Merge pull request #6872 from taosdata/feature/query

[td-5037]<fix>: fix the last query performance worse then previous pr…

src/client/src/tscServer.c

@@ -13,7 +13,10 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
+#include <tscompression.h>
 #include "os.h"
+#include "qPlan.h"
+#include "qTableMeta.h"
 #include "tcmdtype.h"
 #include "tlockfree.h"
 #include "trpc.h"
@@ -21,10 +24,8 @@
 #include "tscLog.h"
 #include "tscProfile.h"
 #include "tscUtil.h"
-#include "qTableMeta.h"
 #include "tsclient.h"
 #include "ttimer.h"
-#include "qPlan.h"
 
 int (*tscBuildMsg[TSDB_SQL_MAX])(SSqlObj *pSql, SSqlInfo *pInfo) = {0};
 
@@ -2048,16 +2049,27 @@ int tscProcessMultiTableMetaRsp(SSqlObj *pSql) {
   }
 
   SSqlCmd *pParentCmd = &pParentSql->cmd;
-
   SHashObj *pSet = taosHashInit(4, taosGetDefaultHashFunction(TSDB_DATA_TYPE_BIGINT), false, HASH_NO_LOCK);
 
   char* pMsg = pMultiMeta->meta;
+  char* buf = NULL;
+  if (pMultiMeta->compressed) {
+    buf = malloc(pMultiMeta->rawLen - sizeof(SMultiTableMeta));
+    int32_t len = tsDecompressString(pMultiMeta->meta, pMultiMeta->contLen - sizeof(SMultiTableMeta), 1,
+        buf, pMultiMeta->rawLen - sizeof(SMultiTableMeta), ONE_STAGE_COMP, NULL, 0);
+    assert(len == pMultiMeta->rawLen - sizeof(SMultiTableMeta));
+
+    pMsg = buf;
+  }
+
   for (int32_t i = 0; i < pMultiMeta->numOfTables; i++) {
     STableMetaMsg *pMetaMsg = (STableMetaMsg *)pMsg;
     int32_t code = tableMetaMsgConvert(pMetaMsg);
     if (code != TSDB_CODE_SUCCESS) {
       taosHashCleanup(pSet);
       taosReleaseRef(tscObjRef, pParentSql->self);
+
+      tfree(buf);
       return code;
     }
 
@@ -2066,6 +2078,8 @@ int tscProcessMultiTableMetaRsp(SSqlObj *pSql) {
       tscError("0x%"PRIx64" invalid table meta from mnode, name:%s", pSql->self, pMetaMsg->tableFname);
       taosHashCleanup(pSet);
       taosReleaseRef(tscObjRef, pParentSql->self);
+
+      tfree(buf);
       return TSDB_CODE_TSC_INVALID_VALUE;
     }
 
@@ -2115,6 +2129,8 @@ int tscProcessMultiTableMetaRsp(SSqlObj *pSql) {
 
   taosHashCleanup(pSet);
   taosReleaseRef(tscObjRef, pParentSql->self);
+
+  tfree(buf);
   return TSDB_CODE_SUCCESS;
 }
 

src/inc/taosmsg.h

@@ -762,7 +762,9 @@ typedef struct STableMetaMsg {
 typedef struct SMultiTableMeta {
   int32_t  numOfTables;
   int32_t  numOfVgroup;
-  int32_t       contLen;
+  uint32_t contLen:31;
+  uint8_t  compressed:1;      // denote if compressed or not
+  uint32_t rawLen;            // size before compress
   char     meta[];
 } SMultiTableMeta;
 

src/mnode/src/mnodeTable.c

@@ -2892,7 +2892,7 @@ static SMultiTableMeta* ensureMsgBufferSpace(SMultiTableMeta *pMultiMeta, SArray
       (*totalMallocLen) *= 2;
     }
 
-    pMultiMeta = rpcReallocCont(pMultiMeta, *totalMallocLen);
+    pMultiMeta = realloc(pMultiMeta, *totalMallocLen);
     if (pMultiMeta == NULL) {
       return NULL;
     }
@@ -2923,8 +2923,8 @@ static int32_t mnodeProcessMultiTableMetaMsg(SMnodeMsg *pMsg) {
   }
 
