIFT(r, PAYLOAD_HEADER_LEN)
+#define payloadType(r) (*(uint8_t *)(r))
+#define payloadSetType(r, t) (payloadType(r) = (t))
+#define payloadTLen(r) (*(TDRowTLenT *)POINTER_SHIFT(r, TD_MEM_ROW_TYPE_SIZE))  // including total header
+#define payloadSetTLen(r, l) (payloadTLen(r) = (l))
+#define payloadNCols(r) (*(TDRowLenT *)POINTER_SHIFT(r, PAYLOAD_NCOLS_OFFSET))
+#define payloadSetNCols(r, n) (payloadNCols(r) = (n))
+#define payloadValuesOffset(r) \
+  (PAYLOAD_HEADER_LEN + payloadNCols(r) * PAYLOAD_COL_HEAD_LEN)    // avoid using the macro in loop
+#define payloadValues(r) POINTER_SHIFT(r, payloadValuesOffset(r))  // avoid using the macro in loop
+#define payloadColId(c) (*(int16_t *)(c))
+#define payloadColType(c) (*(uint8_t *)POINTER_SHIFT(c, PAYLOAD_ID_LEN))
+#define payloadColOffset(c) (*(uint16_t *)POINTER_SHIFT(c, PAYLOAD_ID_TYPE_LEN))
+#define payloadColValue(c) POINTER_SHIFT(c, payloadColOffset(c))
+
+#define payloadColSetId(c, i) (payloadColId(c) = (i))
+#define payloadColSetType(c, t) (payloadColType(c) = (t))
+#define payloadColSetOffset(c, o) (payloadColOffset(c) = (o))
+
+#define payloadTSKey(r) (*(TSKEY *)POINTER_SHIFT(r, payloadValuesOffset(r)))
+#define payloadTKey(r) (*(TKEY *)POINTER_SHIFT(r, payloadValuesOffset(r)))
+#define payloadKey(r) tdGetKey(payloadTKey(r))
+
+
+static FORCE_INLINE char *payloadNextCol(char *pCol) { return (char *)POINTER_SHIFT(pCol, PAYLOAD_COL_HEAD_LEN); }
+
 #ifdef __cplusplus
 }
 #endif
 
-#endif  // _TD_DATA_FORMAT_H_
+#endif  // _TD_DATA_FORMAT_H_

src/common/src/tdataformat.c

@@ -18,6 +18,21 @@
 #include "tcoding.h"
 #include "wchar.h"
 
+const uint8_t      BoolNull = TSDB_DATA_BOOL_NULL;
+const uint8_t      TinyintNull = TSDB_DATA_TINYINT_NULL;
+const uint16_t     SmallintNull = TSDB_DATA_SMALLINT_NULL;
+const uint32_t     IntNull = TSDB_DATA_INT_NULL;
+const uint64_t     BigintNull = TSDB_DATA_BIGINT_NULL;
+const uint64_t     TimestampNull = TSDB_DATA_BIGINT_NULL;
+const uint8_t      UTinyintNull = TSDB_DATA_UTINYINT_NULL;
+const uint16_t     USmallintNull = TSDB_DATA_USMALLINT_NULL;
+const uint32_t     UIntNull = TSDB_DATA_UINT_NULL;
+const uint64_t     UBigintNull = TSDB_DATA_UBIGINT_NULL;
+const uint32_t     FloatNull = TSDB_DATA_FLOAT_NULL;
+const uint64_t     DoubleNull = TSDB_DATA_DOUBLE_NULL;
+const SBinaryNullT BinaryNull = {1, TSDB_DATA_BINARY_NULL};
+const SNCharNullT  NcharNull = {4, TSDB_DATA_NCHAR_NULL};
+
 static void tdMergeTwoDataCols(SDataCols *target, SDataCols *src1, int *iter1, int limit1, SDataCols *src2, int *iter2,
                                int limit2, int tRows);
 
@@ -198,6 +213,14 @@ SDataRow tdDataRowDup(SDataRow row) {
   return trow;
 }
 
+SMemRow tdMemRowDup(SMemRow row) {
+  SMemRow trow = malloc(memRowTLen(row));
+  if (trow == NULL) return NULL;
+
+  memRowCpy(trow, row);
+  return trow;
+}
+
 void dataColInit(SDataCol *pDataCol, STColumn *pCol, void **pBuf, int maxPoints) {
   pDataCol->type = colType(pCol);
   pDataCol->colId = colColId(pCol);
@@ -217,11 +240,22 @@ void dataColInit(SDataCol *pDataCol, STColumn *pCol, void **pBuf, int maxPoints)
     *pBuf = POINTER_SHIFT(*pBuf, pDataCol->spaceSize);
   }
 }
-
 // value from timestamp should be TKEY here instead of TSKEY
-void dataColAppendVal(SDataCol *pCol, void *value, int numOfRows, int maxPoints) {
+void dataColAppendVal(SDataCol *pCol, const void *value, int numOfRows, int maxPoints) {
   ASSERT(pCol != NULL && value != NULL);
 
+  if (isAllRowsNull(pCol)) {
+    if (isNull(value, pCol->type)) {
+      // all null value yet, just return
+      return;
+    }
+
+    if (numOfRows > 0) {
+      // Find the first not null value, fill all previouse values as NULL
+      dataColSetNEleNull(pCol, numOfRows, maxPoints);
+    }
+  }
+
   if (IS_VAR_DATA_TYPE(pCol->type)) {
     // set offset
     pCol->dataOff[numOfRows] = pCol->len;
@@ -243,7 +277,7 @@ bool isNEleNull(SDataCol *pCol, int nEle) {
   return true;
 }
 
-void dataColSetNullAt(SDataCol *pCol, int index) {
+FORCE_INLINE void dataColSetNullAt(SDataCol *pCol, int index) {
   if (IS_VAR_DATA_TYPE(pCol->type)) {
     pCol->dataOff[index] = pCol->len;
     char *ptr = POINTER_SHIFT(pCol->pData, pCol->len);
@@ -399,8 +433,7 @@ void tdResetDataCols(SDataCols *pCols) {
     }
   }
 }
-
+static void tdAppendDataRowToDataCol(SDataRow row, STSchema *pSchema, SDataCols *pCols) {
-void tdAppendDataRowToDataCol(SDataRow row, STSchema *pSchema, SDataCols *pCols) {
   ASSERT(pCols->numOfRows == 0 || dataColsKeyLast(pCols) < dataRowKey(row));
 
   int rcol = 0;
@@ -419,7 +452,8 @@ void tdAppendDataRowToDataCol(SDataRow row, STSchema *pSchema, SDataCols *pCols)
     while (dcol < pCols->numOfCols) {
       SDataCol *pDataCol = &(pCols->cols[dcol]);
       if (rcol >= schemaNCols(pSchema)) {
-        dataColSetNullAt(pDataCol, pCols->numOfRows);
+        // dataColSetNullAt(pDataCol, pCols->numOfRows);
+        dataColAppendVal(pDataCol, tdGetNullVal(pDataCol->type), pCols->numOfRows, pCols->maxPoints);
         dcol++;
         continue;
       }
@@ -433,7 +467,8 @@ void tdAppendDataRowToDataCol(SDataRow row, STSchema *pSchema, SDataCols *pCols)
       } else if (pRowCol->colId < pDataCol->colId) {
         rcol++;
       } else {
-        dataColSetNullAt(pDataCol, pCols->numOfRows);
+        // dataColSetNullAt(pDataCol, pCols->numOfRows);
+        dataColAppendVal(pDataCol, tdGetNullVal(pDataCol->type), pCols->numOfRows, pCols->maxPoints);
         dcol++;
       }
     }
@@ -441,6 +476,62 @@ void tdAppendDataRowToDataCol(SDataRow row, STSchema *pSchema, SDataCols *pCols)
   pCols->numOfRows++;
 }
 
+static void tdAppendKvRowToDataCol(SKVRow row, STSchema *pSchema, SDataCols *pCols) {
+  ASSERT(pCols->numOfRows == 0 || dataColsKeyLast(pCols) < kvRowKey(row));
+
+  int rcol = 0;
+  int dcol = 0;
+
+  if (kvRowDeleted(row)) {
+    for (; dcol < pCols->numOfCols; dcol++) {
+      SDataCol *pDataCol = &(pCols->cols[dcol]);
+      if (dcol == 0) {
+        dataColAppendVal(pDataCol, kvRowValues(row), pCols->numOfRows, pCols->maxPoints);
+      } else {
+        dataColSetNullAt(pDataCol, pCols->numOfRows);
+      }
+    }
+  } else {
+    int nRowCols = kvRowNCols(row);
+
+    while (dcol < pCols->numOfCols) {
+      SDataCol *pDataCol = &(pCols->cols[dcol]);
+      if (rcol >= nRowCols || rcol >= schemaNCols(pSchema)) {
+        // dataColSetNullAt(pDataCol, pCols->numOfRows);
+        dataColAppendVal(pDataCol, tdGetNullVal(pDataCol->type), pCols->numOfRows, pCols->maxPoints);
+        ++dcol;
+        continue;
+      }
+
+      SColIdx *colIdx = kvRowColIdxAt(row, rcol);
+
+      if (colIdx->colId == pDataCol->colId) {
+        void *value = tdGetKvRowDataOfCol(row, colIdx->offset);
+        dataColAppendVal(pDataCol, value, pCols->numOfRows, pCols->maxPoints);
+        ++dcol;
+        ++rcol;
+      } else if (colIdx->colId < pDataCol->colId) {
+        ++rcol;
+      } else {
+        // dataColSetNullAt(pDataCol, pCols->numOfRows);
+        dataColAppendVal(pDataCol, tdGetNullVal(pDataCol->type), pCols->numOfRows, pCols->maxPoints);
+        ++dcol;
+      }
+    }
+  }
+  pCols->numOfRows++;
+}
+
+void tdAppendMemRowToDataCol(SMemRow row, STSchema *pSchema, SDataCols *pCols) {
+  if (isDataRow(row)) {
+    tdAppendDataRowToDataCol(memRowDataBody(row), pSchema, pCols);
+  } else if (isKvRow(row)) {
+    tdAppendKvRowToDataCol(memRowKvBody(row), pSchema, pCols);
+  } else {
+    ASSERT(0);
+  }
+}
+
 int tdMergeDataCols(SDataCols *target, SDataCols *source, int rowsToMerge, int *pOffset) {
   ASSERT(rowsToMerge > 0 && rowsToMerge <= source->numOfRows);
   ASSERT(target->numOfCols == source->numOfCols);
@@ -563,7 +654,7 @@ int tdSetKVRowDataOfCol(SKVRow *orow, int16_t colId, int8_t type, void *value) {
     nrow = malloc(kvRowLen(row) + sizeof(SColIdx) + diff);
     if (nrow == NULL) return -1;
 
