diff --git a/include/common/tglobal.h b/include/common/tglobal.h
index 157e37f080..ced94e0f11 100644
--- a/include/common/tglobal.h
+++ b/include/common/tglobal.h
@@ -29,7 +29,6 @@ extern "C" {
 #define SLOW_LOG_TYPE_OTHERS 0x4
 #define SLOW_LOG_TYPE_ALL    0xFFFFFFFF
 
-
 // cluster
 extern char     tsFirst[];
 extern char     tsSecond[];
@@ -181,6 +180,7 @@ extern bool    tsDisableStream;
 extern int64_t tsStreamBufferSize;
 extern int64_t tsCheckpointInterval;
 extern bool    tsFilterScalarMode;
+extern int32_t tsMaxStreamBackendCache;
 
 // #define NEEDTO_COMPRESSS_MSG(size) (tsCompressMsgSize != -1 && (size) > tsCompressMsgSize)
 
diff --git a/source/common/src/tglobal.c b/source/common/src/tglobal.c
index c648f8551a..c5646d92d1 100644
--- a/source/common/src/tglobal.c
+++ b/source/common/src/tglobal.c
@@ -60,6 +60,7 @@ int32_t tsNumOfQnodeQueryThreads = 4;
 int32_t tsNumOfQnodeFetchThreads = 1;
 int32_t tsNumOfSnodeStreamThreads = 4;
 int32_t tsNumOfSnodeWriteThreads = 1;
+int32_t tsMaxStreamBackendCache = 128;  // M
 
 // sync raft
 int32_t tsElectInterval = 25 * 1000;
@@ -105,7 +106,7 @@ int32_t tsQueryPolicy = 1;
 int32_t tsQueryRspPolicy = 0;
 int64_t tsQueryMaxConcurrentTables = 200;  // unit is TSDB_TABLE_NUM_UNIT
 bool    tsEnableQueryHb = false;
-bool    tsEnableScience = false; // on taos-cli show float and doulbe with scientific notation if true 
+bool    tsEnableScience = false;  // on taos-cli show float and doulbe with scientific notation if true
 int32_t tsQuerySmaOptimize = 0;
 int32_t tsQueryRsmaTolerance = 1000;  // the tolerance time (ms) to judge from which level to query rsma data.
 bool    tsQueryPlannerTrace = false;
@@ -117,8 +118,8 @@ int32_t tsRedirectFactor = 2;
 int32_t tsRedirectMaxPeriod = 1000;
 int32_t tsMaxRetryWaitTime = 10000;
 bool    tsUseAdapter = false;
-int32_t tsMetaCacheMaxSize = -1; // MB
-int32_t tsSlowLogThreshold = 3; // seconds
+int32_t tsMetaCacheMaxSize = -1;  // MB
+int32_t tsSlowLogThreshold = 3;   // seconds
 int32_t tsSlowLogScope = SLOW_LOG_TYPE_ALL;
 
 /*
@@ -349,7 +350,8 @@ static int32_t taosAddClientCfg(SConfig *pCfg) {
   if (cfgAddInt32(pCfg, "maxRetryWaitTime", tsMaxRetryWaitTime, 0, 86400000, 0) != 0) return -1;
   if (cfgAddBool(pCfg, "useAdapter", tsUseAdapter, true) != 0) return -1;
   if (cfgAddBool(pCfg, "crashReporting", tsEnableCrashReport, true) != 0) return -1;
-  if (cfgAddInt64(pCfg, "queryMaxConcurrentTables", tsQueryMaxConcurrentTables, INT64_MIN, INT64_MAX, 1) != 0) return -1;
+  if (cfgAddInt64(pCfg, "queryMaxConcurrentTables", tsQueryMaxConcurrentTables, INT64_MIN, INT64_MAX, 1) != 0)
+    return -1;
   if (cfgAddInt32(pCfg, "metaCacheMaxSize", tsMetaCacheMaxSize, -1, INT32_MAX, 1) != 0) return -1;
   if (cfgAddInt32(pCfg, "slowLogThreshold", tsSlowLogThreshold, 0, INT32_MAX, true) != 0) return -1;
   if (cfgAddString(pCfg, "slowLogScope", "", true) != 0) return -1;
@@ -524,6 +526,7 @@ static int32_t taosAddServerCfg(SConfig *pCfg) {
   if (cfgAddInt32(pCfg, "cacheLazyLoadThreshold", tsCacheLazyLoadThreshold, 0, 100000, 0) != 0) return -1;
 
   if (cfgAddBool(pCfg, "filterScalarMode", tsFilterScalarMode, 0) != 0) return -1;
+  if (cfgAddInt32(pCfg, "maxStreamBackendCache", tsMaxStreamBackendCache, 16, 1024, 0) != 0) return -1;
 