   // first malloc 80KB, subsequent reallocation will expand the size as twice of the original size
-  int32_t totalMallocLen = sizeof(STableMetaMsg) + sizeof(SSchema) * (TSDB_MAX_TAGS + TSDB_MAX_COLUMNS + 16);
-  pMultiMeta = rpcMallocCont(totalMallocLen);
+  int32_t totalMallocLen = sizeof(SMultiTableMeta) + sizeof(STableMetaMsg) + sizeof(SSchema) * (TSDB_MAX_TAGS + TSDB_MAX_COLUMNS + 16);
+  pMultiMeta = calloc(1, totalMallocLen);
   if (pMultiMeta == NULL) {
     code = TSDB_CODE_MND_OUT_OF_MEMORY;
     goto _end;
@@ -2957,7 +2957,7 @@ static int32_t mnodeProcessMultiTableMetaMsg(SMnodeMsg *pMsg) {
     int remain = totalMallocLen - pMultiMeta->contLen;
     if (remain <= sizeof(STableMetaMsg) + sizeof(SSchema) * (TSDB_MAX_TAGS + TSDB_MAX_COLUMNS + 16)) {
       totalMallocLen *= 2;
-      pMultiMeta = rpcReallocCont(pMultiMeta, totalMallocLen);
+      pMultiMeta = realloc(pMultiMeta, totalMallocLen);
       if (pMultiMeta == NULL) {
         mnodeDecTableRef(pMsg->pTable);
         code = TSDB_CODE_MND_OUT_OF_MEMORY;
@@ -3027,16 +3027,41 @@ static int32_t mnodeProcessMultiTableMetaMsg(SMnodeMsg *pMsg) {
   pMsg->rpcRsp.len = pMultiMeta->contLen;
   code = TSDB_CODE_SUCCESS;
 
+  char* tmp = rpcMallocCont(pMultiMeta->contLen + 2);
+  if (tmp == NULL) {
+    code = TSDB_CODE_MND_OUT_OF_MEMORY;
+    goto _end;
+  }
+
+  int32_t len = tsCompressString(pMultiMeta->meta, (int32_t)pMultiMeta->contLen - sizeof(SMultiTableMeta), 1,
+      tmp + sizeof(SMultiTableMeta), (int32_t)pMultiMeta->contLen - sizeof(SMultiTableMeta) + 2, ONE_STAGE_COMP, NULL, 0);
+
+  pMultiMeta->rawLen = pMultiMeta->contLen;
+  if (len == -1 || len + sizeof(SMultiTableMeta) >= pMultiMeta->contLen + 2) { // compress failed, do not compress this binary data
+    pMultiMeta->compressed = 0;
+    memcpy(tmp, pMultiMeta, sizeof(SMultiTableMeta) + pMultiMeta->contLen);
+  } else {
+    pMultiMeta->compressed = 1;
+    pMultiMeta->contLen = sizeof(SMultiTableMeta) + len;
+
+    // copy the header and the compressed payload
+    memcpy(tmp, pMultiMeta, sizeof(SMultiTableMeta));
+  }
+
+  pMsg->rpcRsp.rsp = tmp;
+  pMsg->rpcRsp.len = pMultiMeta->contLen;
+
+  SMultiTableMeta* p = (SMultiTableMeta*) tmp;
+
+  mDebug("multiTable info build completed, original:%d, compressed:%d, comp:%d", p->rawLen, p->contLen, p->compressed);
+
   _end:
   tfree(str);
   tfree(nameList);
   taosArrayDestroy(pList);
   pMsg->pTable  = NULL;
   pMsg->pVgroup = NULL;
-
-  if (code != TSDB_CODE_SUCCESS) {
-    rpcFreeCont(pMultiMeta);
-  }
+  tfree(pMultiMeta);
 
   return code;
 }

src/query/inc/qExecutor.h

@@ -589,7 +589,7 @@ int32_t createIndirectQueryFuncExprFromMsg(SQueryTableMsg *pQueryMsg, int32_t nu
 
 SGroupbyExpr *createGroupbyExprFromMsg(SQueryTableMsg *pQueryMsg, SColIndex *pColIndex, int32_t *code);
 SQInfo *createQInfoImpl(SQueryTableMsg *pQueryMsg, SGroupbyExpr *pGroupbyExpr, SExprInfo *pExprs,
-                        SExprInfo *pSecExprs, STableGroupInfo *pTableGroupInfo, SColumnInfo* pTagCols, int32_t vgId, char* sql, uint64_t *qId);
+                        SExprInfo *pSecExprs, STableGroupInfo *pTableGroupInfo, SColumnInfo* pTagCols, int32_t vgId, char* sql, uint64_t qId);
 
 int32_t initQInfo(STsBufInfo* pTsBufInfo, void* tsdb, void* sourceOptr, SQInfo* pQInfo, SQueryParam* param, char* start,
                   int32_t prevResultLen, void* merger);

src/query/src/qExecutor.c

@@ -2268,10 +2268,11 @@ static int32_t updateBlockLoadStatus(SQueryAttr *pQuery, int32_t status) {
   return status;
 }
 
