Search in sources :

Example 1 with RowKeyJoinPrel

use of org.apache.drill.exec.planner.physical.RowKeyJoinPrel in project drill by apache.

the class NonCoveringIndexPlanGenerator method convertChild.

@Override
public RelNode convertChild(final RelNode topRel, final RelNode input) throws InvalidRelException {
    if (indexGroupScan == null) {
        logger.error("Null indexgroupScan in NonCoveringIndexPlanGenerator.convertChild");
        return null;
    }
    RelDataType dbscanRowType = convertRowType(origScan.getRowType(), origScan.getCluster().getTypeFactory());
    RelDataType indexScanRowType = FunctionalIndexHelper.convertRowTypeForIndexScan(origScan, indexContext.getOrigMarker(), indexGroupScan, functionInfo);
    DrillDistributionTrait partition = IndexPlanUtils.scanIsPartition(IndexPlanUtils.getGroupScan(origScan)) ? DrillDistributionTrait.RANDOM_DISTRIBUTED : DrillDistributionTrait.SINGLETON;
    ScanPrel indexScanPrel = new ScanPrel(origScan.getCluster(), origScan.getTraitSet().plus(Prel.DRILL_PHYSICAL).plus(partition), indexGroupScan, indexScanRowType, origScan.getTable());
    DbGroupScan origDbGroupScan = (DbGroupScan) IndexPlanUtils.getGroupScan(origScan);
    // right (build) side of the rowkey join: do a distribution of project-filter-indexscan subplan
    RexNode convertedIndexCondition = FunctionalIndexHelper.convertConditionForIndexScan(indexCondition, origScan, indexScanRowType, builder, functionInfo);
    FilterPrel rightIndexFilterPrel = new FilterPrel(indexScanPrel.getCluster(), indexScanPrel.getTraitSet(), indexScanPrel, convertedIndexCondition);
    double finalRowCount = indexGroupScan.getRowCount(indexContext.getOrigCondition(), origScan);
    // project the rowkey column from the index scan
    List<RexNode> rightProjectExprs = Lists.newArrayList();
    // indexGroupScan.getRowKeyOrdinal();
    int rightRowKeyIndex = getRowKeyIndex(indexScanPrel.getRowType(), origScan);
    assert rightRowKeyIndex >= 0;
    rightProjectExprs.add(RexInputRef.of(rightRowKeyIndex, indexScanPrel.getRowType()));
    final List<RelDataTypeField> indexScanFields = indexScanPrel.getRowType().getFieldList();
    final RelDataTypeFactory.FieldInfoBuilder rightFieldTypeBuilder = indexScanPrel.getCluster().getTypeFactory().builder();
    // build the row type for the right Project
    final RelDataTypeField rightRowKeyField = indexScanFields.get(rightRowKeyIndex);
    rightFieldTypeBuilder.add(rightRowKeyField);
    final RelDataType rightProjectRowType = rightFieldTypeBuilder.build();
    final ProjectPrel rightIndexProjectPrel = new ProjectPrel(indexScanPrel.getCluster(), indexScanPrel.getTraitSet(), rightIndexFilterPrel, rightProjectExprs, rightProjectRowType);
    // create a RANGE PARTITION on the right side (this could be removed later during ExcessiveExchangeIdentifier phase
    // if the estimated row count is smaller than slice_target
    final RelNode rangeDistRight = createRangeDistRight(rightIndexProjectPrel, rightRowKeyField, origDbGroupScan);
    // the range partitioning adds an extra column for the partition id but in the final plan we already have a
    // renaming Project for the _id field inserted as part of the JoinPrelRenameVisitor. Thus, we are not inserting
    // a separate Project here.
    final RelNode convertedRight = rangeDistRight;
    // left (probe) side of the rowkey join
    List<SchemaPath> cols = new ArrayList<SchemaPath>(origDbGroupScan.getColumns());
    if (!checkRowKey(cols)) {
        cols.add(origDbGroupScan.getRowKeyPath());
    }
    // Create a restricted groupscan from the primary table's groupscan
    DbGroupScan restrictedGroupScan = (DbGroupScan) origDbGroupScan.getRestrictedScan(cols);
    if (restrictedGroupScan == null) {
        logger.error("Null restricted groupscan in NonCoveringIndexPlanGenerator.convertChild");
        return null;
    }
    // Set left side (restricted scan) row count as rows returned from right side (index scan)
    DrillScanRel rightIdxRel = new DrillScanRel(origScan.getCluster(), origScan.getTraitSet(), origScan.getTable(), origScan.getRowType(), indexContext.getScanColumns());
    double rightIdxRowCount = indexGroupScan.getRowCount(indexCondition, rightIdxRel);
    restrictedGroupScan.setRowCount(null, rightIdxRowCount, rightIdxRowCount);
    RelTraitSet origScanTraitSet = origScan.getTraitSet();
