aggregateoperator.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 "filter.h"
#include "function.h"
#include "functionMgt.h"
#include "os.h"
#include "querynodes.h"
#include "tfill.h"
#include "tname.h"

#include "tdatablock.h"
#include "tglobal.h"
#include "tmsg.h"
#include "ttime.h"

#include "executorimpl.h"
#include "index.h"
#include "query.h"
#include "tcompare.h"
#include "thash.h"
#include "ttypes.h"
#include "vnode.h"

typedef struct {
  bool    hasAgg;
  int32_t numOfRows;
  int32_t startOffset;
} SFunctionCtxStatus;

typedef struct SAggOperatorInfo {
  SOptrBasicInfo   binfo;
  SAggSupporter    aggSup;
  STableQueryInfo* current;
  uint64_t         groupId;
  SGroupResInfo    groupResInfo;
  SExprSupp        scalarExprSup;
  bool             groupKeyOptimized;
} SAggOperatorInfo;

static void destroyAggOperatorInfo(void* param);
static void setExecutionContext(SOperatorInfo* pOperator, int32_t numOfOutput, uint64_t groupId);

static int32_t createDataBlockForEmptyInput(SOperatorInfo* pOperator, SSDataBlock** ppBlock);
static void destroyDataBlockForEmptyInput(bool blockAllocated, SSDataBlock** ppBlock);

static int32_t doOpenAggregateOptr(SOperatorInfo* pOperator);
static int32_t doAggregateImpl(SOperatorInfo* pOperator, SqlFunctionCtx* pCtx);
static SSDataBlock* getAggregateResult(SOperatorInfo* pOperator);

static int32_t doInitAggInfoSup(SAggSupporter* pAggSup, SqlFunctionCtx* pCtx, int32_t numOfOutput, size_t keyBufSize,
                         const char* pKey);

static int32_t addNewResultRowBuf(SResultRow* pWindowRes, SDiskbasedBuf* pResultBuf, uint32_t size);

static void doSetTableGroupOutputBuf(SOperatorInfo* pOperator, int32_t numOfOutput, uint64_t groupId);

static void functionCtxSave(SqlFunctionCtx* pCtx, SFunctionCtxStatus* pStatus);
static void functionCtxRestore(SqlFunctionCtx* pCtx, SFunctionCtxStatus* pStatus);

SOperatorInfo* createAggregateOperatorInfo(SOperatorInfo* downstream, SAggPhysiNode* pAggNode,
                                           SExecTaskInfo* pTaskInfo) {
  SAggOperatorInfo* pInfo = taosMemoryCalloc(1, sizeof(SAggOperatorInfo));
  SOperatorInfo*    pOperator = taosMemoryCalloc(1, sizeof(SOperatorInfo));
  if (pInfo == NULL || pOperator == NULL) {
    goto _error;
  }

  SSDataBlock* pResBlock = createDataBlockFromDescNode(pAggNode->node.pOutputDataBlockDesc);
  initBasicInfo(&pInfo->binfo, pResBlock);

  size_t keyBufSize = sizeof(int64_t) + sizeof(int64_t) + POINTER_BYTES;
  initResultSizeInfo(&pOperator->resultInfo, 4096);

  int32_t    num = 0;
  SExprInfo* pExprInfo = createExprInfo(pAggNode->pAggFuncs, pAggNode->pGroupKeys, &num);
  int32_t    code = initAggSup(&pOperator->exprSupp, &pInfo->aggSup, pExprInfo, num, keyBufSize, pTaskInfo->id.str,
                               pTaskInfo->streamInfo.pState);
  if (code != TSDB_CODE_SUCCESS) {
    goto _error;
  }

  int32_t    numOfScalarExpr = 0;
  SExprInfo* pScalarExprInfo = NULL;
  if (pAggNode->pExprs != NULL) {
    pScalarExprInfo = createExprInfo(pAggNode->pExprs, NULL, &numOfScalarExpr);
  }