-    kvRowSetLen(nrow, kvRowLen(row) + (int16_t)sizeof(SColIdx) + diff);
+    kvRowSetLen(nrow, kvRowLen(row) + (uint16_t)sizeof(SColIdx) + diff);
     kvRowSetNCols(nrow, kvRowNCols(row) + 1);
 
     if (ptr == NULL) {
@@ -605,8 +696,8 @@ int tdSetKVRowDataOfCol(SKVRow *orow, int16_t colId, int8_t type, void *value) {
       if (varDataTLen(value) == varDataTLen(pOldVal)) { // just update the column value in place
         memcpy(pOldVal, value, varDataTLen(value));
       } else { // need to reallocate the memory
-        int16_t diff = varDataTLen(value) - varDataTLen(pOldVal);
+        uint16_t diff = varDataTLen(value) - varDataTLen(pOldVal);
-        int16_t nlen = kvRowLen(row) + diff;
+        uint16_t nlen = kvRowLen(row) + diff;
         ASSERT(nlen > 0);
         nrow = malloc(nlen);
         if (nrow == NULL) return -1;
@@ -709,3 +800,39 @@ SKVRow tdGetKVRowFromBuilder(SKVRowBuilder *pBuilder) {
 
   return row;
 }
+
+const void *tdGetNullVal(int8_t type) {
+  switch (type) {
+    case TSDB_DATA_TYPE_BOOL:
+      return &BoolNull;
+    case TSDB_DATA_TYPE_TINYINT:
+      return &TinyintNull;
+    case TSDB_DATA_TYPE_SMALLINT:
+      return &SmallintNull;
+    case TSDB_DATA_TYPE_INT:
+      return &IntNull;
+    case TSDB_DATA_TYPE_BIGINT:
+      return &BigintNull;
+    case TSDB_DATA_TYPE_FLOAT:
+      return &FloatNull;
+    case TSDB_DATA_TYPE_DOUBLE:
+      return &DoubleNull;
+    case TSDB_DATA_TYPE_BINARY:
+      return &BinaryNull;
+    case TSDB_DATA_TYPE_TIMESTAMP:
+      return &TimestampNull;
+    case TSDB_DATA_TYPE_NCHAR:
+      return &NcharNull;
+    case TSDB_DATA_TYPE_UTINYINT:
+      return &UTinyintNull;
+    case TSDB_DATA_TYPE_USMALLINT:
+      return &USmallintNull;
+    case TSDB_DATA_TYPE_UINT:
+      return &UIntNull;
+    case TSDB_DATA_TYPE_UBIGINT:
+      return &UBigintNull;
+    default:
+      ASSERT(0);
+      return NULL;
+  }
+}

src/cq/src/cqMain.c

@@ -476,15 +476,17 @@ static void cqProcessStreamRes(void *param, TAOS_RES *tres, TAOS_ROW row) {
   
   cDebug("vgId:%d, id:%d CQ:%s stream result is ready", pContext->vgId, pObj->tid, pObj->sqlStr);
 
-  int32_t size = sizeof(SWalHead) + sizeof(SSubmitMsg) + sizeof(SSubmitBlk) + TD_DATA_ROW_HEAD_SIZE + pObj->rowSize;
+  int32_t size = sizeof(SWalHead) + sizeof(SSubmitMsg) + sizeof(SSubmitBlk) + TD_MEM_ROW_DATA_HEAD_SIZE + pObj->rowSize;
   char *buffer = calloc(size, 1);
 
   SWalHead   *pHead = (SWalHead *)buffer;
   SSubmitMsg *pMsg = (SSubmitMsg *) (buffer + sizeof(SWalHead));
   SSubmitBlk *pBlk = (SSubmitBlk *) (buffer + sizeof(SWalHead) + sizeof(SSubmitMsg));
 
-  SDataRow trow = (SDataRow)pBlk->data;
+  SMemRow trow = (SMemRow)pBlk->data;
-  tdInitDataRow(trow, pSchema);
+  SDataRow dataRow = (SDataRow)memRowDataBody(trow);
+  memRowSetType(trow, SMEM_ROW_DATA);
+  tdInitDataRow(dataRow, pSchema);
 
   for (int32_t i = 0; i < pSchema->numOfCols; i++) {
     STColumn *c = pSchema->columns + i;
@@ -500,9 +502,9 @@ static void cqProcessStreamRes(void *param, TAOS_RES *tres, TAOS_ROW row) {
       memcpy((char *)val + sizeof(VarDataLenT), buf, len);
       varDataLen(val) = len;
     }
-    tdAppendColVal(trow, val, c->type, c->bytes, c->offset);
+    tdAppendColVal(dataRow, val, c->type, c->offset);
   }
-  pBlk->dataLen = htonl(dataRowLen(trow));
+  pBlk->dataLen = htonl(memRowDataTLen(trow));
   pBlk->schemaLen = 0;
 
   pBlk->uid = htobe64(pObj->uid);
@@ -511,7 +513,7 @@ static void cqProcessStreamRes(void *param, TAOS_RES *tres, TAOS_ROW row) {
   pBlk->sversion = htonl(pSchema->version);
   pBlk->padding = 0;
 
-  pHead->len = sizeof(SSubmitMsg) + sizeof(SSubmitBlk) + dataRowLen(trow);
+  pHead->len = sizeof(SSubmitMsg) + sizeof(SSubmitBlk) + memRowDataTLen(trow);
 
   pMsg->header.vgId = htonl(pContext->vgId);
   pMsg->header.contLen = htonl(pHead->len);

src/inc/taosdef.h

@@ -180,7 +180,7 @@ do { \
 // this is the length of its string representation, including the terminator zero
 #define TSDB_ACCT_ID_LEN          11
 
-#define TSDB_MAX_COLUMNS          1024
+#define TSDB_MAX_COLUMNS          4096
 #define TSDB_MIN_COLUMNS          2       //PRIMARY COLUMN(timestamp) + other columns
 
 #define TSDB_NODE_NAME_LEN        64
@@ -199,7 +199,13 @@ do { \
 
 #define TSDB_APPNAME_LEN          TSDB_UNI_LEN
 
-#define TSDB_MAX_BYTES_PER_ROW    16384
+  /**
+   *  In some scenarios uint16_t (0~65535) is used to store the row len.
+   *  - Firstly, we use 65531(65535 - 4), as the SDataRow/SKVRow contains 4 bits header.
+   *  - Secondly, if all cols are VarDataT type except primary key, we need 4 bits to store the offset, thus
+   *    the final value is 65531-(4096-1)*4 = 49151.
+   */
+#define TSDB_MAX_BYTES_PER_ROW    49151
 #define TSDB_MAX_TAGS_LEN         16384
 #define TSDB_MAX_TAGS             128
 #define TSDB_MAX_TAG_CONDITIONS   1024
@@ -327,8 +333,9 @@ do { \
 #define TSDB_MAX_JOIN_TABLE_NUM         10
 #define TSDB_MAX_UNION_CLAUSE           5
 
-#define TSDB_MAX_BINARY_LEN            (TSDB_MAX_BYTES_PER_ROW-TSDB_KEYSIZE)
+#define TSDB_MAX_FIELD_LEN              16384
-#define TSDB_MAX_NCHAR_LEN             (TSDB_MAX_BYTES_PER_ROW-TSDB_KEYSIZE)
+#define TSDB_MAX_BINARY_LEN            (TSDB_MAX_FIELD_LEN-TSDB_KEYSIZE) // keep 16384
+#define TSDB_MAX_NCHAR_LEN             (TSDB_MAX_FIELD_LEN-TSDB_KEYSIZE) // keep 16384
 #define PRIMARYKEY_TIMESTAMP_COL_INDEX  0
 
 #define TSDB_MAX_RPC_THREADS            5

src/inc/ttype.h

@@ -10,8 +10,10 @@ extern "C" {
 #include "taosdef.h"
 
 // ----------------- For variable data types such as TSDB_DATA_TYPE_BINARY and TSDB_DATA_TYPE_NCHAR
-typedef int32_t VarDataOffsetT;
+typedef int32_t  VarDataOffsetT;
-typedef int16_t VarDataLenT;
+typedef int16_t  VarDataLenT;  // maxVarDataLen: 32767
+typedef uint16_t TDRowLenT;    // not including overhead: 0 ~ 65535
+typedef uint32_t TDRowTLenT;   // total length, including overhead
 
 typedef struct tstr {
   VarDataLenT len;

src/query/inc/qExtbuffer.h

@@ -77,13 +77,13 @@ typedef struct tFilePagesItem {
 
 typedef struct SSchemaEx {
   struct SSchema field;
-  int16_t        offset;
+  int32_t        offset;
 } SSchemaEx;
 
 typedef struct SColumnModel {
   int32_t    capacity;
   int32_t    numOfCols;
-  int16_t    rowSize;
+  int32_t    rowSize;
   SSchemaEx *pFields;
 } SColumnModel;
 

src/query/inc/qUtil.h

@@ -66,7 +66,8 @@ static FORCE_INLINE SResultRow *getResultRow(SResultRowInfo *pResultRowInfo, int
   return pResultRowInfo->pResult[slot];
 }
 
-static FORCE_INLINE char *getPosInResultPage(SQueryAttr *pQueryAttr, tFilePage* page, int32_t rowOffset, int16_t offset) {
+static FORCE_INLINE char* getPosInResultPage(SQueryAttr* pQueryAttr, tFilePage* page, int32_t rowOffset,
+                                             int32_t offset) {
   assert(rowOffset >= 0 && pQueryAttr != NULL);
 
   int32_t numOfRows = (int32_t)GET_ROW_PARAM_FOR_MULTIOUTPUT(pQueryAttr, pQueryAttr->topBotQuery, pQueryAttr->stableQuery);

src/query/src/qAggMain.c

@@ -3982,7 +3982,7 @@ void blockInfo_func(SQLFunctionCtx* pCtx) {
 
   int32_t len = *(int32_t*) pCtx->pInput;
   blockDistInfoFromBinary((char*)pCtx->pInput + sizeof(int32_t), len, pDist);
-  pDist->rowSize = (int16_t) pCtx->param[0].i64;
+  pDist->rowSize = (uint16_t)pCtx->param[0].i64;
 
   memcpy(pCtx->pOutput, pCtx->pInput, sizeof(int32_t) + len);
 
@@ -4129,7 +4129,7 @@ void blockinfo_func_finalizer(SQLFunctionCtx* pCtx) {
   SResultRowCellInfo *pResInfo = GET_RES_INFO(pCtx);
   STableBlockDist* pDist = (STableBlockDist*) GET_ROWCELL_INTERBUF(pResInfo);
 
-  pDist->rowSize = (int16_t)pCtx->param[0].i64;
+  pDist->rowSize = (uint16_t)pCtx->param[0].i64;
   generateBlockDistResult(pDist, pCtx->pOutput);
 
   // cannot set the numOfIteratedElems again since it is set during previous iteration

src/query/src/qExecutor.c

@@ -3583,7 +3583,7 @@ void setResultRowOutputBufInitCtx(SQueryRuntimeEnv *pRuntimeEnv, SResultRow *pRe
   // Note: pResult->pos[i]->num == 0, there is only fixed number of results for each group
   tFilePage* bufPage = getResBufPage(pRuntimeEnv->pResultBuf, pResult->pageId);
 
-  int16_t offset = 0;
+  int32_t offset = 0;
   for (int32_t i = 0; i < numOfOutput; ++i) {
     pCtx[i].resultInfo = getResultCell(pResult, i, rowCellInfoOffset);
 
@@ -3897,7 +3897,7 @@ static int32_t doCopyToSDataBlock(SQueryRuntimeEnv* pRuntimeEnv, SGroupResInfo*
 
     tFilePage *page = getResBufPage(pRuntimeEnv->pResultBuf, pRow->pageId);
 
