homework-jianmu/source/libs/stream/src/streamExec.c

890 lines
33 KiB
C

/*
* Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
*
* This program is free software: you can use, redistribute, and/or modify
* it under the terms of the GNU Affero General Public License, version 3
* or later ("AGPL"), as published by the Free Software Foundation.
*
* This program is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE.
*
* You should have received a copy of the GNU Affero General Public License
* along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#include "streamInt.h"
// maximum allowed processed block batches. One block may include several submit blocks
#define MAX_STREAM_EXEC_BATCH_NUM 32
#define STREAM_RESULT_DUMP_THRESHOLD 300
#define STREAM_RESULT_DUMP_SIZE_THRESHOLD (1048576 * 1) // 1MiB result data
#define STREAM_SCAN_HISTORY_TIMESLICE 1000 // 1000 ms
#define MIN_INVOKE_INTERVAL 50 // 50ms
#define FILL_HISTORY_TASK_EXEC_INTERVAL 5000 // 5 sec
static int32_t streamTransferStateDoPrepare(SStreamTask* pTask);
static void streamTaskExecImpl(SStreamTask* pTask, SStreamQueueItem* pItem, int64_t* totalSize, int32_t* totalBlocks);
bool streamTaskShouldStop(const SStreamTask* pTask) {
SStreamTaskState pState = streamTaskGetStatus(pTask);
return (pState.state == TASK_STATUS__STOP) || (pState.state == TASK_STATUS__DROPPING);
}
bool streamTaskShouldPause(const SStreamTask* pTask) {
return (streamTaskGetStatus(pTask).state == TASK_STATUS__PAUSE);
}
static int32_t doOutputResultBlockImpl(SStreamTask* pTask, SStreamDataBlock* pBlock) {
int32_t code = 0;
int32_t type = pTask->outputInfo.type;
if (type == TASK_OUTPUT__TABLE) {
pTask->outputInfo.tbSink.tbSinkFunc(pTask, pTask->outputInfo.tbSink.vnode, pBlock->blocks);
destroyStreamDataBlock(pBlock);
} else if (type == TASK_OUTPUT__SMA) {
pTask->outputInfo.smaSink.smaSink(pTask->outputInfo.smaSink.vnode, pTask->outputInfo.smaSink.smaId, pBlock->blocks);
destroyStreamDataBlock(pBlock);
} else {
ASSERT(type == TASK_OUTPUT__FIXED_DISPATCH || type == TASK_OUTPUT__SHUFFLE_DISPATCH);
code = streamTaskPutDataIntoOutputQ(pTask, pBlock);
if (code != TSDB_CODE_SUCCESS) {
destroyStreamDataBlock(pBlock);
return code;
}
// not handle error, if dispatch failed, try next time.
// checkpoint trigger will be checked
(void)streamDispatchStreamBlock(pTask);
}
return code;
}
static int32_t doDumpResult(SStreamTask* pTask, SStreamQueueItem* pItem, SArray* pRes, int32_t size, int64_t* totalSize,
int32_t* totalBlocks) {
int32_t numOfBlocks = taosArrayGetSize(pRes);
if (numOfBlocks == 0) {
taosArrayDestroyEx(pRes, (FDelete)blockDataFreeRes);
return TSDB_CODE_SUCCESS;
}
SStreamDataBlock* pStreamBlocks = NULL;
int32_t code = createStreamBlockFromResults(pItem, pTask, size, pRes, &pStreamBlocks);
if (code) {
stError("s-task:%s failed to create result stream data block, code:%s", pTask->id.idStr, tstrerror(terrno));
taosArrayDestroyEx(pRes, (FDelete)blockDataFreeRes);
return TSDB_CODE_OUT_OF_MEMORY;
}
stDebug("s-task:%s dump stream result data blocks, num:%d, size:%.2fMiB", pTask->id.idStr, numOfBlocks,
SIZE_IN_MiB(size));
code = doOutputResultBlockImpl(pTask, pStreamBlocks);
if (code != TSDB_CODE_SUCCESS) { // back pressure and record position
return code;
}
*totalSize += size;
*totalBlocks += numOfBlocks;
return code;
}
void streamTaskExecImpl(SStreamTask* pTask, SStreamQueueItem* pItem, int64_t* totalSize, int32_t* totalBlocks) {
int32_t code = TSDB_CODE_SUCCESS;
void* pExecutor = pTask->exec.pExecutor;
*totalBlocks = 0;
*totalSize = 0;
int32_t size = 0;
int32_t numOfBlocks = 0;
SArray* pRes = NULL;
while (1) {
if (pRes == NULL) {
pRes = taosArrayInit(4, sizeof(SSDataBlock));
}
if (streamTaskShouldStop(pTask)) {
taosArrayDestroyEx(pRes, (FDelete)blockDataFreeRes);
return;
}
SSDataBlock* output = NULL;
uint64_t ts = 0;
