Examples with RelMetadataQuery org.apache.calcite.rel.metadata.RelMetadataQuery used on opensource projects

Example 11 with RelMetadataQuery

use of org.apache.calcite.rel.metadata.RelMetadataQuery in project flink by apache.

the class FlinkAggregateJoinTransposeRule method onMatch.

public void onMatch(RelOptRuleCall call) {
    final Aggregate aggregate = call.rel(0);
    final Join join = call.rel(1);
    final RexBuilder rexBuilder = aggregate.getCluster().getRexBuilder();
    final RelBuilder relBuilder = call.builder();
    // If any aggregate call has a filter, bail out
    for (AggregateCall aggregateCall : aggregate.getAggCallList()) {
        if (aggregateCall.getAggregation().unwrap(SqlSplittableAggFunction.class) == null) {
            return;
        }
        if (aggregateCall.filterArg >= 0) {
            return;
        }
    }
    // aggregate operator
    if (join.getJoinType() != JoinRelType.INNER) {
        return;
    }
    if (!allowFunctions && !aggregate.getAggCallList().isEmpty()) {
        return;
    }
    // Do the columns used by the join appear in the output of the aggregate?
    final ImmutableBitSet aggregateColumns = aggregate.getGroupSet();
    final RelMetadataQuery mq = RelMetadataQuery.instance();
    final ImmutableBitSet keyColumns = keyColumns(aggregateColumns, mq.getPulledUpPredicates(join).pulledUpPredicates);
    final ImmutableBitSet joinColumns = RelOptUtil.InputFinder.bits(join.getCondition());
    final boolean allColumnsInAggregate = keyColumns.contains(joinColumns);
    final ImmutableBitSet belowAggregateColumns = aggregateColumns.union(joinColumns);
    // Split join condition
    final List<Integer> leftKeys = Lists.newArrayList();
    final List<Integer> rightKeys = Lists.newArrayList();
    final List<Boolean> filterNulls = Lists.newArrayList();
    RexNode nonEquiConj = RelOptUtil.splitJoinCondition(join.getLeft(), join.getRight(), join.getCondition(), leftKeys, rightKeys, filterNulls);
    // If it contains non-equi join conditions, we bail out
    if (!nonEquiConj.isAlwaysTrue()) {
        return;
    }
    // Push each aggregate function down to each side that contains all of its
    // arguments. Note that COUNT(*), because it has no arguments, can go to
    // both sides.
    final Map<Integer, Integer> map = new HashMap<>();
    final List<Side> sides = new ArrayList<>();
    int uniqueCount = 0;
    int offset = 0;
    int belowOffset = 0;
    for (int s = 0; s < 2; s++) {
        final Side side = new Side();
        final RelNode joinInput = join.getInput(s);
        int fieldCount = joinInput.getRowType().getFieldCount();
        final ImmutableBitSet fieldSet = ImmutableBitSet.range(offset, offset + fieldCount);
        final ImmutableBitSet belowAggregateKeyNotShifted = belowAggregateColumns.intersect(fieldSet);
        for (Ord<Integer> c : Ord.zip(belowAggregateKeyNotShifted)) {
            map.put(c.e, belowOffset + c.i);
        }
        final ImmutableBitSet belowAggregateKey = belowAggregateKeyNotShifted.shift(-offset);
        final boolean unique;
        if (!allowFunctions) {
            assert aggregate.getAggCallList().isEmpty();
            // If there are no functions, it doesn't matter as much whether we
            // aggregate the inputs before the join, because there will not be
            // any functions experiencing a cartesian product effect.
            //
            // But finding out whether the input is already unique requires a call
            // to areColumnsUnique that currently (until [CALCITE-1048] "Make
            // metadata more robust" is fixed) places a heavy load on
            // the metadata system.
            //
            // So we choose to imagine the the input is already unique, which is
            // untrue but harmless.
            //
            Util.discard(Bug.CALCITE_1048_FIXED);
            unique = true;
        } else {
            final Boolean unique0 = mq.areColumnsUnique(joinInput, belowAggregateKey);
            unique = unique0 != null && unique0;
        }
        if (unique) {
            ++uniqueCount;
            side.aggregate = false;
            side.newInput = joinInput;
        } else {
            side.aggregate = true;
            List<AggregateCall> belowAggCalls = new ArrayList<>();
            final SqlSplittableAggFunction.Registry<AggregateCall> belowAggCallRegistry = registry(belowAggCalls);
            final Mappings.TargetMapping mapping = s == 0 ? Mappings.createIdentity(fieldCount) : Mappings.createShiftMapping(fieldCount + offset, 0, offset, fieldCount);
            for (Ord<AggregateCall> aggCall : Ord.zip(aggregate.getAggCallList())) {
                final SqlAggFunction aggregation = aggCall.e.getAggregation();
                final SqlSplittableAggFunction splitter = Preconditions.checkNotNull(aggregation.unwrap(SqlSplittableAggFunction.class));
                final AggregateCall call1;
                if (fieldSet.contains(ImmutableBitSet.of(aggCall.e.getArgList()))) {
                    call1 = splitter.split(aggCall.e, mapping);
                } else {
                    call1 = splitter.other(rexBuilder.getTypeFactory(), aggCall.e);
                }
                if (call1 != null) {
                    side.split.put(aggCall.i, belowAggregateKey.cardinality() + belowAggCallRegistry.register(call1));
                }
            }
            side.newInput = relBuilder.push(joinInput).aggregate(relBuilder.groupKey(belowAggregateKey, false, null), belowAggCalls).build();
        }
        offset += fieldCount;
        belowOffset += side.newInput.getRowType().getFieldCount();
        sides.add(side);
    }
    if (uniqueCount == 2) {
        // invocation of this rule; if we continue we might loop forever.
        return;
    }
    // Update condition
    final Mapping mapping = (Mapping) Mappings.target(new Function<Integer, Integer>() {