-    int16_t offset = 0;
+    int32_t offset = 0;
     for (int32_t j = 0; j < pBlock->info.numOfCols; ++j) {
       SColumnInfoData* pColInfoData = taosArrayGet(pBlock->pDataBlock, j);
       int32_t bytes = pColInfoData->info.bytes;

src/tsdb/inc/tsdbMemTable.h

@@ -71,27 +71,27 @@ int   tsdbLoadDataFromCache(STable* pTable, SSkipListIterator* pIter, TSKEY maxK
                             TKEY* filterKeys, int nFilterKeys, bool keepDup, SMergeInfo* pMergeInfo);
 void* tsdbCommitData(STsdbRepo* pRepo);
 
-static FORCE_INLINE SDataRow tsdbNextIterRow(SSkipListIterator* pIter) {
+static FORCE_INLINE SMemRow tsdbNextIterRow(SSkipListIterator* pIter) {
   if (pIter == NULL) return NULL;
 
   SSkipListNode* node = tSkipListIterGet(pIter);
   if (node == NULL) return NULL;
 
-  return (SDataRow)SL_GET_NODE_DATA(node);
+  return (SMemRow)SL_GET_NODE_DATA(node);
 }
 
 static FORCE_INLINE TSKEY tsdbNextIterKey(SSkipListIterator* pIter) {
-  SDataRow row = tsdbNextIterRow(pIter);
+  SMemRow row = tsdbNextIterRow(pIter);
   if (row == NULL) return TSDB_DATA_TIMESTAMP_NULL;
 
-  return dataRowKey(row);
+  return memRowKey(row);
 }
 
 static FORCE_INLINE TKEY tsdbNextIterTKey(SSkipListIterator* pIter) {
-  SDataRow row = tsdbNextIterRow(pIter);
+  SMemRow row = tsdbNextIterRow(pIter);
   if (row == NULL) return TKEY_NULL;
 
-  return dataRowTKey(row);
+  return memRowTKey(row);
 }
 
 #endif /* _TD_TSDB_MEMTABLE_H_ */

src/tsdb/inc/tsdbMeta.h

@@ -32,7 +32,7 @@ typedef struct STable {
   void*          eventHandler;   // TODO
   void*          streamHandler;  // TODO
   TSKEY          lastKey;
-  SDataRow       lastRow;
+  SMemRow        lastRow;
   char*          sql;
   void*          cqhandle;
   SRWLatch       latch;  // TODO: implementa latch functions
@@ -148,7 +148,7 @@ static FORCE_INLINE STSchema *tsdbGetTableTagSchema(STable *pTable) {
 }
 
 static FORCE_INLINE TSKEY tsdbGetTableLastKeyImpl(STable* pTable) {
-  ASSERT(pTable->lastRow == NULL || pTable->lastKey == dataRowKey(pTable->lastRow));
+  ASSERT((pTable->lastRow == NULL) || (pTable->lastKey == memRowKey(pTable->lastRow)));
   return pTable->lastKey;
 }
 

src/tsdb/src/tsdbCommit.c

@@ -920,7 +920,8 @@ int tsdbWriteBlockImpl(STsdbRepo *pRepo, STable *pTable, SDFile *pDFile, SDataCo
     SDataCol * pDataCol = pDataCols->cols + ncol;
     SBlockCol *pBlockCol = pBlockData->cols + nColsNotAllNull;
 
-    if (isNEleNull(pDataCol, rowsToWrite)) {  // all data to commit are NULL, just ignore it
+    // if (isNEleNull(pDataCol, rowsToWrite)) {  // all data to commit are NULL, just ignore it
+    if (isAllRowsNull(pDataCol)) {  // all data to commit are NULL, just ignore it
       continue;
     }
 
@@ -1264,12 +1265,12 @@ static void tsdbLoadAndMergeFromCache(SDataCols *pDataCols, int *iter, SCommitIt
   while (true) {
     key1 = (*iter >= pDataCols->numOfRows) ? INT64_MAX : dataColsKeyAt(pDataCols, *iter);
     bool isRowDel = false;
-    SDataRow row = tsdbNextIterRow(pCommitIter->pIter);
+    SMemRow row = tsdbNextIterRow(pCommitIter->pIter);
-    if (row == NULL || dataRowKey(row) > maxKey) {
+    if (row == NULL || memRowKey(row) > maxKey) {
       key2 = INT64_MAX;
     } else {
-      key2 = dataRowKey(row);
+      key2 = memRowKey(row);
-      isRowDel = dataRowDeleted(row);
+      isRowDel = memRowDeleted(row);
     }
 
     if (key1 == INT64_MAX && key2 == INT64_MAX) break;
@@ -1284,24 +1285,24 @@ static void tsdbLoadAndMergeFromCache(SDataCols *pDataCols, int *iter, SCommitIt
       (*iter)++;
     } else if (key1 > key2) {
       if (!isRowDel) {
-        if (pSchema == NULL || schemaVersion(pSchema) != dataRowVersion(row)) {
+        if (pSchema == NULL || schemaVersion(pSchema) != memRowVersion(row)) {
-          pSchema = tsdbGetTableSchemaImpl(pCommitIter->pTable, false, false, dataRowVersion(row));
+          pSchema = tsdbGetTableSchemaImpl(pCommitIter->pTable, false, false, memRowVersion(row));
           ASSERT(pSchema != NULL);
         }
 
-        tdAppendDataRowToDataCol(row, pSchema, pTarget);
+        tdAppendMemRowToDataCol(row, pSchema, pTarget);
       }
 
       tSkipListIterNext(pCommitIter->pIter);
     } else {
       if (update) {
         if (!isRowDel) {
-          if (pSchema == NULL || schemaVersion(pSchema) != dataRowVersion(row)) {
+          if (pSchema == NULL || schemaVersion(pSchema) != memRowVersion(row)) {
-            pSchema = tsdbGetTableSchemaImpl(pCommitIter->pTable, false, false, dataRowVersion(row));
+            pSchema = tsdbGetTableSchemaImpl(pCommitIter->pTable, false, false, memRowVersion(row));
             ASSERT(pSchema != NULL);
           }
 
-          tdAppendDataRowToDataCol(row, pSchema, pTarget);
+          tdAppendMemRowToDataCol(row, pSchema, pTarget);
         }
       } else {
         ASSERT(!isRowDel);

src/tsdb/src/tsdbMain.c

@@ -641,7 +641,7 @@ static int tsdbRestoreLastColumns(STsdbRepo *pRepo, STable *pTable, SReadH* pRea
   int numColumns;
   int32_t blockIdx;
   SDataStatis* pBlockStatis = NULL;
-  SDataRow row = NULL;
+  SMemRow      row = NULL;
   // restore last column data with last schema
   
   int err = 0;
@@ -657,13 +657,15 @@ static int tsdbRestoreLastColumns(STsdbRepo *pRepo, STable *pTable, SReadH* pRea
     }
   }
 
-  row = taosTMalloc(dataRowMaxBytesFromSchema(pSchema));
+  row = taosTMalloc(memRowMaxBytesFromSchema(pSchema));
   if (row == NULL) {
     terrno = TSDB_CODE_TDB_OUT_OF_MEMORY;
     err = -1;
     goto out;
   }
-  tdInitDataRow(row, pSchema);
+
+  memRowSetType(row, SMEM_ROW_DATA);
+  tdInitDataRow(memRowDataBody(row), pSchema);
 
   // first load block index info
   if (tsdbLoadBlockInfo(pReadh, NULL) < 0) {
@@ -720,9 +722,9 @@ static int tsdbRestoreLastColumns(STsdbRepo *pRepo, STable *pTable, SReadH* pRea
       // OK,let's load row from backward to get not-null column
       for (int32_t rowId = pBlock->numOfRows - 1; rowId >= 0; rowId--) {
         SDataCol *pDataCol = pReadh->pDCols[0]->cols + i;
-        tdAppendColVal(row, tdGetColDataOfRow(pDataCol, rowId), pCol->type, pCol->bytes, pCol->offset);
+        tdAppendColVal(memRowDataBody(row), tdGetColDataOfRow(pDataCol, rowId), pCol->type, pCol->offset);
         //SDataCol *pDataCol = readh.pDCols[0]->cols + j;
-        void* value = tdGetRowDataOfCol(row, (int8_t)pCol->type, TD_DATA_ROW_HEAD_SIZE + pCol->offset);
+        void *value = tdGetRowDataOfCol(memRowDataBody(row), (int8_t)pCol->type, TD_DATA_ROW_HEAD_SIZE + pCol->offset);
         if (isNull(value, pCol->type)) {
           continue;
         }
@@ -742,8 +744,8 @@ static int tsdbRestoreLastColumns(STsdbRepo *pRepo, STable *pTable, SReadH* pRea
         // save row ts(in column 0)
         pDataCol = pReadh->pDCols[0]->cols + 0;
         pCol = schemaColAt(pSchema, 0);
-        tdAppendColVal(row, tdGetColDataOfRow(pDataCol, rowId), pCol->type, pCol->bytes, pCol->offset);
+        tdAppendColVal(memRowDataBody(row), tdGetColDataOfRow(pDataCol, rowId), pCol->type, pCol->offset);
-        pLastCol->ts = dataRowKey(row);
+        pLastCol->ts = memRowKey(row);
 
         pTable->restoreColumnNum += 1;
 
@@ -779,18 +781,18 @@ static int tsdbRestoreLastRow(STsdbRepo *pRepo, STable *pTable, SReadH* pReadh,
   // Get the data in row
   
   STSchema *pSchema = tsdbGetTableSchema(pTable);
-  pTable->lastRow = taosTMalloc(dataRowMaxBytesFromSchema(pSchema));
+  pTable->lastRow = taosTMalloc(memRowMaxBytesFromSchema(pSchema));
   if (pTable->lastRow == NULL) {
     terrno = TSDB_CODE_TDB_OUT_OF_MEMORY;
     return -1;
   }
-
+  memRowSetType(pTable->lastRow, SMEM_ROW_DATA);
-  tdInitDataRow(pTable->lastRow, pSchema);
+  tdInitDataRow(memRowDataBody(pTable->lastRow), pSchema);
   for (int icol = 0; icol < schemaNCols(pSchema); icol++) {
     STColumn *pCol = schemaColAt(pSchema, icol);
     SDataCol *pDataCol = pReadh->pDCols[0]->cols + icol;
-    tdAppendColVal(pTable->lastRow, tdGetColDataOfRow(pDataCol, pBlock->numOfRows - 1), pCol->type, pCol->bytes,
+    tdAppendColVal(memRowDataBody(pTable->lastRow), tdGetColDataOfRow(pDataCol, pBlock->numOfRows - 1), pCol->type,
-                    pCol->offset);
+                   pCol->offset);
   }
 
   return 0;

src/tsdb/src/tsdbMemTable.c

@@ -21,7 +21,7 @@
 typedef struct {
   int32_t  totalLen;
   int32_t  len;
-  SDataRow row;
+  SMemRow  row;
 } SSubmitBlkIter;
 