-static void doExchangeTimeWindow(SQInfo* pQInfo, STimeWindow* win) {
-  SQueryAttr* pQueryAttr = &pQInfo->query;
-  size_t t = taosArrayGetSize(pQueryAttr->tableGroupInfo.pGroupList);
-  for(int32_t i = 0; i < t; ++i) {
+static void doUpdateLastKey(SQueryAttr* pQueryAttr) {
+  STimeWindow* win = &pQueryAttr->window;
+
+  size_t num = taosArrayGetSize(pQueryAttr->tableGroupInfo.pGroupList);
+  for(int32_t i = 0; i < num; ++i) {
     SArray* p1 = taosArrayGetP(pQueryAttr->tableGroupInfo.pGroupList, i);
 
     size_t len = taosArrayGetSize(p1);
@@ -2286,7 +2287,7 @@ static void doExchangeTimeWindow(SQInfo* pQInfo, STimeWindow* win) {
   }
 }
 
-static void changeExecuteScanOrder(SQInfo *pQInfo, SQueryTableMsg* pQueryMsg, bool stableQuery) {
+static void updateDataCheckOrder(SQInfo *pQInfo, SQueryTableMsg* pQueryMsg, bool stableQuery) {
   SQueryAttr* pQueryAttr = pQInfo->runtimeEnv.pQueryAttr;
 
   // in case of point-interpolation query, use asc order scan
@@ -2303,6 +2304,7 @@ static void changeExecuteScanOrder(SQInfo *pQInfo, SQueryTableMsg* pQueryMsg, bo
       SWAP(pQueryAttr->window.skey, pQueryAttr->window.ekey, TSKEY);
     }
 
+    pQueryAttr->needReverseScan = false;
     return;
   }
 
@@ -2312,7 +2314,8 @@ static void changeExecuteScanOrder(SQInfo *pQInfo, SQueryTableMsg* pQueryMsg, bo
       SWAP(pQueryAttr->window.skey, pQueryAttr->window.ekey, TSKEY);
     }
 
-    doExchangeTimeWindow(pQInfo, &pQueryAttr->window);
+    pQueryAttr->needReverseScan = false;
+    doUpdateLastKey(pQueryAttr);
     return;
   }
 
@@ -2333,20 +2336,22 @@ static void changeExecuteScanOrder(SQInfo *pQInfo, SQueryTableMsg* pQueryMsg, bo
                pQueryAttr->window.ekey, pQueryAttr->window.ekey, pQueryAttr->window.skey);
 
         SWAP(pQueryAttr->window.skey, pQueryAttr->window.ekey, TSKEY);
-        doExchangeTimeWindow(pQInfo, &pQueryAttr->window);
+        doUpdateLastKey(pQueryAttr);
       }
 
       pQueryAttr->order.order = TSDB_ORDER_ASC;
+      pQueryAttr->needReverseScan = false;
     } else if (onlyLastQuery(pQueryAttr) && notContainSessionOrStateWindow(pQueryAttr)) {
       if (QUERY_IS_ASC_QUERY(pQueryAttr)) {
         qDebug(msg, pQInfo, "only-last", pQueryAttr->order.order, TSDB_ORDER_DESC, pQueryAttr->window.skey,
                pQueryAttr->window.ekey, pQueryAttr->window.ekey, pQueryAttr->window.skey);
 
         SWAP(pQueryAttr->window.skey, pQueryAttr->window.ekey, TSKEY);
-        doExchangeTimeWindow(pQInfo, &pQueryAttr->window);
+        doUpdateLastKey(pQueryAttr);
       }
 
       pQueryAttr->order.order = TSDB_ORDER_DESC;
+      pQueryAttr->needReverseScan = false;
     }
 
   } else {  // interval query
@@ -2357,20 +2362,22 @@ static void changeExecuteScanOrder(SQInfo *pQInfo, SQueryTableMsg* pQueryMsg, bo
                  pQueryAttr->window.skey, pQueryAttr->window.ekey, pQueryAttr->window.ekey, pQueryAttr->window.skey);
 