if ((code = qExecTask(pExecutor, &output, &ts)) < 0) {
if (code == TSDB_CODE_QRY_IN_EXEC) {
resetTaskInfo(pExecutor);
}
stError("unexpected stream execution, s-task:%s since %s", pTask->id.idStr, tstrerror(code));
continue;
}
if (output == NULL) {
if (pItem->type == STREAM_INPUT__DATA_RETRIEVE) {
SSDataBlock block = {0};
const SStreamDataBlock* pRetrieveBlock = (const SStreamDataBlock*)pItem;
ASSERT(taosArrayGetSize(pRetrieveBlock->blocks) == 1);
(void)assignOneDataBlock(&block, taosArrayGet(pRetrieveBlock->blocks, 0));
block.info.type = STREAM_PULL_OVER;
block.info.childId = pTask->info.selfChildId;
void* p = taosArrayPush(pRes, &block);
if (p != NULL) {
numOfBlocks += 1;
} else {
stError("s-task:%s failed to add retrieve block", pTask->id.idStr);
}
stDebug("s-task:%s(child %d) retrieve process completed, QID:0x%" PRIx64 " dump results", pTask->id.idStr,
pTask->info.selfChildId, pRetrieveBlock->reqId);
}
break;
}
if (output->info.type == STREAM_RETRIEVE) {
if (streamBroadcastToUpTasks(pTask, output) < 0) {
// TODO
}
continue;
} else if (output->info.type == STREAM_CHECKPOINT) {
continue; // checkpoint block not dispatch to downstream tasks
}
SSDataBlock block = {0};
code = assignOneDataBlock(&block, output);
if (code) {
stError("s-task:%s failed to build result block due to out of memory", pTask->id.idStr);
continue;
}
block.info.childId = pTask->info.selfChildId;
size += blockDataGetSize(output) + sizeof(SSDataBlock) + sizeof(SColumnInfoData) * blockDataGetNumOfCols(&block);
numOfBlocks += 1;
void* p = taosArrayPush(pRes, &block);
if (p == NULL) {
stError("s-task:%s failed to add computing results, the final res may be incorrect", pTask->id.idStr);
}
stDebug("s-task:%s (child %d) executed and get %d result blocks, size:%.2fMiB", pTask->id.idStr,
pTask->info.selfChildId, numOfBlocks, SIZE_IN_MiB(size));
// current output should be dispatched to down stream nodes
if (numOfBlocks >= STREAM_RESULT_DUMP_THRESHOLD || size >= STREAM_RESULT_DUMP_SIZE_THRESHOLD) {
ASSERT(numOfBlocks == taosArrayGetSize(pRes));
code = doDumpResult(pTask, pItem, pRes, size, totalSize, totalBlocks);
// todo: here we need continue retry to put it into output buffer
if (code != TSDB_CODE_SUCCESS) {
return;
}
pRes = NULL;
size = 0;
numOfBlocks = 0;
}
}
if (numOfBlocks > 0) {
ASSERT(numOfBlocks == taosArrayGetSize(pRes));
code = doDumpResult(pTask, pItem, pRes, size, totalSize, totalBlocks);
} else {
taosArrayDestroyEx(pRes, (FDelete)blockDataFreeRes);
}
}
// todo contiuous try to create result blocks
static int32_t handleScanhistoryResultBlocks(SStreamTask* pTask, SArray* pRes, int32_t size) {
int32_t code = TSDB_CODE_SUCCESS;
if (taosArrayGetSize(pRes) > 0) {
SStreamDataBlock* pStreamBlocks = NULL;
code = createStreamBlockFromResults(NULL, pTask, size, pRes, &pStreamBlocks);
if (code) {
stError("s-task:%s failed to build history result blocks", pTask->id.idStr);
return code;
}
code = doOutputResultBlockImpl(pTask, pStreamBlocks);
if (code != TSDB_CODE_SUCCESS) { // should not have error code
stError("s-task:%s dump fill-history results failed, code:%s", pTask->id.idStr, tstrerror(code));
}
} else {
taosArrayDestroyEx(pRes, (FDelete)blockDataFreeRes);
}
return code;
}
static void streamScanHistoryDataImpl(SStreamTask* pTask, SArray* pRes, int32_t* pSize, bool* pFinish) {
int32_t code = TSDB_CODE_SUCCESS;
void* exec = pTask->exec.pExecutor;
int32_t numOfBlocks = 0;
while (1) {
if (streamTaskShouldStop(pTask)) {
break;
}
if (pTask->inputq.status == TASK_INPUT_STATUS__BLOCKED) {
stDebug("s-task:%s level:%d inputQ is blocked, retry in 5s", pTask->id.idStr, pTask->info.taskLevel);
break;
}
SSDataBlock* output = NULL;
uint64_t ts = 0;
code = qExecTask(exec, &output, &ts);
if (code != TSDB_CODE_TSC_QUERY_KILLED && code != TSDB_CODE_SUCCESS) {