        public Integer apply(Integer a0) {
            return map.get(a0);
        }
    }, join.getRowType().getFieldCount(), belowOffset);
    final RexNode newCondition = RexUtil.apply(mapping, join.getCondition());
    // Create new join
    relBuilder.push(sides.get(0).newInput).push(sides.get(1).newInput).join(join.getJoinType(), newCondition);
    // Aggregate above to sum up the sub-totals
    final List<AggregateCall> newAggCalls = new ArrayList<>();
    final int groupIndicatorCount = aggregate.getGroupCount() + aggregate.getIndicatorCount();
    final int newLeftWidth = sides.get(0).newInput.getRowType().getFieldCount();
    final List<RexNode> projects = new ArrayList<>(rexBuilder.identityProjects(relBuilder.peek().getRowType()));
    for (Ord<AggregateCall> aggCall : Ord.zip(aggregate.getAggCallList())) {
        final SqlAggFunction aggregation = aggCall.e.getAggregation();
        final SqlSplittableAggFunction splitter = Preconditions.checkNotNull(aggregation.unwrap(SqlSplittableAggFunction.class));
        final Integer leftSubTotal = sides.get(0).split.get(aggCall.i);
        final Integer rightSubTotal = sides.get(1).split.get(aggCall.i);
        newAggCalls.add(splitter.topSplit(rexBuilder, registry(projects), groupIndicatorCount, relBuilder.peek().getRowType(), aggCall.e, leftSubTotal == null ? -1 : leftSubTotal, rightSubTotal == null ? -1 : rightSubTotal + newLeftWidth));
    }
    relBuilder.project(projects);
    boolean aggConvertedToProjects = false;
    if (allColumnsInAggregate) {
        // let's see if we can convert aggregate into projects
        List<RexNode> projects2 = new ArrayList<>();
        for (int key : Mappings.apply(mapping, aggregate.getGroupSet())) {
            projects2.add(relBuilder.field(key));
        }
        for (AggregateCall newAggCall : newAggCalls) {
            final SqlSplittableAggFunction splitter = newAggCall.getAggregation().unwrap(SqlSplittableAggFunction.class);
            if (splitter != null) {
                projects2.add(splitter.singleton(rexBuilder, relBuilder.peek().getRowType(), newAggCall));
            }
        }
        if (projects2.size() == aggregate.getGroupSet().cardinality() + newAggCalls.size()) {
            // We successfully converted agg calls into projects.
            relBuilder.project(projects2);
            aggConvertedToProjects = true;
        }
    }
    if (!aggConvertedToProjects) {
        relBuilder.aggregate(relBuilder.groupKey(Mappings.apply(mapping, aggregate.getGroupSet()), aggregate.indicator, Mappings.apply2(mapping, aggregate.getGroupSets())), newAggCalls);
    }
    call.transformTo(relBuilder.build());
}