 typedef struct {
@@ -36,20 +36,19 @@ static STableData *tsdbNewTableData(STsdbCfg *pCfg, STable *pTable);
 static void        tsdbFreeTableData(STableData *pTableData);
 static char *      tsdbGetTsTupleKey(const void *data);
 static int          tsdbAdjustMemMaxTables(SMemTable *pMemTable, int maxTables);
-static int          tsdbAppendTableRowToCols(STable *pTable, SDataCols *pCols, STSchema **ppSchema, SDataRow row);
+static int              tsdbAppendTableRowToCols(STable *pTable, SDataCols *pCols, STSchema **ppSchema, SMemRow row);
 static int          tsdbInitSubmitBlkIter(SSubmitBlk *pBlock, SSubmitBlkIter *pIter);
-static SDataRow     tsdbGetSubmitBlkNext(SSubmitBlkIter *pIter);
+static SMemRow          tsdbGetSubmitBlkNext(SSubmitBlkIter *pIter);
 static int          tsdbScanAndConvertSubmitMsg(STsdbRepo *pRepo, SSubmitMsg *pMsg);
 static int          tsdbInsertDataToTable(STsdbRepo *pRepo, SSubmitBlk *pBlock, int32_t *affectedrows);
-static int          tsdbCopyRowToMem(STsdbRepo *pRepo, SDataRow row, STable *pTable, void **ppRow);
+static int              tsdbCopyRowToMem(STsdbRepo *pRepo, SMemRow row, STable *pTable, void **ppRow);
 static int          tsdbInitSubmitMsgIter(SSubmitMsg *pMsg, SSubmitMsgIter *pIter);
 static int          tsdbGetSubmitMsgNext(SSubmitMsgIter *pIter, SSubmitBlk **pPBlock);
 static int          tsdbCheckTableSchema(STsdbRepo *pRepo, SSubmitBlk *pBlock, STable *pTable);
 static int          tsdbInsertDataToTableImpl(STsdbRepo *pRepo, STable *pTable, void **rows, int rowCounter);
 static void         tsdbFreeRows(STsdbRepo *pRepo, void **rows, int rowCounter);
-static int          tsdbUpdateTableLatestInfo(STsdbRepo *pRepo, STable *pTable, SDataRow row);
+static int              tsdbUpdateTableLatestInfo(STsdbRepo *pRepo, STable *pTable, SMemRow row);
-
+static FORCE_INLINE int tsdbCheckRowRange(STsdbRepo *pRepo, STable *pTable, SMemRow row, TSKEY minKey, TSKEY maxKey,
-static FORCE_INLINE int tsdbCheckRowRange(STsdbRepo *pRepo, STable *pTable, SDataRow row, TSKEY minKey, TSKEY maxKey,
                                           TSKEY now);
 
 int32_t tsdbInsertData(STsdbRepo *repo, SSubmitMsg *pMsg, SShellSubmitRspMsg *pRsp) {
@@ -354,7 +353,7 @@ int tsdbLoadDataFromCache(STable *pTable, SSkipListIterator *pIter, TSKEY maxKey
   TSKEY      fKey = 0;
   bool       isRowDel = false;
   int        filterIter = 0;
-  SDataRow   row = NULL;
+  SMemRow    row = NULL;
   SMergeInfo mInfo;
 
   if (pMergeInfo == NULL) pMergeInfo = &mInfo;
@@ -365,12 +364,12 @@ int tsdbLoadDataFromCache(STable *pTable, SSkipListIterator *pIter, TSKEY maxKey
   if (pCols) tdResetDataCols(pCols);
 
   row = tsdbNextIterRow(pIter);
-  if (row == NULL || dataRowKey(row) > maxKey) {
+  if (row == NULL || memRowKey(row) > maxKey) {
     rowKey = INT64_MAX;
     isRowDel = false;
   } else {
-    rowKey = dataRowKey(row);
+    rowKey = memRowKey(row);
-    isRowDel = dataRowDeleted(row);
+    isRowDel = memRowDeleted(row);
   }
 
   if (filterIter >= nFilterKeys) {
@@ -407,12 +406,12 @@ int tsdbLoadDataFromCache(STable *pTable, SSkipListIterator *pIter, TSKEY maxKey
 
       tSkipListIterNext(pIter);
       row = tsdbNextIterRow(pIter);
-      if (row == NULL || dataRowKey(row) > maxKey) {
+      if (row == NULL || memRowKey(row) > maxKey) {
         rowKey = INT64_MAX;
         isRowDel = false;
       } else {
-        rowKey = dataRowKey(row);
+        rowKey = memRowKey(row);
-        isRowDel = dataRowDeleted(row);
+        isRowDel = memRowDeleted(row);
       }
     } else {
       if (isRowDel) {
@@ -437,12 +436,12 @@ int tsdbLoadDataFromCache(STable *pTable, SSkipListIterator *pIter, TSKEY maxKey
 
       tSkipListIterNext(pIter);
       row = tsdbNextIterRow(pIter);
-      if (row == NULL || dataRowKey(row) > maxKey) {
+      if (row == NULL || memRowKey(row) > maxKey) {
         rowKey = INT64_MAX;
         isRowDel = false;
       } else {
-        rowKey = dataRowKey(row);
+        rowKey = memRowKey(row);
-        isRowDel = dataRowDeleted(row);
+        isRowDel = memRowDeleted(row);
       }
 
       filterIter++;
@@ -548,7 +547,7 @@ static void tsdbFreeTableData(STableData *pTableData) {
   }
 }
 
-static char *tsdbGetTsTupleKey(const void *data) { return dataRowTuple((SDataRow)data); }
+static char *tsdbGetTsTupleKey(const void *data) { return memRowTuple((SMemRow)data); }
 
 static int tsdbAdjustMemMaxTables(SMemTable *pMemTable, int maxTables) {
   ASSERT(pMemTable->maxTables < maxTables);
@@ -572,17 +571,17 @@ static int tsdbAdjustMemMaxTables(SMemTable *pMemTable, int maxTables) {
   return 0;
 }
 
-static int tsdbAppendTableRowToCols(STable *pTable, SDataCols *pCols, STSchema **ppSchema, SDataRow row) {
+static int tsdbAppendTableRowToCols(STable *pTable, SDataCols *pCols, STSchema **ppSchema, SMemRow row) {
   if (pCols) {
-    if (*ppSchema == NULL || schemaVersion(*ppSchema) != dataRowVersion(row)) {
+    if (*ppSchema == NULL || schemaVersion(*ppSchema) != memRowVersion(row)) {
-      *ppSchema = tsdbGetTableSchemaImpl(pTable, false, false, dataRowVersion(row));
+      *ppSchema = tsdbGetTableSchemaImpl(pTable, false, false, memRowVersion(row));
       if (*ppSchema == NULL) {
         ASSERT(false);
         return -1;
       }
     }
 
-    tdAppendDataRowToDataCol(row, *ppSchema, pCols);
+    tdAppendMemRowToDataCol(row, *ppSchema, pCols);
   }
 
   return 0;
@@ -592,31 +591,32 @@ static int tsdbInitSubmitBlkIter(SSubmitBlk *pBlock, SSubmitBlkIter *pIter) {
   if (pBlock->dataLen <= 0) return -1;
   pIter->totalLen = pBlock->dataLen;
   pIter->len = 0;
-  pIter->row = (SDataRow)(pBlock->data+pBlock->schemaLen);
+  pIter->row = (SMemRow)(pBlock->data + pBlock->schemaLen);
   return 0;
 }
 
-static SDataRow tsdbGetSubmitBlkNext(SSubmitBlkIter *pIter) {
+static SMemRow tsdbGetSubmitBlkNext(SSubmitBlkIter *pIter) {
-  SDataRow row = pIter->row;
+  SMemRow row = pIter->row;  // firstly, get current row
   if (row == NULL) return NULL;
 
-  pIter->len += dataRowLen(row);
+  pIter->len += memRowTLen(row);
-  if (pIter->len >= pIter->totalLen) {
+  if (pIter->len >= pIter->totalLen) {  // reach the end
     pIter->row = NULL;
   } else {
-    pIter->row = (char *)row + dataRowLen(row);
+    pIter->row = (char *)row + memRowTLen(row);  // secondly, move to next row
   }
 
   return row;
 }
 
-static FORCE_INLINE int tsdbCheckRowRange(STsdbRepo *pRepo, STable *pTable, SDataRow row, TSKEY minKey, TSKEY maxKey,
+static FORCE_INLINE int tsdbCheckRowRange(STsdbRepo *pRepo, STable *pTable, SMemRow row, TSKEY minKey, TSKEY maxKey,
                                           TSKEY now) {
-  if (dataRowKey(row) < minKey || dataRowKey(row) > maxKey) {
+  TSKEY rowKey = memRowKey(row);
+  if (rowKey < minKey || rowKey > maxKey) {
     tsdbError("vgId:%d table %s tid %d uid %" PRIu64 " timestamp is out of range! now %" PRId64 " minKey %" PRId64
               " maxKey %" PRId64 " row key %" PRId64,
               REPO_ID(pRepo), TABLE_CHAR_NAME(pTable), TABLE_TID(pTable), TABLE_UID(pTable), now, minKey, maxKey,
-              dataRowKey(row));
+              rowKey);
     terrno = TSDB_CODE_TDB_TIMESTAMP_OUT_OF_RANGE;
     return -1;
   }
@@ -630,7 +630,7 @@ static int tsdbScanAndConvertSubmitMsg(STsdbRepo *pRepo, SSubmitMsg *pMsg) {
   SSubmitMsgIter msgIter = {0};
   SSubmitBlk *   pBlock = NULL;
   SSubmitBlkIter blkIter = {0};
-  SDataRow       row = NULL;
+  SMemRow        row = NULL;
   TSKEY          now = taosGetTimestamp(pRepo->config.precision);
   TSKEY          minKey = now - tsTickPerDay[pRepo->config.precision] * pRepo->config.keep;
   TSKEY          maxKey = now + tsTickPerDay[pRepo->config.precision] * pRepo->config.daysPerFile;
@@ -698,7 +698,7 @@ static int tsdbInsertDataToTable(STsdbRepo *pRepo, SSubmitBlk *pBlock, int32_t *
   int64_t        points = 0;
   STable *       pTable = NULL;
   SSubmitBlkIter blkIter = {0};
-  SDataRow       row = NULL;
+  SMemRow        row = NULL;
   void *         rows[TSDB_MAX_INSERT_BATCH] = {0};
   int            rowCounter = 0;
 
@@ -744,10 +744,10 @@ _err:
   return -1;
 }
 
-static int tsdbCopyRowToMem(STsdbRepo *pRepo, SDataRow row, STable *pTable, void **ppRow) {
+static int tsdbCopyRowToMem(STsdbRepo *pRepo, SMemRow row, STable *pTable, void **ppRow) {
   STsdbCfg *  pCfg = &pRepo->config;
-  TKEY        tkey = dataRowTKey(row);
+  TKEY        tkey = memRowTKey(row);
-  TSKEY       key = dataRowKey(row);
+  TSKEY       key = memRowKey(row);
   bool        isRowDelete = TKEY_IS_DELETED(tkey);
 
   if (isRowDelete) {
@@ -765,15 +765,15 @@ static int tsdbCopyRowToMem(STsdbRepo *pRepo, SDataRow row, STable *pTable, void
     }
   }
 
-  void *pRow = tsdbAllocBytes(pRepo, dataRowLen(row));
+  void *pRow = tsdbAllocBytes(pRepo, memRowTLen(row));
   if (pRow == NULL) {
-    tsdbError("vgId:%d failed to insert row with key %" PRId64 " to table %s while allocate %d bytes since %s",
+    tsdbError("vgId:%d failed to insert row with key %" PRId64 " to table %s while allocate %" PRIu32 " bytes since %s",
-              REPO_ID(pRepo), key, TABLE_CHAR_NAME(pTable), dataRowLen(row), tstrerror(terrno));
+              REPO_ID(pRepo), key, TABLE_CHAR_NAME(pTable), memRowTLen(row), tstrerror(terrno));
     return -1;
   }
 