           SWAP(pQueryAttr->window.skey, pQueryAttr->window.ekey, TSKEY);
-          doExchangeTimeWindow(pQInfo, &pQueryAttr->window);
+          doUpdateLastKey(pQueryAttr);
         }
 
         pQueryAttr->order.order = TSDB_ORDER_ASC;
+        pQueryAttr->needReverseScan = false;
       } else if (onlyLastQuery(pQueryAttr)) {
         if (QUERY_IS_ASC_QUERY(pQueryAttr)) {
           qDebug(msg, pQInfo, "only-last stable", pQueryAttr->order.order, TSDB_ORDER_DESC,
                  pQueryAttr->window.skey, pQueryAttr->window.ekey, pQueryAttr->window.ekey, pQueryAttr->window.skey);
 
           SWAP(pQueryAttr->window.skey, pQueryAttr->window.ekey, TSKEY);
-          doExchangeTimeWindow(pQInfo, &pQueryAttr->window);
+          doUpdateLastKey(pQueryAttr);
         }
 
         pQueryAttr->order.order = TSDB_ORDER_DESC;
+        pQueryAttr->needReverseScan = false;
       }
     }
   }
@@ -2388,9 +2395,6 @@ static void getIntermediateBufInfo(SQueryRuntimeEnv* pRuntimeEnv, int32_t* ps, i
   while(((*rowsize) * MIN_ROWS_PER_PAGE) > (*ps) - overhead) {
     *ps = ((*ps) << 1u);
   }
-
-//  pRuntimeEnv->numOfRowsPerPage = ((*ps) - sizeof(tFilePage)) / (*rowsize);
-//  assert(pRuntimeEnv->numOfRowsPerPage <= MAX_ROWS_PER_RESBUF_PAGE);
 }
 
 #define IS_PREFILTER_TYPE(_t) ((_t) != TSDB_DATA_TYPE_BINARY && (_t) != TSDB_DATA_TYPE_NCHAR)
@@ -4382,7 +4386,7 @@ int32_t doInitQInfo(SQInfo* pQInfo, STSBuf* pTsBuf, void* tsdb, void* sourceOptr
       break;
     }
     case OP_DataBlocksOptScan: {
-      pRuntimeEnv->proot = createDataBlocksOptScanInfo(pRuntimeEnv->pQueryHandle, pRuntimeEnv, getNumOfScanTimes(pQueryAttr), 1);
+      pRuntimeEnv->proot = createDataBlocksOptScanInfo(pRuntimeEnv->pQueryHandle, pRuntimeEnv, getNumOfScanTimes(pQueryAttr), pQueryAttr->needReverseScan? 1:0);
       break;
     }
     case OP_TableScan: {
@@ -4420,8 +4424,10 @@ int32_t doInitQInfo(SQInfo* pQInfo, STSBuf* pTsBuf, void* tsdb, void* sourceOptr
   if (pQInfo->summary.queryProfEvents == NULL) {
     qDebug("QInfo:0x%"PRIx64" failed to allocate query prof events array", pQInfo->qId);
   }
+
   pQInfo->summary.operatorProfResults =
       taosHashInit(8, taosGetDefaultHashFunction(TSDB_DATA_TYPE_TINYINT), true, HASH_NO_LOCK);
+
   if (pQInfo->summary.operatorProfResults == NULL) {
     qDebug("QInfo:0x%"PRIx64" failed to allocate operator prof results hash", pQInfo->qId);
   }
@@ -4814,7 +4820,6 @@ SOperatorInfo* createDataBlocksOptScanInfo(void* pTsdbQueryHandle, SQueryRuntime
   pInfo->reverseTimes = reverseTime;
   pInfo->current      = 0;
   pInfo->order        = pRuntimeEnv->pQueryAttr->order.order;
-//  pInfo->prevGroupId  = -1;
 
   SOperatorInfo* pOptr = calloc(1, sizeof(SOperatorInfo));
   pOptr->name          = "DataBlocksOptimizedScanOperator";
@@ -7366,7 +7371,7 @@ FORCE_INLINE bool checkQIdEqual(void *qHandle, uint64_t qId) {
 
 SQInfo* createQInfoImpl(SQueryTableMsg* pQueryMsg, SGroupbyExpr* pGroupbyExpr, SExprInfo* pExprs,
                         SExprInfo* pSecExprs, STableGroupInfo* pTableGroupInfo, SColumnInfo* pTagCols, int32_t vgId,
-                        char* sql, uint64_t *qId) {
+                        char* sql, uint64_t qId) {
   int16_t numOfCols = pQueryMsg->numOfCols;
   int16_t numOfOutput = pQueryMsg->numOfOutput;
 
@@ -7375,7 +7380,7 @@ SQInfo* createQInfoImpl(SQueryTableMsg* pQueryMsg, SGroupbyExpr* pGroupbyExpr, S
     goto _cleanup_qinfo;
   }
 