  code = initExprSupp(&pInfo->scalarExprSup, pScalarExprInfo, numOfScalarExpr);
  if (code != TSDB_CODE_SUCCESS) {
    goto _error;
  }

  code = filterInitFromNode((SNode*)pAggNode->node.pConditions, &pOperator->exprSupp.pFilterInfo, 0);
  if (code != TSDB_CODE_SUCCESS) {
    goto _error;
  }

  pInfo->binfo.mergeResultBlock = pAggNode->mergeDataBlock;
  pInfo->groupKeyOptimized = pAggNode->groupKeyOptimized;
  pInfo->groupId = UINT64_MAX;

  setOperatorInfo(pOperator, "TableAggregate", QUERY_NODE_PHYSICAL_PLAN_HASH_AGG, true, OP_NOT_OPENED, pInfo,
                  pTaskInfo);
  pOperator->fpSet = createOperatorFpSet(doOpenAggregateOptr, getAggregateResult, NULL, destroyAggOperatorInfo,
                                         optrDefaultBufFn, NULL);

  if (downstream->operatorType == QUERY_NODE_PHYSICAL_PLAN_TABLE_SCAN) {
    STableScanInfo* pTableScanInfo = downstream->info;
    pTableScanInfo->base.pdInfo.pExprSup = &pOperator->exprSupp;
    pTableScanInfo->base.pdInfo.pAggSup = &pInfo->aggSup;
  }

  code = appendDownstream(pOperator, &downstream, 1);
  if (code != TSDB_CODE_SUCCESS) {
    goto _error;
  }

  return pOperator;

  _error:
  if (pInfo != NULL) {
    destroyAggOperatorInfo(pInfo);
  }

  if (pOperator != NULL) {
    cleanupExprSupp(&pOperator->exprSupp);
  }

  taosMemoryFreeClear(pOperator);
  pTaskInfo->code = code;
  return NULL;
}

void destroyAggOperatorInfo(void* param) {
  SAggOperatorInfo* pInfo = (SAggOperatorInfo*)param;
  cleanupBasicInfo(&pInfo->binfo);

  cleanupAggSup(&pInfo->aggSup);
  cleanupExprSupp(&pInfo->scalarExprSup);
  cleanupGroupResInfo(&pInfo->groupResInfo);
  taosMemoryFreeClear(param);
}

// this is a blocking operator
int32_t doOpenAggregateOptr(SOperatorInfo* pOperator) {
  if (OPTR_IS_OPENED(pOperator)) {
    return TSDB_CODE_SUCCESS;
  }

  SExecTaskInfo*    pTaskInfo = pOperator->pTaskInfo;
  SAggOperatorInfo* pAggInfo = pOperator->info;

  SExprSupp*     pSup = &pOperator->exprSupp;
  SOperatorInfo* downstream = pOperator->pDownstream[0];

  int64_t st = taosGetTimestampUs();

  int32_t order = TSDB_ORDER_ASC;
  int32_t scanFlag = MAIN_SCAN;

  bool hasValidBlock = false;

  while (1) {
    bool blockAllocated = false;
    SSDataBlock* pBlock = downstream->fpSet.getNextFn(downstream);
    if (pBlock == NULL) {
      if (!hasValidBlock) {
        createDataBlockForEmptyInput(pOperator, &pBlock);
        if (pBlock == NULL) {
          break;
        }
        blockAllocated = true;
      } else {
        break;
      }
    }
    hasValidBlock = true;

    int32_t code = getTableScanInfo(pOperator, &order, &scanFlag, false);
    if (code != TSDB_CODE_SUCCESS) {
      destroyDataBlockForEmptyInput(blockAllocated, &pBlock);
      T_LONG_JMP(pTaskInfo->env, code);
    }

    // there is an scalar expression that needs to be calculated before apply the group aggregation.
    if (pAggInfo->scalarExprSup.pExprInfo != NULL && !blockAllocated) {
      SExprSupp* pSup1 = &pAggInfo->scalarExprSup;
      code = projectApplyFunctions(pSup1->pExprInfo, pBlock, pBlock, pSup1->pCtx, pSup1->numOfExprs, NULL);
      if (code != TSDB_CODE_SUCCESS) {
        destroyDataBlockForEmptyInput(blockAllocated, &pBlock);
        T_LONG_JMP(pTaskInfo->env, code);
      }
    }