stError("s-task:%s scan-history data error occurred code:%s, continue scan-history", pTask->id.idStr,
tstrerror(code));
continue;
}
// the generated results before fill-history task been paused, should be dispatched to sink node
if (output == NULL) {
(*pFinish) = qStreamScanhistoryFinished(exec);
break;
}
SSDataBlock block = {0};
(void)assignOneDataBlock(&block, output);
block.info.childId = pTask->info.selfChildId;
void* p = taosArrayPush(pRes, &block);
if (p == NULL) {
stError("s-task:%s failed to add computing results, the final res may be incorrect", pTask->id.idStr);
}
(*pSize) +=
blockDataGetSize(output) + sizeof(SSDataBlock) + sizeof(SColumnInfoData) * blockDataGetNumOfCols(&block);
numOfBlocks += 1;
if (numOfBlocks >= STREAM_RESULT_DUMP_THRESHOLD || (*pSize) >= STREAM_RESULT_DUMP_SIZE_THRESHOLD) {
stDebug("s-task:%s scan exec numOfBlocks:%d, size:%.2fKiB output num-limit:%d, size-limit:%.2fKiB reached",
pTask->id.idStr, numOfBlocks, SIZE_IN_KiB(*pSize), STREAM_RESULT_DUMP_THRESHOLD,
SIZE_IN_KiB(STREAM_RESULT_DUMP_SIZE_THRESHOLD));
break;
}
}
}
static SScanhistoryDataInfo buildScanhistoryExecRet(EScanHistoryCode code, int32_t idleTime) {
return (SScanhistoryDataInfo){code, idleTime};
}
SScanhistoryDataInfo streamScanHistoryData(SStreamTask* pTask, int64_t st) {
ASSERT(pTask->info.taskLevel == TASK_LEVEL__SOURCE);
void* exec = pTask->exec.pExecutor;
bool finished = false;
const char* id = pTask->id.idStr;
if (!pTask->hTaskInfo.operatorOpen) {
(void)qSetStreamOpOpen(exec);
pTask->hTaskInfo.operatorOpen = true;
}
while (1) {
if (streamTaskShouldPause(pTask)) {
stDebug("s-task:%s paused from the scan-history task", id);
// quit from step1, not continue to handle the step2
return buildScanhistoryExecRet(TASK_SCANHISTORY_QUIT, 0);
}
// output queue is full, idle for 5 sec.
if (streamQueueIsFull(pTask->outputq.queue)) {
stWarn("s-task:%s outputQ is full, idle for 1sec and retry", id);
return buildScanhistoryExecRet(TASK_SCANHISTORY_REXEC, STREAM_SCAN_HISTORY_TIMESLICE);
}
if (pTask->inputq.status == TASK_INPUT_STATUS__BLOCKED) {
stWarn("s-task:%s downstream task inputQ blocked, idle for 5sec and retry", id);
return buildScanhistoryExecRet(TASK_SCANHISTORY_REXEC, FILL_HISTORY_TASK_EXEC_INTERVAL);
}
SArray* pRes = taosArrayInit(0, sizeof(SSDataBlock));
if (pRes == NULL) {
terrno = TSDB_CODE_OUT_OF_MEMORY;
stError("s-task:%s scan-history prepare result block failed, code:%s, retry later", id, tstrerror(terrno));
continue;
}
int32_t size = 0;
streamScanHistoryDataImpl(pTask, pRes, &size, &finished);
if (streamTaskShouldStop(pTask)) {
taosArrayDestroyEx(pRes, (FDelete)blockDataFreeRes);
return buildScanhistoryExecRet(TASK_SCANHISTORY_QUIT, 0);
}
// dispatch the generated results, todo fix error
(void)handleScanhistoryResultBlocks(pTask, pRes, size);
if (finished) {
return buildScanhistoryExecRet(TASK_SCANHISTORY_CONT, 0);
}
int64_t el = taosGetTimestampMs() - st;
if (el >= STREAM_SCAN_HISTORY_TIMESLICE && (pTask->info.fillHistory == 1)) {
stDebug("s-task:%s fill-history:%d time slice exhausted, elapsed time:%.2fs, retry in 100ms", id,
pTask->info.fillHistory, el / 1000.0);
return buildScanhistoryExecRet(TASK_SCANHISTORY_REXEC, 100);
}
}
}
int32_t streamTransferStateDoPrepare(SStreamTask* pTask) {
SStreamMeta* pMeta = pTask->pMeta;
const char* id = pTask->id.idStr;
SStreamTask* pStreamTask = NULL;
int32_t code = streamMetaAcquireTask(pMeta, pTask->streamTaskId.streamId, pTask->streamTaskId.taskId, &pStreamTask);
if (pStreamTask == NULL || code != TSDB_CODE_SUCCESS) {
stError(
"s-task:%s failed to find related stream task:0x%x, it may have been destroyed or closed, destroy the related "
"fill-history task",
id, (int32_t)pTask->streamTaskId.taskId);
// 1. free it and remove fill-history task from disk meta-store
// todo: this function should never be failed.