   GRANT_CFG_ADD;
   return 0;
@@ -781,7 +784,7 @@ static int32_t taosSetClientCfg(SConfig *pCfg) {
   tsNumOfTaskQueueThreads = cfgGetItem(pCfg, "numOfTaskQueueThreads")->i32;
   tsQueryPolicy = cfgGetItem(pCfg, "queryPolicy")->i32;
   tsEnableQueryHb = cfgGetItem(pCfg, "enableQueryHb")->bval;
-  tsEnableScience = cfgGetItem(pCfg, "enableScience")->bval;  
+  tsEnableScience = cfgGetItem(pCfg, "enableScience")->bval;
   tsQuerySmaOptimize = cfgGetItem(pCfg, "querySmaOptimize")->i32;
   tsQueryPlannerTrace = cfgGetItem(pCfg, "queryPlannerTrace")->bval;
   tsQueryNodeChunkSize = cfgGetItem(pCfg, "queryNodeChunkSize")->i32;
@@ -902,7 +905,8 @@ static int32_t taosSetServerCfg(SConfig *pCfg) {
   tsCheckpointInterval = cfgGetItem(pCfg, "checkpointInterval")->i64;
 
   tsFilterScalarMode = cfgGetItem(pCfg, "filterScalarMode")->bval;
-  
+  tsMaxStreamBackendCache = cfgGetItem(pCfg, "maxStreamBackendCache")->i32;
+
   GRANT_CFG_GET;
   return 0;
 }
diff --git a/source/libs/executor/src/scanoperator.c b/source/libs/executor/src/scanoperator.c
index 2d2f377041..f11b083ad6 100644
--- a/source/libs/executor/src/scanoperator.c
+++ b/source/libs/executor/src/scanoperator.c
@@ -1883,7 +1883,7 @@ static SSDataBlock* doStreamScan(SOperatorInfo* pOperator) {
     if (pInfo->pRecoverRes != NULL) {
       pInfo->blockRecoverContiCnt++;
       calBlockTbName(pInfo, pInfo->pRecoverRes);
-      if (pInfo->pUpdateInfo) {
+      if (!pInfo->igCheckUpdate && pInfo->pUpdateInfo) {
         if (pTaskInfo->streamInfo.recoverStep == STREAM_RECOVER_STEP__SCAN1) {
           TSKEY maxTs = pAPI->stateStore.updateInfoFillBlockData(pInfo->pUpdateInfo, pInfo->pRecoverRes, pInfo->primaryTsIndex);
           pInfo->twAggSup.maxTs = TMAX(pInfo->twAggSup.maxTs, maxTs);
diff --git a/source/libs/executor/src/timewindowoperator.c b/source/libs/executor/src/timewindowoperator.c
index aa0aa9799b..2cdf9bb15c 100644
--- a/source/libs/executor/src/timewindowoperator.c
+++ b/source/libs/executor/src/timewindowoperator.c
@@ -1623,7 +1623,7 @@ void initIntervalDownStream(SOperatorInfo* downstream, uint16_t type, SStreamInt
   SStreamScanInfo* pScanInfo = downstream->info;
   pScanInfo->windowSup.parentType = type;
   pScanInfo->windowSup.pIntervalAggSup = &pInfo->aggSup;
-  if (!pScanInfo->igCheckUpdate && !pScanInfo->pUpdateInfo) {
+  if (!pScanInfo->pUpdateInfo) {
     pScanInfo->pUpdateInfo = pAPI->updateInfoInitP(&pInfo->interval, pInfo->twAggSup.waterMark);
   }
 
@@ -2150,28 +2150,29 @@ static void doBuildPullDataBlock(SArray* array, int32_t* pIndex, SSDataBlock* pB
 }
 