    // the pDataBlock are always the same one, no need to call this again
    setExecutionContext(pOperator, pOperator->exprSupp.numOfExprs, pBlock->info.id.groupId);
    setInputDataBlock(pSup, pBlock, order, scanFlag, true);
    code = doAggregateImpl(pOperator, pSup->pCtx);
    if (code != 0) {
      destroyDataBlockForEmptyInput(blockAllocated, &pBlock);
      T_LONG_JMP(pTaskInfo->env, code);
    }

    destroyDataBlockForEmptyInput(blockAllocated, &pBlock);
  }

  // the downstream operator may return with error code, so let's check the code before generating results.
  if (pTaskInfo->code != TSDB_CODE_SUCCESS) {
    T_LONG_JMP(pTaskInfo->env, pTaskInfo->code);
  }

  initGroupedResultInfo(&pAggInfo->groupResInfo, pAggInfo->aggSup.pResultRowHashTable, 0);
  OPTR_SET_OPENED(pOperator);

  pOperator->cost.openCost = (taosGetTimestampUs() - st) / 1000.0;
  return pTaskInfo->code;
}

SSDataBlock* getAggregateResult(SOperatorInfo* pOperator) {
  SAggOperatorInfo* pAggInfo = pOperator->info;
  SOptrBasicInfo*   pInfo = &pAggInfo->binfo;

  if (pOperator->status == OP_EXEC_DONE) {
    return NULL;
  }

  SExecTaskInfo* pTaskInfo = pOperator->pTaskInfo;
  pTaskInfo->code = pOperator->fpSet._openFn(pOperator);
  if (pTaskInfo->code != TSDB_CODE_SUCCESS) {
    setOperatorCompleted(pOperator);
    return NULL;
  }

  blockDataEnsureCapacity(pInfo->pRes, pOperator->resultInfo.capacity);
  while (1) {
    doBuildResultDatablock(pOperator, pInfo, &pAggInfo->groupResInfo, pAggInfo->aggSup.pResultBuf);
    doFilter(pInfo->pRes, pOperator->exprSupp.pFilterInfo, NULL);

    if (!hasRemainResults(&pAggInfo->groupResInfo)) {
      setOperatorCompleted(pOperator);
      break;
    }

    if (pInfo->pRes->info.rows > 0) {
      break;
    }
  }

  size_t rows = blockDataGetNumOfRows(pInfo->pRes);
  pOperator->resultInfo.totalRows += rows;

  return (rows == 0) ? NULL : pInfo->pRes;
}

int32_t doAggregateImpl(SOperatorInfo* pOperator, SqlFunctionCtx* pCtx) {
  for (int32_t k = 0; k < pOperator->exprSupp.numOfExprs; ++k) {
    if (functionNeedToExecute(&pCtx[k])) {
      // todo add a dummy funtion to avoid process check
      if (pCtx[k].fpSet.process == NULL) {
        continue;
      }

      int32_t code = pCtx[k].fpSet.process(&pCtx[k]);
      if (code != TSDB_CODE_SUCCESS) {
        qError("%s aggregate function error happens, code: %s", GET_TASKID(pOperator->pTaskInfo), tstrerror(code));
        return code;
      }
    }
  }

  return TSDB_CODE_SUCCESS;
}

static int32_t createDataBlockForEmptyInput(SOperatorInfo* pOperator, SSDataBlock** ppBlock) {
  if (!tsCountAlwaysReturnValue) {
    return TSDB_CODE_SUCCESS;
  }

  SAggOperatorInfo* pAggInfo = pOperator->info;
  if (pAggInfo->groupKeyOptimized) {
    return TSDB_CODE_SUCCESS;
  }