(void)streamBuildAndSendDropTaskMsg(pTask->pMsgCb, pMeta->vgId, &pTask->id, 0);
// 2. save to disk
streamMetaWLock(pMeta);
if (streamMetaCommit(pMeta) < 0) {
// persist to disk
}
streamMetaWUnLock(pMeta);
return TSDB_CODE_STREAM_TASK_NOT_EXIST;
} else {
double el = (taosGetTimestampMs() - pTask->execInfo.step2Start) / 1000.;
stDebug(
"s-task:%s fill-history task end, status:%s, scan wal elapsed time:%.2fSec, update related stream task:%s "
"info, prepare transfer exec state",
id, streamTaskGetStatus(pTask).name, el, pStreamTask->id.idStr);
}
ETaskStatus status = streamTaskGetStatus(pStreamTask).state;
STimeWindow* pTimeWindow = &pStreamTask->dataRange.window;
// It must be halted for a source stream task, since when the related scan-history-data task start scan the history
// for the step 2.
if (pStreamTask->info.taskLevel == TASK_LEVEL__SOURCE) {
ASSERT(status == TASK_STATUS__HALT || status == TASK_STATUS__DROPPING || status == TASK_STATUS__STOP);
} else {
ASSERT(status == TASK_STATUS__READY || status == TASK_STATUS__PAUSE || status == TASK_STATUS__DROPPING ||
status == TASK_STATUS__STOP);
code = streamTaskHandleEvent(pStreamTask->status.pSM, TASK_EVENT_HALT);
if (code != TSDB_CODE_SUCCESS) {
stError("s-task:%s halt stream task:%s failed, code:%s not transfer state to stream task", id,
pStreamTask->id.idStr, tstrerror(code));
streamMetaReleaseTask(pMeta, pStreamTask);
return code;
} else {
stDebug("s-task:%s halt by related fill-history task:%s", pStreamTask->id.idStr, id);
}
}
// In case of sink tasks, no need to halt them.
// In case of source tasks and agg tasks, we should HALT them, and wait for them to be idle. And then, it's safe to
// start the task state transfer procedure.
SStreamTaskState pState = streamTaskGetStatus(pStreamTask);
status = pState.state;
char* p = pState.name;
if (status == TASK_STATUS__STOP || status == TASK_STATUS__DROPPING) {
stError("s-task:%s failed to transfer state from fill-history task:%s, status:%s", id, pStreamTask->id.idStr, p);
streamMetaReleaseTask(pMeta, pStreamTask);
return TSDB_CODE_STREAM_TASK_IVLD_STATUS;
}
// 1. expand the query time window for stream task of WAL scanner
if (pStreamTask->info.taskLevel == TASK_LEVEL__SOURCE) {
// update the scan data range for source task.
stDebug("s-task:%s level:%d stream task window %" PRId64 " - %" PRId64 " update to %" PRId64 " - %" PRId64
", status:%s, sched-status:%d",
pStreamTask->id.idStr, TASK_LEVEL__SOURCE, pTimeWindow->skey, pTimeWindow->ekey, INT64_MIN,
pTimeWindow->ekey, p, pStreamTask->status.schedStatus);
(void)streamTaskResetTimewindowFilter(pStreamTask);
} else {
stDebug("s-task:%s no need to update/reset filter time window for non-source tasks", pStreamTask->id.idStr);
}
// NOTE: transfer the ownership of executor state before handle the checkpoint block during stream exec
// 2. send msg to mnode to launch a checkpoint to keep the state for current stream
(void)streamTaskSendCheckpointReq(pStreamTask);
// 3. assign the status to the value that will be kept in disk
pStreamTask->status.taskStatus = streamTaskGetStatus(pStreamTask).state;
// 4. open the inputQ for all upstream tasks
streamTaskOpenAllUpstreamInput(pStreamTask);
streamMetaReleaseTask(pMeta, pStreamTask);
return TSDB_CODE_SUCCESS;
}
static int32_t haltCallback(SStreamTask* pTask, void* param) {
streamTaskOpenAllUpstreamInput(pTask);
(void)streamTaskSendCheckpointReq(pTask);
return TSDB_CODE_SUCCESS;
}
int32_t streamTransferStatePrepare(SStreamTask* pTask) {
int32_t code = TSDB_CODE_SUCCESS;
SStreamMeta* pMeta = pTask->pMeta;
ASSERT(pTask->status.appendTranstateBlock == 1);
int32_t level = pTask->info.taskLevel;
if (level == TASK_LEVEL__AGG || level == TASK_LEVEL__SOURCE) { // do transfer task operator states.
code = streamTransferStateDoPrepare(pTask);
} else {
// no state transfer for sink tasks, and drop fill-history task, followed by opening inputQ of sink task.
