Examples with Join org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rel.core.Join used on opensource projects

Example 1 with Join

use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rel.core.Join in project hive by apache.

the class HiveSemiJoinRule method onMatch.

@Override
public void onMatch(RelOptRuleCall call) {
    LOG.debug("Matched HiveSemiJoinRule");
    final Project project = call.rel(0);
    final Join join = call.rel(1);
    final RelNode left = call.rel(2);
    final Aggregate aggregate = call.rel(3);
    final RelOptCluster cluster = join.getCluster();
    final RexBuilder rexBuilder = cluster.getRexBuilder();
    final ImmutableBitSet bits = RelOptUtil.InputFinder.bits(project.getProjects(), null);
    final ImmutableBitSet rightBits = ImmutableBitSet.range(left.getRowType().getFieldCount(), join.getRowType().getFieldCount());
    if (bits.intersects(rightBits)) {
        return;
    }
    final JoinInfo joinInfo = join.analyzeCondition();
    if (!joinInfo.rightSet().equals(ImmutableBitSet.range(aggregate.getGroupCount()))) {
        // By the way, neither a super-set nor a sub-set would work.
        return;
    }
    if (join.getJoinType() == JoinRelType.LEFT) {
        // since for LEFT join we are only interested in rows from LEFT we can get rid of right side
        call.transformTo(call.builder().push(left).project(project.getProjects(), project.getRowType().getFieldNames()).build());
        return;
    }
    if (join.getJoinType() != JoinRelType.INNER) {
        return;
    }
    if (!joinInfo.isEqui()) {
        return;
    }
    LOG.debug("All conditions matched for HiveSemiJoinRule. Going to apply transformation.");
    final List<Integer> newRightKeyBuilder = Lists.newArrayList();
    final List<Integer> aggregateKeys = aggregate.getGroupSet().asList();
    for (int key : joinInfo.rightKeys) {
        newRightKeyBuilder.add(aggregateKeys.get(key));
    }
    final ImmutableIntList newRightKeys = ImmutableIntList.copyOf(newRightKeyBuilder);
    final RelNode newRight = aggregate.getInput();
    final RexNode newCondition = RelOptUtil.createEquiJoinCondition(left, joinInfo.leftKeys, newRight, newRightKeys, rexBuilder);
    RelNode semi = null;
    // is not expected further down the pipeline. see jira for more details
    if (aggregate.getInput() instanceof HepRelVertex && ((HepRelVertex) aggregate.getInput()).getCurrentRel() instanceof Join) {
        Join rightJoin = (Join) (((HepRelVertex) aggregate.getInput()).getCurrentRel());
        List<RexNode> projects = new ArrayList<>();
        for (int i = 0; i < rightJoin.getRowType().getFieldCount(); i++) {
            projects.add(rexBuilder.makeInputRef(rightJoin, i));
        }
        RelNode topProject = call.builder().push(rightJoin).project(projects, rightJoin.getRowType().getFieldNames(), true).build();
        semi = call.builder().push(left).push(topProject).semiJoin(newCondition).build();
    } else {
        semi = call.builder().push(left).push(aggregate.getInput()).semiJoin(newCondition).build();
    }
    call.transformTo(call.builder().push(semi).project(project.getProjects(), project.getRowType().getFieldNames()).build());
}

Example 2 with Join

use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rel.core.Join in project druid by druid-io.

the class DruidSemiJoinRule method onMatch.