Example 12 with RelMetadataQuery

use of org.apache.calcite.rel.metadata.RelMetadataQuery in project hive by apache.

the class HiveAlgorithmsUtil method isFittingIntoMemory.

public static boolean isFittingIntoMemory(Double maxSize, RelNode input, int buckets) {
    final RelMetadataQuery mq = input.getCluster().getMetadataQuery();
    Double currentMemory = mq.cumulativeMemoryWithinPhase(input);
    if (currentMemory != null) {
        if (currentMemory / buckets > maxSize) {
            return false;
        }
        return true;
    }
    return false;
}

Example 13 with RelMetadataQuery

use of org.apache.calcite.rel.metadata.RelMetadataQuery in project hive by apache.

the class HiveAlgorithmsUtil method getJoinMemory.

public static Double getJoinMemory(HiveJoin join, MapJoinStreamingRelation streamingSide) {
    Double memory = 0.0;
    final RelMetadataQuery mq = join.getCluster().getMetadataQuery();
    if (streamingSide == MapJoinStreamingRelation.NONE || streamingSide == MapJoinStreamingRelation.RIGHT_RELATION) {
        // Left side
        final Double leftAvgRowSize = mq.getAverageRowSize(join.getLeft());
        final Double leftRowCount = mq.getRowCount(join.getLeft());
        if (leftAvgRowSize == null || leftRowCount == null) {
            return null;
        }
        memory += leftAvgRowSize * leftRowCount;
    }
    if (streamingSide == MapJoinStreamingRelation.NONE || streamingSide == MapJoinStreamingRelation.LEFT_RELATION) {
        // Right side
        final Double rightAvgRowSize = mq.getAverageRowSize(join.getRight());
        final Double rightRowCount = mq.getRowCount(join.getRight());
        if (rightAvgRowSize == null || rightRowCount == null) {
            return null;
        }
        memory += rightAvgRowSize * rightRowCount;
    }
    return memory;
}

Example 14 with RelMetadataQuery

use of org.apache.calcite.rel.metadata.RelMetadataQuery in project hive by apache.

the class HiveAlgorithmsUtil method getSplitCountWithRepartition.

public static Integer getSplitCountWithRepartition(HiveJoin join) {
    final Double maxSplitSize = join.getCluster().getPlanner().getContext().unwrap(HiveAlgorithmsConf.class).getMaxSplitSize();
    // We repartition: new number of splits
    final RelMetadataQuery mq = join.getCluster().getMetadataQuery();
    final Double averageRowSize = mq.getAverageRowSize(join);
    final Double rowCount = mq.getRowCount(join);
    if (averageRowSize == null || rowCount == null) {
        return null;
    }
    final Double totalSize = averageRowSize * rowCount;
    final Double splitCount = totalSize / maxSplitSize;
    return splitCount.intValue();
}

Example 15 with RelMetadataQuery

use of org.apache.calcite.rel.metadata.RelMetadataQuery in project hive by apache.

the class HiveOnTezCostModel method getAggregateCost.

@Override
public RelOptCost getAggregateCost(HiveAggregate aggregate) {
    if (aggregate.isBucketedInput()) {
        return HiveCost.FACTORY.makeZeroCost();
    } else {
        final RelMetadataQuery mq = aggregate.getCluster().getMetadataQuery();
        // 1. Sum of input cardinalities
        final Double rCount = mq.getRowCount(aggregate.getInput());
        if (rCount == null) {
            return null;
        }
        // 2. CPU cost = sorting cost
        final double cpuCost = algoUtils.computeSortCPUCost(rCount);
        // 3. IO cost = cost of writing intermediary results to local FS +
        // cost of reading from local FS for transferring to GBy +
        // cost of transferring map outputs to GBy operator
        final Double rAverageSize = mq.getAverageRowSize(aggregate.getInput());
        if (rAverageSize == null) {
            return null;
        }
        final double ioCost = algoUtils.computeSortIOCost(new Pair<Double, Double>(rCount, rAverageSize));
        // 4. Result
        return HiveCost.FACTORY.makeCost(rCount, cpuCost, ioCost);
    }
}