SStreamTask* pStreamTask = NULL;
code = streamMetaAcquireTask(pMeta, pTask->streamTaskId.streamId, pTask->streamTaskId.taskId, &pStreamTask);
if (pStreamTask != NULL) {
// halt the related stream sink task
code = streamTaskHandleEventAsync(pStreamTask->status.pSM, TASK_EVENT_HALT, haltCallback, NULL);
if (code != TSDB_CODE_SUCCESS) {
stError("s-task:%s halt stream task:%s failed, code:%s not transfer state to stream task", pTask->id.idStr,
pStreamTask->id.idStr, tstrerror(code));
streamMetaReleaseTask(pMeta, pStreamTask);
return code;
} else {
stDebug("s-task:%s sink task halt by related fill-history task:%s", pStreamTask->id.idStr, pTask->id.idStr);
}
streamMetaReleaseTask(pMeta, pStreamTask);
}
}
return code;
}
// set input
static int32_t doSetStreamInputBlock(SStreamTask* pTask, const void* pInput, int64_t* pVer, const char* id) {
void* pExecutor = pTask->exec.pExecutor;
int32_t code = 0;
const SStreamQueueItem* pItem = pInput;
if (pItem->type == STREAM_INPUT__GET_RES) {
const SStreamTrigger* pTrigger = (const SStreamTrigger*)pInput;
code = qSetMultiStreamInput(pExecutor, pTrigger->pBlock, 1, STREAM_INPUT__DATA_BLOCK);
} else if (pItem->type == STREAM_INPUT__DATA_SUBMIT) {
ASSERT(pTask->info.taskLevel == TASK_LEVEL__SOURCE);
const SStreamDataSubmit* pSubmit = (const SStreamDataSubmit*)pInput;
code = qSetMultiStreamInput(pExecutor, &pSubmit->submit, 1, STREAM_INPUT__DATA_SUBMIT);
stDebug("s-task:%s set submit blocks as source block completed, %p %p len:%d ver:%" PRId64, id, pSubmit,
pSubmit->submit.msgStr, pSubmit->submit.msgLen, pSubmit->submit.ver);
ASSERT((*pVer) <= pSubmit->submit.ver);
(*pVer) = pSubmit->submit.ver;
} else if (pItem->type == STREAM_INPUT__DATA_BLOCK || pItem->type == STREAM_INPUT__DATA_RETRIEVE) {
const SStreamDataBlock* pBlock = (const SStreamDataBlock*)pInput;
SArray* pBlockList = pBlock->blocks;
int32_t numOfBlocks = taosArrayGetSize(pBlockList);
stDebug("s-task:%s set sdata blocks as input num:%d, ver:%" PRId64, id, numOfBlocks, pBlock->sourceVer);
code = qSetMultiStreamInput(pExecutor, pBlockList->pData, numOfBlocks, STREAM_INPUT__DATA_BLOCK);
} else if (pItem->type == STREAM_INPUT__MERGED_SUBMIT) {
const SStreamMergedSubmit* pMerged = (const SStreamMergedSubmit*)pInput;
SArray* pBlockList = pMerged->submits;
int32_t numOfBlocks = taosArrayGetSize(pBlockList);
stDebug("s-task:%s %p set (merged) submit blocks as a batch, numOfBlocks:%d, ver:%" PRId64, id, pTask, numOfBlocks,
pMerged->ver);
code = qSetMultiStreamInput(pExecutor, pBlockList->pData, numOfBlocks, STREAM_INPUT__MERGED_SUBMIT);
ASSERT((*pVer) <= pMerged->ver);
(*pVer) = pMerged->ver;
} else if (pItem->type == STREAM_INPUT__REF_DATA_BLOCK) {
const SStreamRefDataBlock* pRefBlock = (const SStreamRefDataBlock*)pInput;
code = qSetMultiStreamInput(pExecutor, pRefBlock->pBlock, 1, STREAM_INPUT__DATA_BLOCK);
} else if (pItem->type == STREAM_INPUT__CHECKPOINT || pItem->type == STREAM_INPUT__CHECKPOINT_TRIGGER) {
const SStreamDataBlock* pCheckpoint = (const SStreamDataBlock*)pInput;
code = qSetMultiStreamInput(pExecutor, pCheckpoint->blocks, 1, pItem->type);
} else {
ASSERT(0);
}
return code;
}
void streamProcessTransstateBlock(SStreamTask* pTask, SStreamDataBlock* pBlock) {
const char* id = pTask->id.idStr;
int32_t code = TSDB_CODE_SUCCESS;
int32_t level = pTask->info.taskLevel;
if (level == TASK_LEVEL__AGG || level == TASK_LEVEL__SINK) {
int32_t remain = streamAlignTransferState(pTask);
if (remain > 0) {
streamFreeQitem((SStreamQueueItem*)pBlock);
stDebug("s-task:%s receive upstream trans-state msg, not sent remain:%d", id, remain);
return;
}
}
// dispatch the tran-state block to downstream task immediately
int32_t type = pTask->outputInfo.type;
// transfer the ownership of executor state
if (type == TASK_OUTPUT__FIXED_DISPATCH || type == TASK_OUTPUT__SHUFFLE_DISPATCH) {
if (level == TASK_LEVEL__SOURCE) {
stDebug("s-task:%s add transfer-state block into outputQ", id);
} else {
stDebug("s-task:%s all upstream tasks send transfer-state block, add transfer-state block into outputQ", id);
ASSERT(pTask->streamTaskId.taskId != 0 && pTask->info.fillHistory == 1);
}
// agg task should dispatch trans-state msg to sink task, to flush all data to sink task.