 void processPullOver(SSDataBlock* pBlock, SHashObj* pMap, SInterval* pInterval) {
-  SColumnInfoData* pStartCol = taosArrayGet(pBlock->pDataBlock, START_TS_COLUMN_INDEX);
+  SColumnInfoData* pStartCol = taosArrayGet(pBlock->pDataBlock, CALCULATE_START_TS_COLUMN_INDEX);
   TSKEY*           tsData = (TSKEY*)pStartCol->pData;
-  SColumnInfoData* pEndCol = taosArrayGet(pBlock->pDataBlock, END_TS_COLUMN_INDEX);
+  SColumnInfoData* pEndCol = taosArrayGet(pBlock->pDataBlock, CALCULATE_END_TS_COLUMN_INDEX);
   TSKEY*           tsEndData = (TSKEY*)pEndCol->pData;
   SColumnInfoData* pGroupCol = taosArrayGet(pBlock->pDataBlock, GROUPID_COLUMN_INDEX);
   uint64_t*        groupIdData = (uint64_t*)pGroupCol->pData;
   int32_t          chId = getChildIndex(pBlock);
   for (int32_t i = 0; i < pBlock->info.rows; i++) {
     TSKEY winTs = tsData[i];
-    while (winTs < tsEndData[i]) {
+    while (winTs <= tsEndData[i]) {
       SWinKey winRes = {.ts = winTs, .groupId = groupIdData[i]};
       void*   chIds = taosHashGet(pMap, &winRes, sizeof(SWinKey));
       if (chIds) {
         SArray* chArray = *(SArray**)chIds;
         int32_t index = taosArraySearchIdx(chArray, &chId, compareInt32Val, TD_EQ);
         if (index != -1) {
-          qDebug("===stream===window %" PRId64 " delete child id %d", winRes.ts, chId);
+          qDebug("===stream===retrive window %" PRId64 " delete child id %d", winRes.ts, chId);
           taosArrayRemove(chArray, index);
           if (taosArrayGetSize(chArray) == 0) {
             // pull data is over
             taosArrayDestroy(chArray);
             taosHashRemove(pMap, &winRes, sizeof(SWinKey));
+            qDebug("===stream===retrive pull data over.window %" PRId64 , winRes.ts);
           }
         }
       }
diff --git a/source/libs/stream/src/stream.c b/source/libs/stream/src/stream.c
index d3e4b23ad1..8bca0e6ec2 100644
--- a/source/libs/stream/src/stream.c
+++ b/source/libs/stream/src/stream.c
@@ -132,8 +132,6 @@ int32_t streamTaskEnqueueBlocks(SStreamTask* pTask, const SStreamDispatchReq* pR
     int32_t code = tAppendDataToInputQueue(pTask, (SStreamQueueItem*)pBlock);
     // input queue is full, upstream is blocked now
     status = (code == TSDB_CODE_SUCCESS)? TASK_INPUT_STATUS__NORMAL:TASK_INPUT_STATUS__BLOCKED;
-
-
   }
 
   // rsp by input status
@@ -235,8 +233,7 @@ int32_t streamProcessDispatchMsg(SStreamTask* pTask, SStreamDispatchReq* pReq, S
 }
 
 int32_t streamProcessDispatchRsp(SStreamTask* pTask, SStreamDispatchRsp* pRsp, int32_t code) {
-  ASSERT(pRsp->inputStatus == TASK_OUTPUT_STATUS__NORMAL || pRsp->inputStatus == TASK_OUTPUT_STATUS__BLOCKED);
-  qDebug("s-task:%s receive dispatch rsp, code: %x", pTask->id.idStr, code);
+  qDebug("s-task:%s receive dispatch rsp, status:%d code:%d", pTask->id.idStr, pRsp->inputStatus, code);
 
   if (pTask->outputType == TASK_OUTPUT__SHUFFLE_DISPATCH) {
     int32_t leftRsp = atomic_sub_fetch_32(&pTask->shuffleDispatcher.waitingRspCnt, 1);
@@ -248,13 +245,16 @@ int32_t streamProcessDispatchRsp(SStreamTask* pTask, SStreamDispatchRsp* pRsp, i
 
   int8_t old = atomic_exchange_8(&pTask->outputStatus, pRsp->inputStatus);
   ASSERT(old == TASK_OUTPUT_STATUS__WAIT);
+
+  // the input queue of the (down stream) task that receive the output data is full, so the TASK_INPUT_STATUS_BLOCKED is rsp
+  // todo we need to send EMPTY PACKAGE to detect if the input queue is available for output of upstream task, every 50 ms.
   if (pRsp->inputStatus == TASK_INPUT_STATUS__BLOCKED) {
     // TODO: init recover timer
-    ASSERT(0);
+    qError("s-task:%s inputQ of downstream task:0x%x is full, need to block output", pTask->id.idStr, pRsp->downstreamTaskId);
     return 0;
   }
 
-  // continue dispatch one block to down stream in pipeline
+  // otherwise, continue dispatch the first block to down stream task in pipeline
   streamDispatchStreamBlock(pTask);
   return 0;
 }
@@ -304,23 +304,32 @@ int32_t tAppendDataToInputQueue(SStreamTask* pTask, SStreamQueueItem* pItem) {
       return -1;
     }
 