    RelTraitSet restrictedScanTraitSet = origScanTraitSet.plus(Prel.DRILL_PHYSICAL);
    // Create the collation traits for restricted scan based on the index columns under the
    // conditions that (a) the index actually has collation property (e.g hash indexes don't)
    // and (b) if an explicit sort operation is not enforced
    RelCollation collation = null;
    if (indexDesc.getCollation() != null && !settings.isIndexForceSortNonCovering()) {
        collation = IndexPlanUtils.buildCollationNonCoveringIndexScan(indexDesc, indexScanRowType, dbscanRowType, indexContext);
        if (restrictedScanTraitSet.contains(RelCollationTraitDef.INSTANCE)) {
            // replace existing trait
            restrictedScanTraitSet = restrictedScanTraitSet.plus(partition).replace(collation);
        } else {
            // add new one
            restrictedScanTraitSet = restrictedScanTraitSet.plus(partition).plus(collation);
        }
    }
    ScanPrel dbScan = new ScanPrel(origScan.getCluster(), restrictedScanTraitSet, restrictedGroupScan, dbscanRowType, origScan.getTable());
    RelNode lastLeft = dbScan;
    // build the row type for the left Project
    List<RexNode> leftProjectExprs = Lists.newArrayList();
    int leftRowKeyIndex = getRowKeyIndex(dbScan.getRowType(), origScan);
    final RelDataTypeField leftRowKeyField = dbScan.getRowType().getFieldList().get(leftRowKeyIndex);
    final RelDataTypeFactory.FieldInfoBuilder leftFieldTypeBuilder = dbScan.getCluster().getTypeFactory().builder();
    // We are applying the same index condition to primary table's restricted scan. The reason is, the index may be an async
    // index .. i.e it is not synchronously updated along with the primary table update as part of a single transaction, so it
    // is possible that after or during index scan, the primary table rows may have been updated and no longer satisfy the index
    // condition. By re-applying the index condition here, we will ensure non-qualifying records are filtered out.
    // The remainder condition will be applied on top of RowKeyJoin.
    FilterPrel leftIndexFilterPrel = null;
    if (indexDesc.isAsyncIndex()) {
        leftIndexFilterPrel = new FilterPrel(dbScan.getCluster(), dbScan.getTraitSet(), dbScan, indexContext.getOrigCondition());
        lastLeft = leftIndexFilterPrel;
    }
    RelDataType origRowType = origProject == null ? origScan.getRowType() : origProject.getRowType();
    if (origProject != null) {
        // then we also  don't need a project
        // new Project's rowtype is original Project's rowtype [plus rowkey if rowkey is not in original rowtype]
        List<RelDataTypeField> origProjFields = origRowType.getFieldList();
        leftFieldTypeBuilder.addAll(origProjFields);
        // get the exprs from the original Project
        leftProjectExprs.addAll(IndexPlanUtils.getProjects(origProject));
        // add the rowkey IFF rowkey is not in orig scan
        if (getRowKeyIndex(origRowType, origScan) < 0) {
            leftFieldTypeBuilder.add(leftRowKeyField);
            leftProjectExprs.add(RexInputRef.of(leftRowKeyIndex, dbScan.getRowType()));
        }
        final RelDataType leftProjectRowType = leftFieldTypeBuilder.build();
        // build collation in project
        if (!settings.isIndexForceSortNonCovering()) {
            collation = IndexPlanUtils.buildCollationProject(leftProjectExprs, null, dbScan, functionInfo, indexContext);
        }
        final ProjectPrel leftIndexProjectPrel = new ProjectPrel(dbScan.getCluster(), collation != null ? dbScan.getTraitSet().plus(collation) : dbScan.getTraitSet(), leftIndexFilterPrel == null ? dbScan : leftIndexFilterPrel, leftProjectExprs, leftProjectRowType);
        lastLeft = leftIndexProjectPrel;
    }
    final RelTraitSet leftTraits = dbScan.getTraitSet().plus(Prel.DRILL_PHYSICAL);
    // final RelNode convertedLeft = convert(leftIndexProjectPrel, leftTraits);
    final RelNode convertedLeft = Prule.convert(lastLeft, leftTraits);
    // find the rowkey column on the left side of join
    final int leftRowKeyIdx = getRowKeyIndex(convertedLeft.getRowType(), origScan);
    // only rowkey field is being projected from right side
    final int rightRowKeyIdx = 0;
    assert leftRowKeyIdx >= 0;
    List<Integer> leftJoinKeys = ImmutableList.of(leftRowKeyIdx);
    List<Integer> rightJoinKeys = ImmutableList.of(rightRowKeyIdx);