-  dataRowCpy(pRow, row);
+  memRowCpy(pRow, row);
-  ppRow[0] = pRow;
+  ppRow[0] = pRow;  // save the memory address of data rows
 
   tsdbTrace("vgId:%d a row is %s table %s tid %d uid %" PRIu64 " key %" PRIu64, REPO_ID(pRepo),
             isRowDelete ? "deleted from" : "updated in", TABLE_CHAR_NAME(pTable), TABLE_TID(pTable), TABLE_UID(pTable),
@@ -932,13 +932,15 @@ static int tsdbInsertDataToTableImpl(STsdbRepo *pRepo, STable *pTable, void **ro
   int64_t osize = SL_SIZE(pTableData->pData);
   tSkipListPutBatch(pTableData->pData, rows, rowCounter);
   int64_t dsize = SL_SIZE(pTableData->pData) - osize;
+  TSKEY   keyFirstRow = memRowKey(rows[0]);
+  TSKEY   keyLastRow = memRowKey(rows[rowCounter - 1]);
 
-  if (pMemTable->keyFirst > dataRowKey(rows[0])) pMemTable->keyFirst = dataRowKey(rows[0]);
+  if (pMemTable->keyFirst > keyFirstRow) pMemTable->keyFirst = keyFirstRow;
-  if (pMemTable->keyLast < dataRowKey(rows[rowCounter - 1])) pMemTable->keyLast = dataRowKey(rows[rowCounter - 1]);
+  if (pMemTable->keyLast < keyLastRow) pMemTable->keyLast = keyLastRow;
   pMemTable->numOfRows += dsize;
 
-  if (pTableData->keyFirst > dataRowKey(rows[0])) pTableData->keyFirst = dataRowKey(rows[0]);
+  if (pTableData->keyFirst > keyFirstRow) pTableData->keyFirst = keyFirstRow;
-  if (pTableData->keyLast < dataRowKey(rows[rowCounter - 1])) pTableData->keyLast = dataRowKey(rows[rowCounter - 1]);
+  if (pTableData->keyLast < keyLastRow) pTableData->keyLast = keyLastRow;
   pTableData->numOfRows += dsize;
 
   // update table latest info
@@ -954,8 +956,8 @@ static void tsdbFreeRows(STsdbRepo *pRepo, void **rows, int rowCounter) {
   STsdbBufPool *pBufPool = pRepo->pPool;
 
   for (int i = rowCounter - 1; i >= 0; --i) {
-    SDataRow row = (SDataRow)rows[i];
+    SMemRow row = (SMemRow)rows[i];
-    int      bytes = (int)dataRowLen(row);
+    int     bytes = (int)memRowTLen(row);
 
     if (pRepo->mem->extraBuffList == NULL) {
       STsdbBufBlock *pBufBlock = tsdbGetCurrBufBlock(pRepo);
@@ -988,21 +990,23 @@ static void tsdbFreeRows(STsdbRepo *pRepo, void **rows, int rowCounter) {
   }
 }
 
-static void updateTableLatestColumn(STsdbRepo *pRepo, STable *pTable, SDataRow row) {
+static void updateTableLatestColumn(STsdbRepo *pRepo, STable *pTable, SMemRow row) {
-  tsdbDebug("vgId:%d updateTableLatestColumn, %s row version:%d", REPO_ID(pRepo), pTable->name->data, dataRowVersion(row));
+  tsdbDebug("vgId:%d updateTableLatestColumn, %s row version:%d", REPO_ID(pRepo), pTable->name->data,
+            memRowVersion(row));
 
   STSchema* pSchema = tsdbGetTableLatestSchema(pTable);
   if (tsdbUpdateLastColSchema(pTable, pSchema) < 0) {
     return;
   }
 
-  pSchema = tsdbGetTableSchemaByVersion(pTable, dataRowVersion(row));
+  pSchema = tsdbGetTableSchemaByVersion(pTable, memRowVersion(row));
   if (pSchema == NULL) {
     return;
   }
 
   SDataCol *pLatestCols = pTable->lastCols;
 
+  bool isDataRow = isDataRow(row);
   for (int16_t j = 0; j < schemaNCols(pSchema); j++) {
     STColumn *pTCol = schemaColAt(pSchema, j);
     // ignore not exist colId
@@ -1010,9 +1014,21 @@ static void updateTableLatestColumn(STsdbRepo *pRepo, STable *pTable, SDataRow r
     if (idx == -1) {
       continue;
     }
-    
+
-    void* value = tdGetRowDataOfCol(row, (int8_t)pTCol->type, TD_DATA_ROW_HEAD_SIZE + pSchema->columns[j].offset);
+    void *value = NULL;
-    if (isNull(value, pTCol->type)) {
+
+    if (isDataRow) {
+      value = tdGetRowDataOfCol(memRowDataBody(row), (int8_t)pTCol->type,
+                                TD_DATA_ROW_HEAD_SIZE + pSchema->columns[j].offset);
+    } else {
+      // SKVRow
+      SColIdx *pColIdx = tdGetKVRowIdxOfCol(memRowKvBody(row), pTCol->colId);
+      if (pColIdx) {
+        value = tdGetKvRowDataOfCol(memRowKvBody(row), pColIdx->offset);
+      }
+    }
+
+    if ((value == NULL) || isNull(value, pTCol->type)) {
       continue;
     }
 
@@ -1027,11 +1043,11 @@ static void updateTableLatestColumn(STsdbRepo *pRepo, STable *pTable, SDataRow r
 
     memcpy(pDataCol->pData, value, pDataCol->bytes);
     //tsdbInfo("updateTableLatestColumn vgId:%d cache column %d for %d,%s", REPO_ID(pRepo), j, pDataCol->bytes, (char*)pDataCol->pData);
-    pDataCol->ts = dataRowKey(row);
+    pDataCol->ts = memRowKey(row);
   }
 }
 
-static int tsdbUpdateTableLatestInfo(STsdbRepo *pRepo, STable *pTable, SDataRow row) {
+static int tsdbUpdateTableLatestInfo(STsdbRepo *pRepo, STable *pTable, SMemRow row) {
   STsdbCfg *pCfg = &pRepo->config;
 
   // if cacheLastRow config has been reset, free the lastRow
@@ -1042,31 +1058,31 @@ static int tsdbUpdateTableLatestInfo(STsdbRepo *pRepo, STable *pTable, SDataRow
     TSDB_WUNLOCK_TABLE(pTable);
   }
 
-  if (tsdbGetTableLastKeyImpl(pTable) < dataRowKey(row)) {
+  if (tsdbGetTableLastKeyImpl(pTable) < memRowKey(row)) {
     if (CACHE_LAST_ROW(pCfg) || pTable->lastRow != NULL) {
-      SDataRow nrow = pTable->lastRow;
+      SMemRow nrow = pTable->lastRow;
-      if (taosTSizeof(nrow) < dataRowLen(row)) {
+      if (taosTSizeof(nrow) < memRowTLen(row)) {
-        SDataRow orow = nrow;
+        SMemRow orow = nrow;
-        nrow = taosTMalloc(dataRowLen(row));
+        nrow = taosTMalloc(memRowTLen(row));
         if (nrow == NULL) {
           terrno = TSDB_CODE_TDB_OUT_OF_MEMORY;
           return -1;
         }
 
-        dataRowCpy(nrow, row);
+        memRowCpy(nrow, row);
         TSDB_WLOCK_TABLE(pTable);
-        pTable->lastKey = dataRowKey(row);
+        pTable->lastKey = memRowKey(row);
         pTable->lastRow = nrow;
         TSDB_WUNLOCK_TABLE(pTable);
         taosTZfree(orow);
       } else {
         TSDB_WLOCK_TABLE(pTable);
-        pTable->lastKey = dataRowKey(row);
+        pTable->lastKey = memRowKey(row);
-        dataRowCpy(nrow, row);
+        memRowCpy(nrow, row);
         TSDB_WUNLOCK_TABLE(pTable);
       }
     } else {
-      pTable->lastKey = dataRowKey(row);
+      pTable->lastKey = memRowKey(row);
     }
 
     if (CACHE_LAST_NULL_COLUMN(pCfg)) {

src/tsdb/src/tsdbRead.c

@@ -147,7 +147,7 @@ typedef struct STableGroupSupporter {
 static STimeWindow updateLastrowForEachGroup(STableGroupInfo *groupList);
 static int32_t checkForCachedLastRow(STsdbQueryHandle* pQueryHandle, STableGroupInfo *groupList);
 static int32_t checkForCachedLast(STsdbQueryHandle* pQueryHandle);
-static int32_t tsdbGetCachedLastRow(STable* pTable, SDataRow* pRes, TSKEY* lastKey);
+static int32_t tsdbGetCachedLastRow(STable* pTable, SMemRow* pRes, TSKEY* lastKey);
 
 static void    changeQueryHandleForInterpQuery(TsdbQueryHandleT pHandle);
 static void    doMergeTwoLevelData(STsdbQueryHandle* pQueryHandle, STableCheckInfo* pCheckInfo, SBlock* pBlock);
@@ -734,8 +734,8 @@ static bool initTableMemIterator(STsdbQueryHandle* pHandle, STableCheckInfo* pCh
     SSkipListNode* node = tSkipListIterGet(pCheckInfo->iter);
     assert(node != NULL);
 
-    SDataRow row = (SDataRow)SL_GET_NODE_DATA(node);
+    SMemRow row = (SMemRow)SL_GET_NODE_DATA(node);
-    TSKEY key = dataRowKey(row);  // first timestamp in buffer
+    TSKEY   key = memRowKey(row);  // first timestamp in buffer
     tsdbDebug("%p uid:%" PRId64 ", tid:%d check data in mem from skey:%" PRId64 ", order:%d, ts range in buf:%" PRId64
               "-%" PRId64 ", lastKey:%" PRId64 ", numOfRows:%"PRId64", 0x%"PRIx64,
               pHandle, pCheckInfo->tableId.uid, pCheckInfo->tableId.tid, key, order, pMem->keyFirst, pMem->keyLast,
@@ -756,8 +756,8 @@ static bool initTableMemIterator(STsdbQueryHandle* pHandle, STableCheckInfo* pCh
     SSkipListNode* node = tSkipListIterGet(pCheckInfo->iiter);
     assert(node != NULL);
 
-    SDataRow row = (SDataRow)SL_GET_NODE_DATA(node);
+    SMemRow row = (SMemRow)SL_GET_NODE_DATA(node);
-    TSKEY key = dataRowKey(row);  // first timestamp in buffer
+    TSKEY   key = memRowKey(row);  // first timestamp in buffer
     tsdbDebug("%p uid:%" PRId64 ", tid:%d check data in imem from skey:%" PRId64 ", order:%d, ts range in buf:%" PRId64
               "-%" PRId64 ", lastKey:%" PRId64 ", numOfRows:%"PRId64", 0x%"PRIx64,
               pHandle, pCheckInfo->tableId.uid, pCheckInfo->tableId.tid, key, order, pIMem->keyFirst, pIMem->keyLast,
@@ -781,19 +781,19 @@ static void destroyTableMemIterator(STableCheckInfo* pCheckInfo) {
   tSkipListDestroyIter(pCheckInfo->iiter);
 }
 