-    taosWriteQitem(pTask->inputQueue->queue, pItem);
+    int32_t code = taosWriteQitem(pTask->inputQueue->queue, pItem);
+    if (code != TSDB_CODE_SUCCESS) {
+      streamDataSubmitDestroy(px);
+      taosFreeQitem(pItem);
+      return code;
+    }
   } else if (type == STREAM_INPUT__DATA_BLOCK || type == STREAM_INPUT__DATA_RETRIEVE ||
              type == STREAM_INPUT__REF_DATA_BLOCK) {
-    if ((pTask->taskLevel == TASK_LEVEL__SOURCE) && (tInputQueueIsFull(pTask))) {
+    if (/*(pTask->taskLevel == TASK_LEVEL__SOURCE) && */(tInputQueueIsFull(pTask))) {
       qError("s-task:%s input queue is full, capacity:%d size:%d MiB, current(blocks:%d, size:%.2fMiB) abort",
              pTask->id.idStr, STREAM_TASK_INPUT_QUEUEU_CAPACITY, STREAM_TASK_INPUT_QUEUEU_CAPACITY_IN_SIZE, total,
              size);
       destroyStreamDataBlock((SStreamDataBlock*) pItem);
-      taosFreeQitem(pItem);
       return -1;
     }
 
     qDebug("s-task:%s data block enqueue, current(blocks:%d, size:%.2fMiB)", pTask->id.idStr, total, size);
-    taosWriteQitem(pTask->inputQueue->queue, pItem);
+    int32_t code = taosWriteQitem(pTask->inputQueue->queue, pItem);
+    if (code != TSDB_CODE_SUCCESS) {
+      destroyStreamDataBlock((SStreamDataBlock*) pItem);
+      return code;
+    }
   } else if (type == STREAM_INPUT__CHECKPOINT) {
     taosWriteQitem(pTask->inputQueue->queue, pItem);
   } else if (type == STREAM_INPUT__GET_RES) {
+    // use the default memory limit, refactor later.
     taosWriteQitem(pTask->inputQueue->queue, pItem);
     qDebug("s-task:%s data res enqueue, current(blocks:%d, size:%.2fMiB)", pTask->id.idStr, total, size);
   }
diff --git a/source/libs/stream/src/streamBackendRocksdb.c b/source/libs/stream/src/streamBackendRocksdb.c
index f7638a42ae..c8a6597bad 100644
--- a/source/libs/stream/src/streamBackendRocksdb.c
+++ b/source/libs/stream/src/streamBackendRocksdb.c
@@ -38,6 +38,15 @@ typedef struct {
   rocksdb_comparator_t**           pCompares;
 } RocksdbCfInst;
 
+uint32_t nextPow2(uint32_t x) {
+  x = x - 1;
+  x = x | (x >> 1);
+  x = x | (x >> 2);
+  x = x | (x >> 4);
+  x = x | (x >> 8);
+  x = x | (x >> 16);
+  return x + 1;
+}
 int32_t streamStateOpenBackendCf(void* backend, char* name, char** cfs, int32_t nCf);
 
 void destroyRocksdbCfInst(RocksdbCfInst* inst);
@@ -92,6 +101,8 @@ void* streamBackendInit(const char* path) {
   rocksdb_options_set_recycle_log_file_num(opts, 6);
   rocksdb_options_set_max_write_buffer_number(opts, 2);
   rocksdb_options_set_info_log_level(opts, 0);
+  uint32_t dbLimit = nextPow2(tsMaxStreamBackendCache);
+  rocksdb_options_set_db_write_buffer_size(opts, dbLimit << 20);
 
   pHandle->env = env;
   pHandle->dbOpt = opts;
diff --git a/source/libs/stream/src/streamState.c b/source/libs/stream/src/streamState.c
index ba24a581e9..71a21ac150 100644
--- a/source/libs/stream/src/streamState.c
+++ b/source/libs/stream/src/streamState.c
@@ -1062,7 +1062,7 @@ _end:
 }
 
 int32_t streamStatePutParName(SStreamState* pState, int64_t groupId, const char tbname[TSDB_TABLE_NAME_LEN]) {
-  qWarn("try to write to cf parname");
+  qDebug("try to write to cf parname");
 #ifdef USE_ROCKSDB
   if (tSimpleHashGetSize(pState->parNameMap) > MAX_TABLE_NAME_NUM) {
     if (tSimpleHashGet(pState->parNameMap, &groupId, sizeof(int64_t)) == NULL) {