    RexNode joinCondition = RelOptUtil.createEquiJoinCondition(convertedLeft, leftJoinKeys, convertedRight, rightJoinKeys, builder);
    RelNode newRel;
    if (settings.isIndexUseHashJoinNonCovering()) {
        // for hash join, collation will be cleared
        HashJoinPrel hjPrel = new HashJoinPrel(topRel.getCluster(), leftTraits, convertedLeft, convertedRight, joinCondition, JoinRelType.INNER, false, /* no swap */
        null, /* no runtime filter */
        true, /* useful for join-restricted scans */
        JoinControl.DEFAULT);
        newRel = hjPrel;
    } else {
        // if there is collation, add to rowkey join
        RowKeyJoinPrel rjPrel = new RowKeyJoinPrel(topRel.getCluster(), collation != null ? leftTraits.plus(collation) : leftTraits, convertedLeft, convertedRight, joinCondition, JoinRelType.INNER);
        rjPrel.setEstimatedRowCount(finalRowCount);
        newRel = rjPrel;
    }
    final RelDataTypeFactory.FieldInfoBuilder finalFieldTypeBuilder = origScan.getCluster().getTypeFactory().builder();
    List<RelDataTypeField> rjRowFields = newRel.getRowType().getFieldList();
    int toRemoveRowKeyCount = 1;
    if (getRowKeyIndex(origRowType, origScan) < 0) {
        toRemoveRowKeyCount = 2;
    }
    finalFieldTypeBuilder.addAll(rjRowFields.subList(0, rjRowFields.size() - toRemoveRowKeyCount));
    final RelDataType finalProjectRowType = finalFieldTypeBuilder.build();
    List<RexNode> resetExprs = Lists.newArrayList();
    for (int idx = 0; idx < rjRowFields.size() - toRemoveRowKeyCount; ++idx) {
        resetExprs.add(RexInputRef.of(idx, newRel.getRowType()));
    }
    // rewrite the collation for this projectPrel
    final ProjectPrel resetProjectPrel = new ProjectPrel(newRel.getCluster(), newRel.getTraitSet(), newRel, resetExprs, finalProjectRowType);
    newRel = resetProjectPrel;
    if (upperProject != null) {
        RelCollation newCollation = RelCollations.of(RelCollations.EMPTY.getFieldCollations());
        DrillDistributionTrait newDist = null;
        newDist = upperProject.getInput().getTraitSet().getTrait(DrillDistributionTraitDef.INSTANCE);
        if (!settings.isIndexForceSortNonCovering()) {
            newCollation = IndexPlanUtils.buildCollationProject(IndexPlanUtils.getProjects(upperProject), origProject, origScan, functionInfo, indexContext);
        }
        RelTraitSet newProjectTraits = newTraitSet(Prel.DRILL_PHYSICAL, newDist, newCollation);
        ProjectPrel cap = new ProjectPrel(upperProject.getCluster(), newProjectTraits, newRel, IndexPlanUtils.getProjects(upperProject), upperProject.getRowType());
        newRel = cap;
    }
    // whether to remove sort
    if (indexContext.getSort() != null) {
        // unsorted input because ordering is not guaranteed across different parallel inputs.
        if (toRemoveSort(indexContext.getCollation(), newRel.getTraitSet().getTrait(RelCollationTraitDef.INSTANCE))) {
            ((IndexGroupScan) indexScanPrel.getGroupScan()).setParallelizationWidth(1);
        }
        newRel = getSortNode(indexContext, newRel, false, true, true);
        Preconditions.checkArgument(newRel != null);
    }
    RelNode finalRel = Prule.convert(newRel, newRel.getTraitSet());
    logger.debug("NonCoveringIndexPlanGenerator got finalRel {} from origScan {}", finalRel.toString(), origScan.toString());
    return finalRel;
}
Also used : DrillScanRel(org.apache.drill.exec.planner.logical.DrillScanRel) ArrayList(java.util.ArrayList) RelDataType(org.apache.calcite.rel.type.RelDataType) RelTraitSet(org.apache.calcite.plan.RelTraitSet) RowKeyJoinPrel(org.apache.drill.exec.planner.physical.RowKeyJoinPrel) SchemaPath(org.apache.drill.common.expression.SchemaPath) RelDataTypeFactory(org.apache.calcite.rel.type.RelDataTypeFactory) FilterPrel(org.apache.drill.exec.planner.physical.FilterPrel) ProjectPrel(org.apache.drill.exec.planner.physical.ProjectPrel) ScanPrel(org.apache.drill.exec.planner.physical.ScanPrel) DrillDistributionTrait(org.apache.drill.exec.planner.physical.DrillDistributionTrait) IndexGroupScan(org.apache.drill.exec.physical.base.IndexGroupScan) RelDataTypeField(org.apache.calcite.rel.type.RelDataTypeField) RelCollation(org.apache.calcite.rel.RelCollation) RelNode(org.apache.calcite.rel.RelNode) HashJoinPrel(org.apache.drill.exec.planner.physical.HashJoinPrel) DbGroupScan(org.apache.drill.exec.physical.base.DbGroupScan) RexNode(org.apache.calcite.rex.RexNode)