-static SDataRow getSDataRowInTableMem(STableCheckInfo* pCheckInfo, int32_t order, int32_t update) {
+static SMemRow getSDataRowInTableMem(STableCheckInfo* pCheckInfo, int32_t order, int32_t update) {
-  SDataRow rmem = NULL, rimem = NULL;
+  SMemRow rmem = NULL, rimem = NULL;
   if (pCheckInfo->iter) {
     SSkipListNode* node = tSkipListIterGet(pCheckInfo->iter);
     if (node != NULL) {
-      rmem = (SDataRow)SL_GET_NODE_DATA(node);
+      rmem = (SMemRow)SL_GET_NODE_DATA(node);
     }
   }
 
   if (pCheckInfo->iiter) {
     SSkipListNode* node = tSkipListIterGet(pCheckInfo->iiter);
     if (node != NULL) {
-      rimem = (SDataRow)SL_GET_NODE_DATA(node);
+      rimem = (SMemRow)SL_GET_NODE_DATA(node);
     }
   }
 
@@ -811,8 +811,8 @@ static SDataRow getSDataRowInTableMem(STableCheckInfo* pCheckInfo, int32_t order
     return rimem;
   }
 
-  TSKEY r1 = dataRowKey(rmem);
+  TSKEY r1 = memRowKey(rmem);
-  TSKEY r2 = dataRowKey(rimem);
+  TSKEY r2 = memRowKey(rimem);
 
   if (r1 == r2) { // data ts are duplicated, ignore the data in mem
     if (!update) {
@@ -891,12 +891,12 @@ static bool hasMoreDataInCache(STsdbQueryHandle* pHandle) {
     initTableMemIterator(pHandle, pCheckInfo);
   }
 
-  SDataRow row = getSDataRowInTableMem(pCheckInfo, pHandle->order, pCfg->update);
+  SMemRow row = getSDataRowInTableMem(pCheckInfo, pHandle->order, pCfg->update);
   if (row == NULL) {
     return false;
   }
 
-  pCheckInfo->lastKey = dataRowKey(row);  // first timestamp in buffer
+  pCheckInfo->lastKey = memRowKey(row);  // first timestamp in buffer
   tsdbDebug("%p uid:%" PRId64", tid:%d check data in buffer from skey:%" PRId64 ", order:%d, 0x%"PRIx64, pHandle,
       pCheckInfo->tableId.uid, pCheckInfo->tableId.tid, pCheckInfo->lastKey, pHandle->order, pHandle->qId);
 
@@ -1155,11 +1155,11 @@ static int32_t handleDataMergeIfNeeded(STsdbQueryHandle* pQueryHandle, SBlock* p
   int32_t code = TSDB_CODE_SUCCESS;
 
   /*bool hasData = */ initTableMemIterator(pQueryHandle, pCheckInfo);
-  SDataRow row = getSDataRowInTableMem(pCheckInfo, pQueryHandle->order, pCfg->update);
+  SMemRow row = getSDataRowInTableMem(pCheckInfo, pQueryHandle->order, pCfg->update);
 
   assert(cur->pos >= 0 && cur->pos <= binfo.rows);
 
-  TSKEY key = (row != NULL)? dataRowKey(row):TSKEY_INITIAL_VAL;
+  TSKEY key = (row != NULL) ? memRowKey(row) : TSKEY_INITIAL_VAL;
   if (key != TSKEY_INITIAL_VAL) {
     tsdbDebug("%p key in mem:%"PRId64", 0x%"PRIx64, pQueryHandle, key, pQueryHandle->qId);
   } else {
@@ -1401,7 +1401,7 @@ int32_t doCopyRowsFromFileBlock(STsdbQueryHandle* pQueryHandle, int32_t capacity
 
         // todo refactor, only copy one-by-one
         for (int32_t k = start; k < num + start; ++k) {
-          char* p = tdGetColDataOfRow(src, k);
+          const char* p = tdGetColDataOfRow(src, k);
           memcpy(dst, p, varDataTLen(p));
           dst += bytes;
         }
@@ -1452,88 +1452,170 @@ int32_t doCopyRowsFromFileBlock(STsdbQueryHandle* pQueryHandle, int32_t capacity
   return numOfRows + num;
 }
 
-static void copyOneRowFromMem(STsdbQueryHandle* pQueryHandle, int32_t capacity, int32_t numOfRows, SDataRow row,
+static void copyOneRowFromMem(STsdbQueryHandle* pQueryHandle, int32_t capacity, int32_t numOfRows, SMemRow row,
                               int32_t numOfCols, STable* pTable, STSchema* pSchema) {
   char* pData = NULL;
 
-  // the schema version info is embeded in SDataRow
+  // the schema version info is embedded in SDataRow, and use latest schema version for SKVRow
   int32_t numOfRowCols = 0;
   if (pSchema == NULL) {
-    pSchema = tsdbGetTableSchemaByVersion(pTable, dataRowVersion(row));
+    pSchema = tsdbGetTableSchemaByVersion(pTable, memRowVersion(row));
     numOfRowCols = schemaNCols(pSchema);
   } else {
     numOfRowCols = schemaNCols(pSchema);
   }
-  
-  int32_t i = 0, j = 0;
-  while(i < numOfCols && j < numOfRowCols) {
-    SColumnInfoData* pColInfo = taosArrayGet(pQueryHandle->pColumns, i);
-    if (pSchema->columns[j].colId < pColInfo->info.colId) {
-      j++;
-      continue;
-    }
 
-    if (ASCENDING_TRAVERSE(pQueryHandle->order)) {
+  int32_t i = 0;
-      pData = (char*)pColInfo->pData + numOfRows * pColInfo->info.bytes;
-    } else {
-      pData = (char*)pColInfo->pData + (capacity - numOfRows - 1) * pColInfo->info.bytes;
-    }
 
-    if (pSchema->columns[j].colId == pColInfo->info.colId) {
+  if (isDataRow(row)) {
-      void* value = tdGetRowDataOfCol(row, (int8_t)pColInfo->info.type, TD_DATA_ROW_HEAD_SIZE + pSchema->columns[j].offset);
+    SDataRow dataRow = memRowDataBody(row);
-      switch (pColInfo->info.type) {
+    int32_t j = 0;
-        case TSDB_DATA_TYPE_BINARY:
+    while (i < numOfCols && j < numOfRowCols) {
-        case TSDB_DATA_TYPE_NCHAR:
+      SColumnInfoData* pColInfo = taosArrayGet(pQueryHandle->pColumns, i);
-          memcpy(pData, value, varDataTLen(value));
+      if (pSchema->columns[j].colId < pColInfo->info.colId) {
-          break;
+        j++;
-        case TSDB_DATA_TYPE_NULL:
+        continue;
-        case TSDB_DATA_TYPE_BOOL:
-        case TSDB_DATA_TYPE_TINYINT:
-        case TSDB_DATA_TYPE_UTINYINT:  
-          *(uint8_t *)pData = *(uint8_t *)value;
-          break;
-        case TSDB_DATA_TYPE_SMALLINT:
-        case TSDB_DATA_TYPE_USMALLINT:
-          *(uint16_t *)pData = *(uint16_t *)value;
-          break;
-        case TSDB_DATA_TYPE_INT:
-        case TSDB_DATA_TYPE_UINT:
-          *(uint32_t *)pData = *(uint32_t *)value;
-          break;
-        case TSDB_DATA_TYPE_BIGINT:
-        case TSDB_DATA_TYPE_UBIGINT:
-          *(uint64_t *)pData = *(uint64_t *)value;
-          break;
-        case TSDB_DATA_TYPE_FLOAT:
-          SET_FLOAT_PTR(pData, value);
-          break;
-        case TSDB_DATA_TYPE_DOUBLE:
-          SET_DOUBLE_PTR(pData, value);
-          break;
-        case TSDB_DATA_TYPE_TIMESTAMP:
-          if (pColInfo->info.colId == PRIMARYKEY_TIMESTAMP_COL_INDEX) {
-            *(TSKEY *)pData = tdGetKey(*(TKEY *)value);
-          } else {
-            *(TSKEY *)pData = *(TSKEY *)value;
-          }
-          break;
-        default:
-          memcpy(pData, value, pColInfo->info.bytes);
       }
 
-      j++;
+      if (ASCENDING_TRAVERSE(pQueryHandle->order)) {
-      i++;
+        pData = (char*)pColInfo->pData + numOfRows * pColInfo->info.bytes;
-    } else { // pColInfo->info.colId < pSchema->columns[j].colId, it is a NULL data
+      } else {
+        pData = (char*)pColInfo->pData + (capacity - numOfRows - 1) * pColInfo->info.bytes;
+      }
+
+      if (pSchema->columns[j].colId == pColInfo->info.colId) {
+        void* value =
+            tdGetRowDataOfCol(dataRow, (int8_t)pColInfo->info.type, TD_DATA_ROW_HEAD_SIZE + pSchema->columns[j].offset);
+        switch (pColInfo->info.type) {
+          case TSDB_DATA_TYPE_BINARY:
+          case TSDB_DATA_TYPE_NCHAR:
+            memcpy(pData, value, varDataTLen(value));
+            break;
+          case TSDB_DATA_TYPE_NULL:
+          case TSDB_DATA_TYPE_BOOL:
+          case TSDB_DATA_TYPE_TINYINT:
+          case TSDB_DATA_TYPE_UTINYINT:
+            *(uint8_t*)pData = *(uint8_t*)value;
+            break;
+          case TSDB_DATA_TYPE_SMALLINT:
+          case TSDB_DATA_TYPE_USMALLINT:
+            *(uint16_t*)pData = *(uint16_t*)value;
+            break;
+          case TSDB_DATA_TYPE_INT:
+          case TSDB_DATA_TYPE_UINT:
+            *(uint32_t*)pData = *(uint32_t*)value;
+            break;
+          case TSDB_DATA_TYPE_BIGINT:
+          case TSDB_DATA_TYPE_UBIGINT:
+            *(uint64_t*)pData = *(uint64_t*)value;
+            break;
+          case TSDB_DATA_TYPE_FLOAT:
+            SET_FLOAT_PTR(pData, value);
+            break;
+          case TSDB_DATA_TYPE_DOUBLE:
+            SET_DOUBLE_PTR(pData, value);
+            break;
+          case TSDB_DATA_TYPE_TIMESTAMP:
+            if (pColInfo->info.colId == PRIMARYKEY_TIMESTAMP_COL_INDEX) {
+              *(TSKEY*)pData = tdGetKey(*(TKEY*)value);
+            } else {
+              *(TSKEY*)pData = *(TSKEY*)value;
+            }
+            break;
+          default:
+            memcpy(pData, value, pColInfo->info.bytes);
+        }
+
+        j++;
+        i++;
+      } else {  // pColInfo->info.colId < pSchema->columns[j].colId, it is a NULL data
+        if (pColInfo->info.type == TSDB_DATA_TYPE_BINARY || pColInfo->info.type == TSDB_DATA_TYPE_NCHAR) {
+          setVardataNull(pData, pColInfo->info.type);
+        } else {
+          setNull(pData, pColInfo->info.type, pColInfo->info.bytes);
+        }
+        i++;
+      }
+    }
+  } else if (isKvRow(row)) {
+    SKVRow  kvRow = memRowKvBody(row);
+    int32_t k = 0;
+    int32_t nKvRowCols = kvRowNCols(kvRow);
+
+    while (i < numOfCols && k < nKvRowCols) {
+      SColumnInfoData* pColInfo = taosArrayGet(pQueryHandle->pColumns, i);
+      SColIdx*         pColIdx = kvRowColIdxAt(kvRow, k);
+
+      if (pColIdx->colId < pColInfo->info.colId) {
+        ++k;
+        continue;
+      }
+
+      if (ASCENDING_TRAVERSE(pQueryHandle->order)) {
+        pData = (char*)pColInfo->pData + numOfRows * pColInfo->info.bytes;
+      } else {
+        pData = (char*)pColInfo->pData + (capacity - numOfRows - 1) * pColInfo->info.bytes;
+      }
+
+      if (pColIdx->colId == pColInfo->info.colId) {
+        // offset of pColIdx for SKVRow including the TD_KV_ROW_HEAD_SIZE
+        void* value = tdGetKvRowDataOfCol(kvRow, pColIdx->offset);
+        switch (pColInfo->info.type) {
+          case TSDB_DATA_TYPE_BINARY:
+          case TSDB_DATA_TYPE_NCHAR:
+            memcpy(pData, value, varDataTLen(value));
+            break;
+          case TSDB_DATA_TYPE_NULL:
+          case TSDB_DATA_TYPE_BOOL:
+          case TSDB_DATA_TYPE_TINYINT:
+          case TSDB_DATA_TYPE_UTINYINT:
+            *(uint8_t*)pData = *(uint8_t*)value;
+            break;
+          case TSDB_DATA_TYPE_SMALLINT:
+          case TSDB_DATA_TYPE_USMALLINT:
+            *(uint16_t*)pData = *(uint16_t*)value;
+            break;
+          case TSDB_DATA_TYPE_INT:
+          case TSDB_DATA_TYPE_UINT:
+            *(uint32_t*)pData = *(uint32_t*)value;
+            break;
+          case TSDB_DATA_TYPE_BIGINT:
+          case TSDB_DATA_TYPE_UBIGINT:
+            *(uint64_t*)pData = *(uint64_t*)value;
+            break;
+          case TSDB_DATA_TYPE_FLOAT:
+            SET_FLOAT_PTR(pData, value);
+            break;
+          case TSDB_DATA_TYPE_DOUBLE:
+            SET_DOUBLE_PTR(pData, value);
+            break;
+          case TSDB_DATA_TYPE_TIMESTAMP:
+            if (pColInfo->info.colId == PRIMARYKEY_TIMESTAMP_COL_INDEX) {
+              *(TSKEY*)pData = tdGetKey(*(TKEY*)value);
+            } else {
+              *(TSKEY*)pData = *(TSKEY*)value;
+            }
+            break;
+          default:
+            memcpy(pData, value, pColInfo->info.bytes);
+        }
+        ++k;
+        ++i;
+        continue;
+      }
+      // If (pColInfo->info.colId < pColIdx->colId), it is NULL data
       if (pColInfo->info.type == TSDB_DATA_TYPE_BINARY || pColInfo->info.type == TSDB_DATA_TYPE_NCHAR) {
         setVardataNull(pData, pColInfo->info.type);
       } else {
         setNull(pData, pColInfo->info.type, pColInfo->info.bytes);
       }
-      i++;
+      ++i;
     }
+  } else {
+    ASSERT(0);
   }
 