if (level == TASK_LEVEL__AGG || level == TASK_LEVEL__SOURCE) {
pBlock->srcVgId = pTask->pMeta->vgId;
code = taosWriteQitem(pTask->outputq.queue->pQueue, pBlock);
if (code == 0) {
(void)streamDispatchStreamBlock(pTask);
} else { // todo put into queue failed, retry
streamFreeQitem((SStreamQueueItem*)pBlock);
}
} else { // level == TASK_LEVEL__SINK
streamFreeQitem((SStreamQueueItem*)pBlock);
}
} else { // non-dispatch task, do task state transfer directly
streamFreeQitem((SStreamQueueItem*)pBlock);
stDebug("s-task:%s non-dispatch task, level:%d start to transfer state directly", id, level);
ASSERT(pTask->info.fillHistory == 1);
code = streamTransferStatePrepare(pTask);
if (code != TSDB_CODE_SUCCESS) {
(void)streamTaskSetSchedStatusInactive(pTask);
}
}
}
// static void streamTaskSetIdleInfo(SStreamTask* pTask, int32_t idleTime) { pTask->status.schedIdleTime = idleTime; }
static void setLastExecTs(SStreamTask* pTask, int64_t ts) { pTask->status.lastExecTs = ts; }
static void doStreamTaskExecImpl(SStreamTask* pTask, SStreamQueueItem* pBlock, int32_t num) {
const char* id = pTask->id.idStr;
int32_t blockSize = 0;
int64_t st = taosGetTimestampMs();
SCheckpointInfo* pInfo = &pTask->chkInfo;
int64_t ver = pInfo->processedVer;
stDebug("s-task:%s start to process batch blocks, num:%d, type:%s", id, num, streamQueueItemGetTypeStr(pBlock->type));
int32_t code = doSetStreamInputBlock(pTask, pBlock, &ver, id);
if (code) {
stError("s-task:%s failed to set input block, not exec for these blocks", id);
return;
}
int64_t totalSize = 0;
int32_t totalBlocks = 0;
streamTaskExecImpl(pTask, pBlock, &totalSize, &totalBlocks);
double el = (taosGetTimestampMs() - st) / 1000.0;
stDebug("s-task:%s batch of input blocks exec end, elapsed time:%.2fs, result size:%.2fMiB, numOfBlocks:%d", id, el,
SIZE_IN_MiB(totalSize), totalBlocks);
pTask->execInfo.outputDataBlocks += totalBlocks;
pTask->execInfo.outputDataSize += totalSize;
if (fabs(el - 0.0) <= DBL_EPSILON) {
pTask->execInfo.procsThroughput = 0;
pTask->execInfo.outputThroughput = 0;
} else {
pTask->execInfo.outputThroughput = (totalSize / el);
pTask->execInfo.procsThroughput = (blockSize / el);
}
// update the currentVer if processing the submit blocks.
ASSERT(pInfo->checkpointVer <= pInfo->nextProcessVer && ver >= pInfo->checkpointVer);
if (ver != pInfo->processedVer) {
stDebug("s-task:%s update processedVer(unsaved) from %" PRId64 " to %" PRId64 " nextProcessVer:%" PRId64
" ckpt:%" PRId64,
id, pInfo->processedVer, ver, pInfo->nextProcessVer, pInfo->checkpointVer);
pInfo->processedVer = ver;
}
}
void flushStateDataInExecutor(SStreamTask* pTask, SStreamQueueItem* pCheckpointBlock) {
const char* id = pTask->id.idStr;
// 1. transfer the ownership of executor state
bool dropRelHTask = (streamTaskGetPrevStatus(pTask) == TASK_STATUS__HALT);
if (dropRelHTask) {
ASSERT(HAS_RELATED_FILLHISTORY_TASK(pTask));
STaskId* pHTaskId = &pTask->hTaskInfo.id;
SStreamTask* pHTask = NULL;
int32_t code = streamMetaAcquireTask(pTask->pMeta, pHTaskId->streamId, pHTaskId->taskId, &pHTask);
if (code == TSDB_CODE_SUCCESS) { // ignore the error code.
(void)streamTaskReleaseState(pHTask);
(void)streamTaskReloadState(pTask);
stDebug("s-task:%s transfer state from fill-history task:%s, status:%s completed", id, pHTask->id.idStr,
streamTaskGetStatus(pHTask).name);
// todo execute qExecTask to fetch the reload-generated result, if this is stream is for session window query.