Example 2 with RowKeyJoinPrel

use of org.apache.drill.exec.planner.physical.RowKeyJoinPrel in project drill by apache.

the class IndexIntersectPlanGenerator method buildRowKeyJoin.

public RelNode buildRowKeyJoin(RelNode left, RelNode right, boolean isRowKeyJoin, int htControl) throws InvalidRelException {
    final int leftRowKeyIdx = getRowKeyIndex(left.getRowType(), origScan);
    // only rowkey field is being projected from right side
    final int rightRowKeyIdx = 0;
    assert leftRowKeyIdx >= 0;
    List<Integer> leftJoinKeys = ImmutableList.of(leftRowKeyIdx);
    List<Integer> rightJoinKeys = ImmutableList.of(rightRowKeyIdx);
    logger.trace(String.format("buildRowKeyJoin: leftIdx: %d, rightIdx: %d", leftRowKeyIdx, rightRowKeyIdx));
    RexNode joinCondition = RelOptUtil.createEquiJoinCondition(left, leftJoinKeys, right, rightJoinKeys, builder);
    if (isRowKeyJoin) {
        RelNode newRel;
        if (settings.isIndexUseHashJoinNonCovering()) {
            HashJoinPrel hjPrel = new HashJoinPrel(left.getCluster(), left.getTraitSet(), left, right, joinCondition, JoinRelType.INNER, false, /* no swap */
            null, /* no runtime filter */
            isRowKeyJoin, htControl);
            newRel = hjPrel;
        } else {
            RowKeyJoinPrel rjPrel = new RowKeyJoinPrel(left.getCluster(), left.getTraitSet(), left, right, joinCondition, JoinRelType.INNER);
            newRel = rjPrel;
        }
        // since there is a restricted Scan on left side, assume original project
        return buildOriginalProject(newRel);
    } else {
        // there is no restricted scan on left, do a regular rowkey join
        HashJoinPrel hjPrel = new HashJoinPrel(left.getCluster(), left.getTraitSet(), left, right, joinCondition, JoinRelType.INNER, false, /* no swap */
        null, /* no runtime filter */
        isRowKeyJoin, htControl);
        return buildRowKeyProject(hjPrel, leftRowKeyIdx);
    }
}
Also used : RelNode(org.apache.calcite.rel.RelNode) RowKeyJoinPrel(org.apache.drill.exec.planner.physical.RowKeyJoinPrel) HashJoinPrel(org.apache.drill.exec.planner.physical.HashJoinPrel) RexNode(org.apache.calcite.rex.RexNode)