-  while (i < numOfCols) { // the remain columns are all null data
+  while (i < numOfCols) {  // the remain columns are all null data
     SColumnInfoData* pColInfo = taosArrayGet(pQueryHandle->pColumns, i);
     if (ASCENDING_TRAVERSE(pQueryHandle->order)) {
       pData = (char*)pColInfo->pData + numOfRows * pColInfo->info.bytes;
@@ -1550,7 +1632,6 @@ static void copyOneRowFromMem(STsdbQueryHandle* pQueryHandle, int32_t capacity,
     i++;
   }
 }
-
 static void moveDataToFront(STsdbQueryHandle* pQueryHandle, int32_t numOfRows, int32_t numOfCols) {
   if (numOfRows == 0 || ASCENDING_TRAVERSE(pQueryHandle->order)) {
     return;
@@ -1730,12 +1811,12 @@ static void doMergeTwoLevelData(STsdbQueryHandle* pQueryHandle, STableCheckInfo*
   } else if (pCheckInfo->iter != NULL || pCheckInfo->iiter != NULL) {
     SSkipListNode* node = NULL;
     do {
-      SDataRow row = getSDataRowInTableMem(pCheckInfo, pQueryHandle->order, pCfg->update);
+      SMemRow row = getSDataRowInTableMem(pCheckInfo, pQueryHandle->order, pCfg->update);
       if (row == NULL) {
         break;
       }
 
-      TSKEY key = dataRowKey(row);
+      TSKEY key = memRowKey(row);
       if ((key > pQueryHandle->window.ekey && ASCENDING_TRAVERSE(pQueryHandle->order)) ||
           (key < pQueryHandle->window.ekey && !ASCENDING_TRAVERSE(pQueryHandle->order))) {
         break;
@@ -1748,11 +1829,11 @@ static void doMergeTwoLevelData(STsdbQueryHandle* pQueryHandle, STableCheckInfo*
 
       if ((key < tsArray[pos] && ASCENDING_TRAVERSE(pQueryHandle->order)) ||
           (key > tsArray[pos] && !ASCENDING_TRAVERSE(pQueryHandle->order))) {
-        if (rv != dataRowVersion(row)) {
+        if (rv != memRowVersion(row)) {
-          pSchema = tsdbGetTableSchemaByVersion(pTable, dataRowVersion(row));
+          pSchema = tsdbGetTableSchemaByVersion(pTable, memRowVersion(row));
-          rv = dataRowVersion(row);
+          rv = memRowVersion(row);
-        }  
+        }
-        
+
         copyOneRowFromMem(pQueryHandle, pQueryHandle->outputCapacity, numOfRows, row, numOfCols, pTable, pSchema);
         numOfRows += 1;
         if (cur->win.skey == TSKEY_INITIAL_VAL) {
@@ -1766,11 +1847,11 @@ static void doMergeTwoLevelData(STsdbQueryHandle* pQueryHandle, STableCheckInfo*
         moveToNextRowInMem(pCheckInfo);
       } else if (key == tsArray[pos]) {  // data in buffer has the same timestamp of data in file block, ignore it
         if (pCfg->update) {
-          if (rv != dataRowVersion(row)) {
+          if (rv != memRowVersion(row)) {
-            pSchema = tsdbGetTableSchemaByVersion(pTable, dataRowVersion(row));
+            pSchema = tsdbGetTableSchemaByVersion(pTable, memRowVersion(row));
-            rv = dataRowVersion(row);
+            rv = memRowVersion(row);
           }
-          
+
           copyOneRowFromMem(pQueryHandle, pQueryHandle->outputCapacity, numOfRows, row, numOfCols, pTable, pSchema);
           numOfRows += 1;
           if (cur->win.skey == TSKEY_INITIAL_VAL) {
@@ -1820,8 +1901,10 @@ static void doMergeTwoLevelData(STsdbQueryHandle* pQueryHandle, STableCheckInfo*
        * copy them all to result buffer, since it may be overlapped with file data block.
        */
       if (node == NULL ||
-          ((dataRowKey((SDataRow)SL_GET_NODE_DATA(node)) > pQueryHandle->window.ekey) && ASCENDING_TRAVERSE(pQueryHandle->order)) ||
+          ((memRowKey((SMemRow)SL_GET_NODE_DATA(node)) > pQueryHandle->window.ekey) &&
-          ((dataRowKey((SDataRow)SL_GET_NODE_DATA(node)) < pQueryHandle->window.ekey) && !ASCENDING_TRAVERSE(pQueryHandle->order))) {
+           ASCENDING_TRAVERSE(pQueryHandle->order)) ||
+          ((memRowKey((SMemRow)SL_GET_NODE_DATA(node)) < pQueryHandle->window.ekey) &&
+           !ASCENDING_TRAVERSE(pQueryHandle->order))) {
         // no data in cache or data in cache is greater than the ekey of time window, load data from file block
         if (cur->win.skey == TSKEY_INITIAL_VAL) {
           cur->win.skey = tsArray[pos];
@@ -2407,12 +2490,12 @@ static int tsdbReadRowsFromCache(STableCheckInfo* pCheckInfo, TSKEY maxKey, int
   STSchema* pSchema = NULL;
 
   do {
-    SDataRow row = getSDataRowInTableMem(pCheckInfo, pQueryHandle->order, pCfg->update);
+    SMemRow row = getSDataRowInTableMem(pCheckInfo, pQueryHandle->order, pCfg->update);
     if (row == NULL) {
       break;
     }
 
-    TSKEY key = dataRowKey(row);
+    TSKEY key = memRowKey(row);
     if ((key > maxKey && ASCENDING_TRAVERSE(pQueryHandle->order)) || (key < maxKey && !ASCENDING_TRAVERSE(pQueryHandle->order))) {
       tsdbDebug("%p key:%"PRIu64" beyond qrange:%"PRId64" - %"PRId64", no more data in buffer", pQueryHandle, key, pQueryHandle->window.skey,
                 pQueryHandle->window.ekey);
@@ -2425,9 +2508,9 @@ static int tsdbReadRowsFromCache(STableCheckInfo* pCheckInfo, TSKEY maxKey, int
     }
 
     win->ekey = key;
-    if (rv != dataRowVersion(row)) {
+    if (rv != memRowVersion(row)) {
-      pSchema = tsdbGetTableSchemaByVersion(pTable, dataRowVersion(row));
+      pSchema = tsdbGetTableSchemaByVersion(pTable, memRowVersion(row));
-      rv = dataRowVersion(row);
+      rv = memRowVersion(row);
     }
     copyOneRowFromMem(pQueryHandle, maxRowsToRead, numOfRows, row, numOfCols, pTable, pSchema);
 
@@ -2544,7 +2627,7 @@ static bool loadCachedLastRow(STsdbQueryHandle* pQueryHandle) {
 
   SQueryFilePos* cur = &pQueryHandle->cur;
 
-  SDataRow pRow = NULL;
+  SMemRow  pRow = NULL;
   TSKEY    key  = TSKEY_INITIAL_VAL;
   int32_t  step = ASCENDING_TRAVERSE(pQueryHandle->order)? 1:-1;
 
@@ -2954,7 +3037,7 @@ bool tsdbGetExternalRow(TsdbQueryHandleT pHandle) {
  * if lastRow == NULL, return TSDB_CODE_TDB_NO_CACHE_LAST_ROW
  * else set pRes and return TSDB_CODE_SUCCESS and save lastKey
  */
-int32_t tsdbGetCachedLastRow(STable* pTable, SDataRow* pRes, TSKEY* lastKey) {
+int32_t tsdbGetCachedLastRow(STable* pTable, SMemRow* pRes, TSKEY* lastKey) {
   int32_t code = TSDB_CODE_SUCCESS;
 