@Override
public void onMatch(RelOptRuleCall call) {
    final Project project = call.rel(0);
    final Join join = call.rel(1);
    final DruidRel left = call.rel(2);
    final DruidRel right = call.rel(3);
    final ImmutableBitSet bits = RelOptUtil.InputFinder.bits(project.getProjects(), null);
    final ImmutableBitSet rightBits = ImmutableBitSet.range(left.getRowType().getFieldCount(), join.getRowType().getFieldCount());
    if (bits.intersects(rightBits)) {
        return;
    }
    final JoinInfo joinInfo = join.analyzeCondition();
    final List<Integer> rightDimsOut = new ArrayList<>();
    for (DimensionSpec dimensionSpec : right.getQueryBuilder().getGrouping().getDimensions()) {
        rightDimsOut.add(right.getOutputRowSignature().getRowOrder().indexOf(dimensionSpec.getOutputName()));
    }
    if (!joinInfo.isEqui() || !joinInfo.rightSet().equals(ImmutableBitSet.of(rightDimsOut))) {
        // By the way, neither a super-set nor a sub-set would work.
        return;
    }
    final RelBuilder relBuilder = call.builder();
    final PlannerConfig plannerConfig = left.getPlannerContext().getPlannerConfig();
    if (join.getJoinType() == JoinRelType.LEFT) {
        // Join can be eliminated since the right-hand side cannot have any effect (nothing is being selected,
        // and LEFT means even if there is no match, a left-hand row will still be included).
        relBuilder.push(left);
    } else {
        final DruidSemiJoin druidSemiJoin = DruidSemiJoin.from(left, right, joinInfo.leftKeys, joinInfo.rightKeys, plannerConfig);
        if (druidSemiJoin == null) {
            return;
        }
        // Check maxQueryCount.
        if (plannerConfig.getMaxQueryCount() > 0 && druidSemiJoin.getQueryCount() > plannerConfig.getMaxQueryCount()) {
            return;
        }
        relBuilder.push(druidSemiJoin);
    }
    call.transformTo(relBuilder.project(project.getProjects(), project.getRowType().getFieldNames()).build());
}

Example 3 with Join

use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rel.core.Join in project calcite by apache.

the class AggregateJoinTransposeRule 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 = call.getMetadataQuery();
    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 Mappings.TargetMapping mapping = s == 0 ? Mappings.createIdentity(fieldCount) : Mappings.createShiftMapping(fieldCount + offset, 0, offset, fieldCount);
        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;
            relBuilder.push(joinInput);
            final List<RexNode> projects = new ArrayList<>();
            for (Integer i : belowAggregateKey) {
                projects.add(relBuilder.field(i));
            }
            for (Ord<AggregateCall> aggCall : Ord.zip(aggregate.getAggCallList())) {
                final SqlAggFunction aggregation = aggCall.e.getAggregation();
                final SqlSplittableAggFunction splitter = Preconditions.checkNotNull(aggregation.unwrap(SqlSplittableAggFunction.class));
                if (!aggCall.e.getArgList().isEmpty() && fieldSet.contains(ImmutableBitSet.of(aggCall.e.getArgList()))) {
                    final RexNode singleton = splitter.singleton(rexBuilder, joinInput.getRowType(), aggCall.e.transform(mapping));
                    if (singleton instanceof RexInputRef) {
                        side.split.put(aggCall.i, ((RexInputRef) singleton).getIndex());
                    } else {
                        projects.add(singleton);
                        side.split.put(aggCall.i, projects.size() - 1);
                    }
                }
            }
            relBuilder.project(projects);
            side.newInput = relBuilder.build();
        } else {
            side.aggregate = true;
            List<AggregateCall> belowAggCalls = new ArrayList<>();
            final SqlSplittableAggFunction.Registry<AggregateCall> belowAggCallRegistry = registry(belowAggCalls);
            final int oldGroupKeyCount = aggregate.getGroupCount();
            final int newGroupKeyCount = belowAggregateKey.cardinality();
            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()))) {
                    final AggregateCall splitCall = splitter.split(aggCall.e, mapping);
                    call1 = splitCall.adaptTo(joinInput, splitCall.getArgList(), splitCall.filterArg, oldGroupKeyCount, newGroupKeyCount);
                } 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, 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) {
                final RelDataType rowType = relBuilder.peek().getRowType();
                projects2.add(splitter.singleton(rexBuilder, rowType, 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()), Mappings.apply2(mapping, aggregate.getGroupSets())), newAggCalls);
    }
    call.transformTo(relBuilder.build());
}

Example 4 with Join

use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rel.core.Join in project calcite by apache.

the class JoinPushTransitivePredicatesRule method onMatch.