Example 3 with RowKeyJoinPrel

use of org.apache.drill.exec.planner.physical.RowKeyJoinPrel in project drill by apache.

the class MapRDBPushLimitIntoScan method doPushLimitIntoRowKeyJoin.

protected void doPushLimitIntoRowKeyJoin(RelOptRuleCall call, LimitPrel limit, final ProjectPrel project, RowKeyJoinPrel join) {
    final RelNode newChild;
    try {
        RelNode left = join.getLeft();
        RelNode right = join.getRight();
        final RelNode limitOnLeft = new LimitPrel(left.getCluster(), left.getTraitSet(), left, limit.getOffset(), limit.getFetch());
        RowKeyJoinPrel newJoin = new RowKeyJoinPrel(join.getCluster(), join.getTraitSet(), limitOnLeft, right, join.getCondition(), join.getJoinType());
        if (project != null) {
            final ProjectPrel newProject = new ProjectPrel(project.getCluster(), project.getTraitSet(), newJoin, project.getProjects(), project.getRowType());
            newChild = newProject;
        } else {
            newChild = newJoin;
        }
        call.transformTo(newChild);
        logger.debug("pushLimitIntoRowKeyJoin: Pushed limit on left side of Join " + join.toString());
    } catch (Exception e) {
        logger.warn("pushLimitIntoRowKeyJoin: Exception while trying limit pushdown!", e);
    }
}
Also used : ProjectPrel(org.apache.drill.exec.planner.physical.ProjectPrel) RelNode(org.apache.calcite.rel.RelNode) RowKeyJoinPrel(org.apache.drill.exec.planner.physical.RowKeyJoinPrel) LimitPrel(org.apache.drill.exec.planner.physical.LimitPrel)

Aggregations

RelNode (org.apache.calcite.rel.RelNode)3 RowKeyJoinPrel (org.apache.drill.exec.planner.physical.RowKeyJoinPrel)3 RexNode (org.apache.calcite.rex.RexNode)2 HashJoinPrel (org.apache.drill.exec.planner.physical.HashJoinPrel)2 ProjectPrel (org.apache.drill.exec.planner.physical.ProjectPrel)2 ArrayList (java.util.ArrayList)1 RelTraitSet (org.apache.calcite.plan.RelTraitSet)1 RelCollation (org.apache.calcite.rel.RelCollation)1 RelDataType (org.apache.calcite.rel.type.RelDataType)1 RelDataTypeFactory (org.apache.calcite.rel.type.RelDataTypeFactory)1 RelDataTypeField (org.apache.calcite.rel.type.RelDataTypeField)1 SchemaPath (org.apache.drill.common.expression.SchemaPath)1 DbGroupScan (org.apache.drill.exec.physical.base.DbGroupScan)1 IndexGroupScan (org.apache.drill.exec.physical.base.IndexGroupScan)1 DrillScanRel (org.apache.drill.exec.planner.logical.DrillScanRel)1 DrillDistributionTrait (org.apache.drill.exec.planner.physical.DrillDistributionTrait)1 FilterPrel (org.apache.drill.exec.planner.physical.FilterPrel)1 LimitPrel (org.apache.drill.exec.planner.physical.LimitPrel)1 ScanPrel (org.apache.drill.exec.planner.physical.ScanPrel)1