/*
* while(1) {
* qExecTask()
* }
* // put into the output queue.
*/
streamMetaReleaseTask(pTask->pMeta, pHTask);
} else {
stError("s-task:%s related fill-history task:0x%x failed to acquire, transfer state failed", id,
(int32_t)pHTaskId->taskId);
}
} else {
stDebug("s-task:%s no transfer-state needed", id);
}
// 2. flush data in executor to K/V store, which should be completed before do checkpoint in the K/V.
doStreamTaskExecImpl(pTask, pCheckpointBlock, 1);
}
/**
* todo: the batch of blocks should be tuned dynamic, according to the total elapsed time of each batch of blocks, the
* appropriate batch of blocks should be handled in 5 to 10 sec.
*/
static int32_t doStreamExecTask(SStreamTask* pTask) {
const char* id = pTask->id.idStr;
// merge multiple input data if possible in the input queue.
stDebug("s-task:%s start to extract data block from inputQ", id);
while (1) {
int32_t blockSize = 0;
int32_t numOfBlocks = 0;
SStreamQueueItem* pInput = NULL;
if (streamTaskShouldStop(pTask) || (streamTaskGetStatus(pTask).state == TASK_STATUS__UNINIT)) {
stDebug("s-task:%s stream task is stopped", id);
return 0;
}
if (streamQueueIsFull(pTask->outputq.queue)) {
stTrace("s-task:%s outputQ is full, idle for 500ms and retry", id);
streamTaskSetIdleInfo(pTask, 1000);
return 0;
}
if (pTask->inputq.status == TASK_INPUT_STATUS__BLOCKED) {
stTrace("s-task:%s downstream task inputQ blocked, idle for 1sec and retry", id);
streamTaskSetIdleInfo(pTask, 1000);
return 0;
}
if (taosGetTimestampMs() - pTask->status.lastExecTs < MIN_INVOKE_INTERVAL) {
stDebug("s-task:%s invoke exec too fast, idle and retry in 50ms", id);
streamTaskSetIdleInfo(pTask, MIN_INVOKE_INTERVAL);
return 0;
}
EExtractDataCode ret = streamTaskGetDataFromInputQ(pTask, &pInput, &numOfBlocks, &blockSize);
if (ret == EXEC_AFTER_IDLE) {
ASSERT(pInput == NULL && numOfBlocks == 0);
streamTaskSetIdleInfo(pTask, MIN_INVOKE_INTERVAL);
return 0;
} else {
if (pInput == NULL) {
ASSERT(numOfBlocks == 0);
return 0;
}
}
pTask->execInfo.inputDataBlocks += numOfBlocks;
pTask->execInfo.inputDataSize += blockSize;
// dispatch checkpoint msg to all downstream tasks
int32_t type = pInput->type;
if (type == STREAM_INPUT__CHECKPOINT_TRIGGER) {
(void)streamProcessCheckpointTriggerBlock(pTask, (SStreamDataBlock*)pInput);
continue;
}
if (type == STREAM_INPUT__TRANS_STATE) {
streamProcessTransstateBlock(pTask, (SStreamDataBlock*)pInput);
continue;
}
if (pTask->info.taskLevel == TASK_LEVEL__SINK) {
ASSERT(type == STREAM_INPUT__DATA_BLOCK || type == STREAM_INPUT__CHECKPOINT);
int64_t st = taosGetTimestampMs();
// here only handle the data block sink operation
if (type == STREAM_INPUT__DATA_BLOCK) {
pTask->execInfo.sink.dataSize += blockSize;
stDebug("s-task:%s sink task start to sink %d blocks, size:%.2fKiB", id, numOfBlocks, SIZE_IN_KiB(blockSize));
int32_t code = doOutputResultBlockImpl(pTask, (SStreamDataBlock*)pInput);
if (code != TSDB_CODE_SUCCESS) {
// todo handle error.
}
double el = (taosGetTimestampMs() - st) / 1000.0;
if (fabs(el - 0.0) <= DBL_EPSILON) {
pTask->execInfo.procsThroughput = 0;
} else {
pTask->execInfo.procsThroughput = (blockSize / el);
}
continue;
}
}
if (type != STREAM_INPUT__CHECKPOINT) {
doStreamTaskExecImpl(pTask, pInput, numOfBlocks);
streamFreeQitem(pInput);
} else { // todo other thread may change the status
// do nothing after sync executor state to storage backend, untill the vnode-level checkpoint is completed.