@Override
public void onMatch(RelOptRuleCall call) {
    Join join = call.rel(0);
    final RelMetadataQuery mq = call.getMetadataQuery();
    RelOptPredicateList preds = mq.getPulledUpPredicates(join);
    if (preds.leftInferredPredicates.isEmpty() && preds.rightInferredPredicates.isEmpty()) {
        return;
    }
    final RexBuilder rexBuilder = join.getCluster().getRexBuilder();
    final RelBuilder relBuilder = call.builder();
    RelNode lChild = join.getLeft();
    if (preds.leftInferredPredicates.size() > 0) {
        RelNode curr = lChild;
        lChild = relBuilder.push(lChild).filter(preds.leftInferredPredicates).build();
        call.getPlanner().onCopy(curr, lChild);
    }
    RelNode rChild = join.getRight();
    if (preds.rightInferredPredicates.size() > 0) {
        RelNode curr = rChild;
        rChild = relBuilder.push(rChild).filter(preds.rightInferredPredicates).build();
        call.getPlanner().onCopy(curr, rChild);
    }
    RelNode newRel = join.copy(join.getTraitSet(), join.getCondition(), lChild, rChild, join.getJoinType(), join.isSemiJoinDone());
    call.getPlanner().onCopy(join, newRel);
    call.transformTo(newRel);
}

Example 5 with Join

use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rel.core.Join in project calcite by apache.

the class JoinToMultiJoinRule method onMatch.

// ~ Methods ----------------------------------------------------------------
public void onMatch(RelOptRuleCall call) {
    final Join origJoin = call.rel(0);
    final RelNode left = call.rel(1);
    final RelNode right = call.rel(2);
    // combine the children MultiJoin inputs into an array of inputs
    // for the new MultiJoin
    final List<ImmutableBitSet> projFieldsList = Lists.newArrayList();
    final List<int[]> joinFieldRefCountsList = Lists.newArrayList();
    final List<RelNode> newInputs = combineInputs(origJoin, left, right, projFieldsList, joinFieldRefCountsList);
    // combine the outer join information from the left and right
    // inputs, and include the outer join information from the current
    // join, if it's a left/right outer join
    final List<Pair<JoinRelType, RexNode>> joinSpecs = Lists.newArrayList();
    combineOuterJoins(origJoin, newInputs, left, right, joinSpecs);
    // pull up the join filters from the children MultiJoinRels and
    // combine them with the join filter associated with this LogicalJoin to
    // form the join filter for the new MultiJoin
    List<RexNode> newJoinFilters = combineJoinFilters(origJoin, left, right);
    // add on the join field reference counts for the join condition
    // associated with this LogicalJoin
    final ImmutableMap<Integer, ImmutableIntList> newJoinFieldRefCountsMap = addOnJoinFieldRefCounts(newInputs, origJoin.getRowType().getFieldCount(), origJoin.getCondition(), joinFieldRefCountsList);
    List<RexNode> newPostJoinFilters = combinePostJoinFilters(origJoin, left, right);
    final RexBuilder rexBuilder = origJoin.getCluster().getRexBuilder();
    RelNode multiJoin = new MultiJoin(origJoin.getCluster(), newInputs, RexUtil.composeConjunction(rexBuilder, newJoinFilters, false), origJoin.getRowType(), origJoin.getJoinType() == JoinRelType.FULL, Pair.right(joinSpecs), Pair.left(joinSpecs), projFieldsList, newJoinFieldRefCountsMap, RexUtil.composeConjunction(rexBuilder, newPostJoinFilters, true));
    call.transformTo(multiJoin);
}