  SOperatorInfo* downstream = pOperator->pDownstream[0];
  if (downstream->operatorType == QUERY_NODE_PHYSICAL_PLAN_PARTITION ||
      (downstream->operatorType == QUERY_NODE_PHYSICAL_PLAN_TABLE_SCAN &&
       ((STableScanInfo*)downstream->info)->hasGroupByTag == true)) {
    return TSDB_CODE_SUCCESS;
  }

  SqlFunctionCtx* pCtx = pOperator->exprSupp.pCtx;
  bool            hasCountFunc = false;

  for (int32_t i = 0; i < pOperator->exprSupp.numOfExprs; ++i) {
    const char* pName = pCtx[i].pExpr->pExpr->_function.functionName;
    if ((strcmp(pName, "count") == 0) || (strcmp(pName, "hyperloglog") == 0) ||
        (strcmp(pName, "_hyperloglog_partial") == 0) || (strcmp(pName, "_hyperloglog_merge") == 0)) {
      hasCountFunc = true;
      break;
    }
  }

  if (!hasCountFunc) {
    return TSDB_CODE_SUCCESS;
  }

  SSDataBlock* pBlock = createDataBlock();
  pBlock->info.rows = 1;
  pBlock->info.capacity = 0;

  for (int32_t i = 0; i < pOperator->exprSupp.numOfExprs; ++i) {
    SColumnInfoData colInfo = {0};
    colInfo.hasNull = true;
    colInfo.info.type = TSDB_DATA_TYPE_NULL;
    colInfo.info.bytes = 1;

    SExprInfo* pOneExpr = &pOperator->exprSupp.pExprInfo[i];
    for (int32_t j = 0; j < pOneExpr->base.numOfParams; ++j) {
      SFunctParam* pFuncParam = &pOneExpr->base.pParam[j];
      if (pFuncParam->type == FUNC_PARAM_TYPE_COLUMN) {
        int32_t slotId = pFuncParam->pCol->slotId;
        int32_t numOfCols = taosArrayGetSize(pBlock->pDataBlock);
        if (slotId >= numOfCols) {
          taosArrayEnsureCap(pBlock->pDataBlock, slotId + 1);
          for (int32_t k = numOfCols; k < slotId + 1; ++k) {
            taosArrayPush(pBlock->pDataBlock, &colInfo);
          }
        }
      } else if (pFuncParam->type == FUNC_PARAM_TYPE_VALUE) {
        // do nothing
      }
    }
  }

  blockDataEnsureCapacity(pBlock, pBlock->info.rows);
  *ppBlock = pBlock;

  return TSDB_CODE_SUCCESS;
}

void destroyDataBlockForEmptyInput(bool blockAllocated, SSDataBlock** ppBlock) {
  if (!blockAllocated) {
    return;
  }

  blockDataDestroy(*ppBlock);
  *ppBlock = NULL;
}

void setExecutionContext(SOperatorInfo* pOperator, int32_t numOfOutput, uint64_t groupId) {
  SAggOperatorInfo* pAggInfo = pOperator->info;
  if (pAggInfo->groupId != UINT64_MAX && pAggInfo->groupId == groupId) {
    return;
  }

  doSetTableGroupOutputBuf(pOperator, numOfOutput, groupId);

  // record the current active group id
  pAggInfo->groupId = groupId;
}

void doSetTableGroupOutputBuf(SOperatorInfo* pOperator, int32_t numOfOutput, uint64_t groupId) {
  // for simple group by query without interval, all the tables belong to one group result.
  SExecTaskInfo*    pTaskInfo = pOperator->pTaskInfo;
  SAggOperatorInfo* pAggInfo = pOperator->info;

  SResultRowInfo* pResultRowInfo = &pAggInfo->binfo.resultRowInfo;
  SqlFunctionCtx* pCtx = pOperator->exprSupp.pCtx;
  int32_t*        rowEntryInfoOffset = pOperator->exprSupp.rowEntryInfoOffset;