-  pQInfo->qId = *qId;
+  pQInfo->qId = qId;
 
   // to make sure third party won't overwrite this structure
   pQInfo->signature = pQInfo;
@@ -7485,7 +7490,7 @@ SQInfo* createQInfoImpl(SQueryTableMsg* pQueryMsg, SGroupbyExpr* pGroupbyExpr, S
   tsem_init(&pQInfo->ready, 0, 0);
 
   pQueryAttr->window = pQueryMsg->window;
-  changeExecuteScanOrder(pQInfo, pQueryMsg, pQueryAttr->stableQuery);
+  updateDataCheckOrder(pQInfo, pQueryMsg, pQueryAttr->stableQuery);
 
   SQueryRuntimeEnv* pRuntimeEnv = &pQInfo->runtimeEnv;
   STimeWindow window = pQueryAttr->window;

src/query/src/queryMain.c

@@ -162,7 +162,7 @@ int32_t qCreateQueryInfo(void* tsdb, int32_t vgId, SQueryTableMsg* pQueryMsg, qi
 
   assert(pQueryMsg->stableQuery == isSTableQuery);
   (*pQInfo) = createQInfoImpl(pQueryMsg, param.pGroupbyExpr, param.pExprs, param.pSecExprs, &tableGroupInfo,
-                              param.pTagColumnInfo, vgId, param.sql, qId);
+                              param.pTagColumnInfo, vgId, param.sql, *qId);
 
   param.sql    = NULL;
   param.pExprs = NULL;

src/tsdb/src/tsdbRead.c

@@ -98,6 +98,8 @@ typedef struct SIOCostSummary {
   int64_t blockLoadTime;
   int64_t statisInfoLoadTime;
   int64_t checkForNextTime;
+  int64_t headFileLoad;
+  int64_t headFileLoadTime;
 } SIOCostSummary;
 
 typedef struct STsdbQueryHandle {
@@ -1045,6 +1047,9 @@ static int32_t getFileCompInfo(STsdbQueryHandle* pQueryHandle, int32_t* numOfBlo
   int32_t code = TSDB_CODE_SUCCESS;
   *numOfBlocks = 0;
 
+  pQueryHandle->cost.headFileLoad += 1;
+  int64_t s = taosGetTimestampUs();
+
   size_t numOfTables = 0;
   if (pQueryHandle->loadType == BLOCK_LOAD_TABLE_SEQ_ORDER) {
     code = loadBlockInfo(pQueryHandle, pQueryHandle->activeIndex, numOfBlocks);
@@ -1054,6 +1059,9 @@ static int32_t getFileCompInfo(STsdbQueryHandle* pQueryHandle, int32_t* numOfBlo
     for (int32_t i = 0; i < numOfTables; ++i) {
       code = loadBlockInfo(pQueryHandle, i, numOfBlocks);
       if (code != TSDB_CODE_SUCCESS) {
+        int64_t e = taosGetTimestampUs();
+
+        pQueryHandle->cost.headFileLoadTime += (e - s);
         return code;
       }
     }
@@ -1061,6 +1069,8 @@ static int32_t getFileCompInfo(STsdbQueryHandle* pQueryHandle, int32_t* numOfBlo
     assert(0);
   }
 
+  int64_t e = taosGetTimestampUs();
+  pQueryHandle->cost.headFileLoadTime += (e - s);
   return code;
 }
 
@@ -3731,8 +3741,8 @@ void tsdbCleanupQueryHandle(TsdbQueryHandleT queryHandle) {
   pQueryHandle->next = doFreeColumnInfoData(pQueryHandle->next);
 
   SIOCostSummary* pCost = &pQueryHandle->cost;
-  tsdbDebug("%p :io-cost summary: statis-info:%"PRId64" us, datablock:%" PRId64" us, check data:%"PRId64" us, 0x%"PRIx64,
-      pQueryHandle, pCost->statisInfoLoadTime, pCost->blockLoadTime, pCost->checkForNextTime, pQueryHandle->qId);
+  tsdbDebug("%p :io-cost summary: head-file read cnt:%"PRIu64", head-file time:%"PRIu64" us, statis-info:%"PRId64" us, datablock:%" PRId64" us, check data:%"PRId64" us, 0x%"PRIx64,
+      pQueryHandle, pCost->headFileLoad, pCost->headFileLoadTime, pCost->statisInfoLoadTime, pCost->blockLoadTime, pCost->checkForNextTime, pQueryHandle->qId);
 
   tfree(pQueryHandle);
 }