streamMutexLock(&pTask->lock);
SStreamTaskState pState = streamTaskGetStatus(pTask);
if (pState.state == TASK_STATUS__CK) {
stDebug("s-task:%s checkpoint block received, set status:%s", id, pState.name);
(void)streamTaskBuildCheckpoint(pTask); // ignore this error msg, and continue
} else { // todo refactor
int32_t code = 0;
if (pTask->info.taskLevel == TASK_LEVEL__SOURCE) {
code = streamTaskSendCheckpointSourceRsp(pTask);
} else {
code = streamTaskSendCheckpointReadyMsg(pTask);
}
if (code != TSDB_CODE_SUCCESS) {
// todo: let's retry send rsp to upstream/mnode
stError("s-task:%s failed to send checkpoint rsp to upstream, checkpointId:%d, code:%s", id, 0,
tstrerror(code));
}
}
streamMutexUnlock(&pTask->lock);
streamFreeQitem(pInput);
return 0;
}
}
}
// the task may be set dropping/stopping, while it is still in the task queue, therefore, the sched-status can not
// be updated by tryExec function, therefore, the schedStatus will always be the TASK_SCHED_STATUS__WAITING.
bool streamTaskIsIdle(const SStreamTask* pTask) {
ETaskStatus status = streamTaskGetStatus(pTask).state;
return (pTask->status.schedStatus == TASK_SCHED_STATUS__INACTIVE || status == TASK_STATUS__STOP ||
status == TASK_STATUS__DROPPING);
}
bool streamTaskReadyToRun(const SStreamTask* pTask, char** pStatus) {
SStreamTaskState pState = streamTaskGetStatus(pTask);
ETaskStatus st = pState.state;
if (pStatus != NULL) {
*pStatus = pState.name;
}
// pause & halt will still run for sink tasks.
if (streamTaskIsSinkTask(pTask)) {
return (st == TASK_STATUS__READY || st == TASK_STATUS__SCAN_HISTORY || st == TASK_STATUS__CK ||
st == TASK_STATUS__PAUSE || st == TASK_STATUS__HALT);
} else {
return (st == TASK_STATUS__READY || st == TASK_STATUS__SCAN_HISTORY || st == TASK_STATUS__CK ||
st == TASK_STATUS__HALT);
}
}
void streamResumeTask(SStreamTask* pTask) {
ASSERT(pTask->status.schedStatus == TASK_SCHED_STATUS__ACTIVE);
const char* id = pTask->id.idStr;
while (1) {
(void)doStreamExecTask(pTask);
// check if continue
streamMutexLock(&pTask->lock);
int32_t numOfItems = streamQueueGetNumOfItems(pTask->inputq.queue);
if ((numOfItems == 0) || streamTaskShouldStop(pTask) || streamTaskShouldPause(pTask)) {
atomic_store_8(&pTask->status.schedStatus, TASK_SCHED_STATUS__INACTIVE);
streamTaskClearSchedIdleInfo(pTask);
streamMutexUnlock(&pTask->lock);
setLastExecTs(pTask, taosGetTimestampMs());
char* p = streamTaskGetStatus(pTask).name;
stDebug("s-task:%s exec completed, status:%s, sched-status:%d, lastExecTs:%" PRId64, id, p,
pTask->status.schedStatus, pTask->status.lastExecTs);
return;
} else {
// check if this task needs to be idle for a while
if (pTask->status.schedIdleTime > 0) {
streamTaskResumeInFuture(pTask);
streamMutexUnlock(&pTask->lock);
setLastExecTs(pTask, taosGetTimestampMs());
return;
}
}
streamMutexUnlock(&pTask->lock);
}
}
int32_t streamExecTask(SStreamTask* pTask) {
// this function may be executed by multi-threads, so status check is required.
const char* id = pTask->id.idStr;
int8_t schedStatus = streamTaskSetSchedStatusActive(pTask);
if (schedStatus == TASK_SCHED_STATUS__WAITING) {
streamResumeTask(pTask);
} else {
char* p = streamTaskGetStatus(pTask).name;
stDebug("s-task:%s already started to exec by other thread, status:%s, sched-status:%d", id, p,
pTask->status.schedStatus);
}
return 0;
}
int32_t streamTaskReleaseState(SStreamTask* pTask) {
stDebug("s-task:%s release exec state", pTask->id.idStr);
void* pExecutor = pTask->exec.pExecutor;
int32_t code = TSDB_CODE_SUCCESS;
if (pExecutor != NULL) {
code = qStreamOperatorReleaseState(pExecutor);
}
return code;
}
int32_t streamTaskReloadState(SStreamTask* pTask) {
stDebug("s-task:%s reload exec state", pTask->id.idStr);
void* pExecutor = pTask->exec.pExecutor;
int32_t code = TSDB_CODE_SUCCESS;
if (pExecutor != NULL) {
code = qStreamOperatorReloadState(pExecutor);
}
return code;
}
int32_t streamAlignTransferState(SStreamTask* pTask) {
int32_t numOfUpstream = taosArrayGetSize(pTask->upstreamInfo.pList);
int32_t old = atomic_val_compare_exchange_32(&pTask->transferStateAlignCnt, 0, numOfUpstream);
if (old == 0) {
stDebug("s-task:%s set the transfer state aligncnt %d", pTask->id.idStr, numOfUpstream);
}
return atomic_sub_fetch_32(&pTask->transferStateAlignCnt, 1);
}