  SResultRow* pResultRow = doSetResultOutBufByKey(pAggInfo->aggSup.pResultBuf, pResultRowInfo, (char*)&groupId,
                                                  sizeof(groupId), true, groupId, pTaskInfo, false, &pAggInfo->aggSup, true);
  /*
   * not assign result buffer yet, add new result buffer
   * all group belong to one result set, and each group result has different group id so set the id to be one
   */
  if (pResultRow->pageId == -1) {
    int32_t ret = addNewResultRowBuf(pResultRow, pAggInfo->aggSup.pResultBuf, pAggInfo->binfo.pRes->info.rowSize);
    if (ret != TSDB_CODE_SUCCESS) {
      T_LONG_JMP(pTaskInfo->env, terrno);
    }
  }

  setResultRowInitCtx(pResultRow, pCtx, numOfOutput, rowEntryInfoOffset);
}

// a new buffer page for each table. Needs to opt this design
int32_t addNewResultRowBuf(SResultRow* pWindowRes, SDiskbasedBuf* pResultBuf, uint32_t size) {
  if (pWindowRes->pageId != -1) {
    return 0;
  }

  SFilePage* pData = NULL;

  // in the first scan, new space needed for results
  int32_t pageId = -1;
  SArray* list = getDataBufPagesIdList(pResultBuf);

  if (taosArrayGetSize(list) == 0) {
    pData = getNewBufPage(pResultBuf, &pageId);
    pData->num = sizeof(SFilePage);
  } else {
    SPageInfo* pi = getLastPageInfo(list);
    pData = getBufPage(pResultBuf, getPageId(pi));
    if (pData == NULL) {
      qError("failed to get buffer, code:%s", tstrerror(terrno));
      return terrno;
    }

    pageId = getPageId(pi);

    if (pData->num + size > getBufPageSize(pResultBuf)) {
      // release current page first, and prepare the next one
      releaseBufPageInfo(pResultBuf, pi);

      pData = getNewBufPage(pResultBuf, &pageId);
      if (pData != NULL) {
        pData->num = sizeof(SFilePage);
      }
    }
  }

  if (pData == NULL) {
    return -1;
  }

  // set the number of rows in current disk page
  if (pWindowRes->pageId == -1) {  // not allocated yet, allocate new buffer
    pWindowRes->pageId = pageId;
    pWindowRes->offset = (int32_t)pData->num;

    pData->num += size;
  }

  return 0;
}

int32_t doInitAggInfoSup(SAggSupporter* pAggSup, SqlFunctionCtx* pCtx, int32_t numOfOutput, size_t keyBufSize,
                         const char* pKey) {
  int32_t code = 0;
  //  _hash_fn_t hashFn = taosGetDefaultHashFunction(TSDB_DATA_TYPE_BINARY);

  pAggSup->currentPageId = -1;
  pAggSup->resultRowSize = getResultRowSize(pCtx, numOfOutput);
  pAggSup->keyBuf = taosMemoryCalloc(1, keyBufSize + POINTER_BYTES + sizeof(int64_t));
  pAggSup->pResultRowHashTable = tSimpleHashInit(100, taosFastHash);

  if (pAggSup->keyBuf == NULL || pAggSup->pResultRowHashTable == NULL) {
    return TSDB_CODE_OUT_OF_MEMORY;
  }

  uint32_t defaultPgsz = 0;
  uint32_t defaultBufsz = 0;
  getBufferPgSize(pAggSup->resultRowSize, &defaultPgsz, &defaultBufsz);

  if (!osTempSpaceAvailable()) {
    code = TSDB_CODE_NO_DISKSPACE;
    qError("Init stream agg supporter failed since %s, key:%s, tempDir:%s", terrstr(code), pKey, tsTempDir);
    return code;
  }

  code = createDiskbasedBuf(&pAggSup->pResultBuf, defaultPgsz, defaultBufsz, pKey, tsTempDir);
  if (code != TSDB_CODE_SUCCESS) {
    qError("Create agg result buf failed since %s, %s", tstrerror(code), pKey);
    return code;
  }