   TSDB_RLOCK_TABLE(pTable);
@@ -2965,7 +3048,7 @@ int32_t tsdbGetCachedLastRow(STable* pTable, SDataRow* pRes, TSKEY* lastKey) {
   }
 
   if (pRes) {
-    *pRes = tdDataRowDup(pTable->lastRow);
+    *pRes = tdMemRowDup(pTable->lastRow);
     if (*pRes == NULL) {
       code = TSDB_CODE_TDB_OUT_OF_MEMORY;
     }

src/tsdb/tests/tsdbTests.cpp

@@ -55,10 +55,10 @@ static int insertData(SInsertInfo *pInfo) {
       for (int j = 0; j < schemaNCols(pInfo->pSchema); j++) {
         STColumn *pTCol = schemaColAt(pInfo->pSchema, j);
         if (j == 0) {  // Just for timestamp
-          tdAppendColVal(row, (void *)(&start_time), pTCol->type, pTCol->bytes, pTCol->offset);
+          tdAppendColVal(row, (void *)(&start_time), pTCol->type, pTCol->offset);
         } else {  // For int
           int val = 10;
-          tdAppendColVal(row, (void *)(&val), pTCol->type, pTCol->bytes, pTCol->offset);
+          tdAppendColVal(row, (void *)(&val), pTCol->type, pTCol->offset);
         }
       }
       pBlock->dataLen += dataRowLen(row);

tests/pytest/fulltest.sh

@@ -352,7 +352,7 @@ python3 ./test.py -f alter/alter_debugFlag.py
 python3 ./test.py -f query/queryBetweenAnd.py
 python3 ./test.py -f tag_lite/alter_tag.py
 
-python3 test.py -f tools/taosdemoAllTest/taosdemoTestInsertWithJson.py 
+# python3 test.py -f tools/taosdemoAllTest/taosdemoTestInsertWithJson.py 
 python3 test.py -f tools/taosdemoAllTest/taosdemoTestQueryWithJson.py
 python3 test.py -f tools/taosdemoAllTest/TD-4985/query-limit-offset.py
 python3 ./test.py -f tag_lite/drop_auto_create.py

tests/pytest/table/max_table_length.py

@@ -42,7 +42,7 @@ class TDTestCase:
         
         print("==============step3")
         tdLog.info("check int & binary")
-        tdSql.error("create table anal2 (ts timestamp ,i binary(16371),j int)")
+        # tdSql.error("create table anal2 (ts timestamp ,i binary(16371),j int)")
         tdSql.execute("create table anal2 (ts timestamp ,i binary(16370),j int)")
         tdSql.execute("create table anal3 (ts timestamp ,i binary(16366), j int, k int)")
         

tests/pytest/tools/taosdemoAllTest/TD-5213/insertSigcolumnsNum4096.csv

@@ -0,0 +1,10 @@
+0,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
+1,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
+2,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
+3,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
+4,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
+5,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
+6,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
+7,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
+8,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
+9,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,

tests/pytest/tools/taosdemoAllTest/TD-5213/insertSigcolumnsNum4096.json

@@ -0,0 +1,112 @@
+{
+    "filetype": "insert",
+    "cfgdir": "/etc/taos",
+    "host": "127.0.0.1",
+    "port": 6030,
+    "user": "root",
+    "password": "taosdata",
+    "thread_count": 10,
+    "thread_count_create_tbl": 10,
+    "result_file": "./insert_res.txt",
+    "confirm_parameter_prompt": "no",
+    "insert_interval": 0,
+    "interlace_rows": 10,
+    "num_of_records_per_req": 1,
+    "max_sql_len": 102400000,
+    "databases": [{
+        "dbinfo": {
+            "name": "db",
+            "drop": "yes",
+            "replica": 1,
+            "days": 10,
+            "cache": 50,
+            "blocks": 8,
+            "precision": "ms",
+            "keep": 365,
+            "minRows": 100,
+            "maxRows": 4096,
+            "comp":2,
+            "walLevel":1,
+            "cachelast":0,
+            "quorum":1,
+            "fsync":3000,
+            "update": 0
+        },
+        "super_tables": [{
+            "name": "stb_old",
+            "child_table_exists":"no",
+            "childtable_count": 10,
+            "childtable_prefix": "stb_old_",
+            "auto_create_table": "no",
+            "batch_create_tbl_num": 5,
+            "data_source": "sample",
+            "insert_mode": "taosc",
+            "insert_rows": 100,
+            "childtable_limit": 0,
+            "childtable_offset":0,
+            "multi_thread_write_one_tbl": "no",
+            "interlace_rows": 0,
+            "insert_interval":0,
+            "max_sql_len": 1024000,
+            "disorder_ratio": 0,
+            "disorder_range": 1000,
+            "timestamp_step": 1,
+            "start_timestamp": "2020-10-01 00:00:00.000",
+            "sample_format": "csv",
+            "sample_file": "./tools/taosdemoAllTest/TD-5213/insertSigcolumnsNum4096.csv",
+            "tags_file": "",
+            "columns": [{"type": "INT","count":4000}, {"type": "BINARY", "len": 16, "count":1}],
+            "tags": [{"type": "TINYINT", "count":2}, {"type": "BINARY", "len": 16, "count":1}]
+        },{
+            "name": "stb_new",
+            "child_table_exists":"no",
+            "childtable_count": 10,
+            "childtable_prefix": "stb_new_",
+            "auto_create_table": "no",
+            "batch_create_tbl_num": 5,
+            "data_source": "rand",
+            "insert_mode": "taosc",
+            "insert_rows": 100,
+            "childtable_limit": 0,
+            "childtable_offset":0,
+            "multi_thread_write_one_tbl": "no",
+            "interlace_rows": 0,
+            "insert_interval":0,
+            "max_sql_len": 1024000,
+            "disorder_ratio": 0,
+            "disorder_range": 1000,
+            "timestamp_step": 1,
+            "start_timestamp": "2020-10-01 00:00:00.000",
+            "sample_format": "csv",
+            "sample_file": "./tools/taosdemoAllTest/sample.csv",
+            "tags_file": "",
+            "columns": [{"type": "DOUBLE","count":1020}],
+            "tags": [{"type": "TINYINT", "count":2}, {"type": "BINARY", "len": 16, "count":1}]
+        },{
+            "name": "stb_int",
+            "child_table_exists":"no",
+            "childtable_count": 10,
+            "childtable_prefix": "stb_int_",
+            "auto_create_table": "no",
+            "batch_create_tbl_num": 5,
+            "data_source": "rand",
+            "insert_mode": "taosc",
+            "insert_rows": 100,
+            "childtable_limit": 0,
+            "childtable_offset":0,
+            "multi_thread_write_one_tbl": "no",
+            "interlace_rows": 0,
+            "insert_interval":0,
+            "max_sql_len": 1024000,
+            "disorder_ratio": 0,
+            "disorder_range": 1000,
+            "timestamp_step": 1,
+            "start_timestamp": "2020-10-01 00:00:00.000",
+            "sample_format": "csv",
+            "sample_file": "./tools/taosdemoAllTest/sample.csv",
+            "tags_file": "",
+            "columns": [{"type": "int","count":1020}],
+            "tags": [{"type": "TINYINT", "count":2}, {"type": "BINARY", "len": 16, "count":1}]
+        }]
+    }]
+}

tests/pytest/tools/taosdemoAllTest/TD-5213/insertSigcolumnsNum4096.py

@@ -0,0 +1,137 @@
+###################################################################
+#           Copyright (c) 2016 by TAOS Technologies, Inc.
+#                     All rights reserved.
+#
+#  This file is proprietary and confidential to TAOS Technologies.
+#  No part of this file may be reproduced, stored, transmitted,
+#  disclosed or used in any form or by any means other than as
+#  expressly provided by the written permission from Jianhui Tao
+#
+###################################################################
+
+# -*- coding: utf-8 -*-
+
+import sys
+import os
+from util.log import *
+from util.cases import *
+from util.sql import *
+from util.dnodes import *
+
+
+class TDTestCase:
+    def init(self, conn, logSql):
+        tdLog.debug("start to execute %s" % __file__)
+        tdSql.init(conn.cursor(), logSql)
+        
+    def getBuildPath(self):
+        selfPath = os.path.dirname(os.path.realpath(__file__))
+
+        if ("community" in selfPath):
+            projPath = selfPath[:selfPath.find("community")]
+        else:
+            projPath = selfPath[:selfPath.find("tests")]
+
+        for root, dirs, files in os.walk(projPath):
+            if ("taosd" in files):
+                rootRealPath = os.path.dirname(os.path.realpath(root))
+                if ("packaging" not in rootRealPath):
+                    buildPath = root[:len(root)-len("/build/bin")]
+                    break
+        return buildPath
+        
+    def run(self):
+        buildPath = self.getBuildPath()
+        if (buildPath == ""):
+            tdLog.exit("taosd not found!")
+        else:
+            tdLog.info("taosd found in %s" % buildPath)
+        binPath = buildPath+ "/build/bin/"
+
+        # insert: create one  or mutiple tables per sql and insert multiple rows per sql 
+        # test case for https://jira.taosdata.com:18080/browse/TD-5213
+        os.system("%staosdemo -f tools/taosdemoAllTest/TD-5213/insertSigcolumnsNum4096.json -y " % binPath)
+        tdSql.execute("use db")
+        tdSql.query("select count (tbname) from stb_old")
+        tdSql.checkData(0, 0, 10)
+
+        # tdSql.query("select * from stb_old")
+        # tdSql.checkRows(10)
+        # tdSql.checkCols(1024)
+            
+        # tdSql.query("select count (tbname) from stb_new")
+        # tdSql.checkData(0, 0, 10)
+
+        # tdSql.query("select * from stb_new")
+        # tdSql.checkRows(10)
+        # tdSql.checkCols(4096)
+        
+        # tdLog.info("stop dnode to commit data to disk")
+        # tdDnodes.stop(1)
+        # tdDnodes.start(1)     
+
+        #regular table
+        sql = "create table tb(ts timestamp, "
+        for i in range(1022):
+            sql += "col%d binary(14), " % (i + 1)
+        sql += "col1023 binary(22))"        
+        tdSql.execute(sql)
+
+        for i in range(4):
+            sql = "insert into tb values(%d, "
+            for j in range(1022):
+                str = "'%s', " % self.get_random_string(14)                
+                sql += str
+            sql += "'%s')" % self.get_random_string(22)
+            tdSql.execute(sql % (self.ts + i))
+
+        time.sleep(10)
+        tdSql.query("select count(*) from tb")
+        tdSql.checkData(0, 0, 4)
+
+        tdDnodes.stop(1)
+        tdDnodes.start(1)
+
+        time.sleep(1)
+        tdSql.query("select count(*) from tb")
+        tdSql.checkData(0, 0, 4)  
+
+
+        sql = "create table tb1(ts timestamp, "
+        for i in range(4094):
+            sql += "col%d binary(14), " % (i + 1)
+        sql += "col4095 binary(22))"        
+        tdSql.execute(sql)
+
+        for i in range(4):
+            sql = "insert into tb1 values(%d, "
+            for j in range(4094):
+                str = "'%s', " % self.get_random_string(14)                
+                sql += str
+            sql += "'%s')" % self.get_random_string(22)
+            tdSql.execute(sql % (self.ts + i))
+
+        time.sleep(10)
+        tdSql.query("select count(*) from tb1")
+        tdSql.checkData(0, 0, 4)
+
+        tdDnodes.stop(1)
+        tdDnodes.start(1)
+
+        time.sleep(1)
+        tdSql.query("select count(*) from tb1")
+        tdSql.checkData(0, 0, 4)  
+    
+
+
+        #os.system("rm -rf tools/taosdemoAllTest/TD-5213/insertSigcolumnsNum4096.py.sql")        
+        
+        
+        
+    def stop(self):
+        tdSql.close()
+        tdLog.success("%s successfully executed" % __file__)
+
+
+tdCases.addWindows(__file__, TDTestCase())
+tdCases.addLinux(__file__, TDTestCase())