  return code;
}

void cleanupAggSup(SAggSupporter* pAggSup) {
  taosMemoryFreeClear(pAggSup->keyBuf);
  tSimpleHashCleanup(pAggSup->pResultRowHashTable);
  destroyDiskbasedBuf(pAggSup->pResultBuf);
}

int32_t initAggSup(SExprSupp* pSup, SAggSupporter* pAggSup, SExprInfo* pExprInfo, int32_t numOfCols, size_t keyBufSize,
                   const char* pkey, void* pState) {
  int32_t code = initExprSupp(pSup, pExprInfo, numOfCols);
  if (code != TSDB_CODE_SUCCESS) {
    return code;
  }

  code = doInitAggInfoSup(pAggSup, pSup->pCtx, numOfCols, keyBufSize, pkey);
  if (code != TSDB_CODE_SUCCESS) {
    return code;
  }

  for (int32_t i = 0; i < numOfCols; ++i) {
    if (pState) {
      pSup->pCtx[i].saveHandle.pBuf = NULL;
      pSup->pCtx[i].saveHandle.pState = pState;
      pSup->pCtx[i].exprIdx = i;
    } else {
      pSup->pCtx[i].saveHandle.pBuf = pAggSup->pResultBuf;
    }
  }

  return TSDB_CODE_SUCCESS;
}

void applyAggFunctionOnPartialTuples(SExecTaskInfo* taskInfo, SqlFunctionCtx* pCtx, SColumnInfoData* pTimeWindowData,
                                     int32_t offset, int32_t forwardStep, int32_t numOfTotal, int32_t numOfOutput) {
  for (int32_t k = 0; k < numOfOutput; ++k) {
    // keep it temporarily
    SFunctionCtxStatus status = {0};
    functionCtxSave(&pCtx[k], &status);

    pCtx[k].input.startRowIndex = offset;
    pCtx[k].input.numOfRows = forwardStep;

    // not a whole block involved in query processing, statistics data can not be used
    // NOTE: the original value of isSet have been changed here
    if (pCtx[k].input.colDataSMAIsSet && forwardStep < numOfTotal) {
      pCtx[k].input.colDataSMAIsSet = false;
    }

    if (pCtx[k].isPseudoFunc) {
      SResultRowEntryInfo* pEntryInfo = GET_RES_INFO(&pCtx[k]);

      char* p = GET_ROWCELL_INTERBUF(pEntryInfo);

      SColumnInfoData idata = {0};
      idata.info.type = TSDB_DATA_TYPE_BIGINT;
      idata.info.bytes = tDataTypes[TSDB_DATA_TYPE_BIGINT].bytes;
      idata.pData = p;

      SScalarParam out = {.columnData = &idata};
      SScalarParam tw = {.numOfRows = 5, .columnData = pTimeWindowData};
      pCtx[k].sfp.process(&tw, 1, &out);
      pEntryInfo->numOfRes = 1;
    } else {
      int32_t code = TSDB_CODE_SUCCESS;
      if (functionNeedToExecute(&pCtx[k]) && pCtx[k].fpSet.process != NULL) {
        code = pCtx[k].fpSet.process(&pCtx[k]);

        if (code != TSDB_CODE_SUCCESS) {
          qError("%s apply functions error, code: %s", GET_TASKID(taskInfo), tstrerror(code));
          taskInfo->code = code;
          T_LONG_JMP(taskInfo->env, code);
        }
      }

      // restore it
      functionCtxRestore(&pCtx[k], &status);
    }
  }
}

void functionCtxSave(SqlFunctionCtx* pCtx, SFunctionCtxStatus* pStatus) {
  pStatus->hasAgg = pCtx->input.colDataSMAIsSet;
  pStatus->numOfRows = pCtx->input.numOfRows;
  pStatus->startOffset = pCtx->input.startRowIndex;
}

void functionCtxRestore(SqlFunctionCtx* pCtx, SFunctionCtxStatus* pStatus) {
  pCtx->input.colDataSMAIsSet = pStatus->hasAgg;
  pCtx->input.numOfRows = pStatus->numOfRows;
  pCtx->input.startRowIndex = pStatus->startOffset;
}