Examples with Aggregate org.apache.calcite.rel.core.Aggregate used on opensource projects

Example 1 with Aggregate

use of org.apache.calcite.rel.core.Aggregate 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 Aggregate

use of org.apache.calcite.rel.core.Aggregate in project flink by apache.

the class FlinkAggregateExpandDistinctAggregatesRule method onMatch.

//~ Methods ----------------------------------------------------------------
public void onMatch(RelOptRuleCall call) {
    final Aggregate aggregate = call.rel(0);
    if (!aggregate.containsDistinctCall()) {
        return;
    }
    // Find all of the agg expressions. We use a LinkedHashSet to ensure
    // determinism.
    int nonDistinctCount = 0;
    int distinctCount = 0;
    int filterCount = 0;
    int unsupportedAggCount = 0;
    final Set<Pair<List<Integer>, Integer>> argLists = new LinkedHashSet<>();
    for (AggregateCall aggCall : aggregate.getAggCallList()) {
        if (aggCall.filterArg >= 0) {
            ++filterCount;
        }
        if (!aggCall.isDistinct()) {
            ++nonDistinctCount;
            if (!(aggCall.getAggregation() instanceof SqlCountAggFunction || aggCall.getAggregation() instanceof SqlSumAggFunction || aggCall.getAggregation() instanceof SqlMinMaxAggFunction)) {
                ++unsupportedAggCount;
            }
            continue;
        }
        ++distinctCount;
        argLists.add(Pair.of(aggCall.getArgList(), aggCall.filterArg));
    }
    Preconditions.checkState(argLists.size() > 0, "containsDistinctCall lied");
    // arguments then we can use a more efficient form.
    if (nonDistinctCount == 0 && argLists.size() == 1) {
        final Pair<List<Integer>, Integer> pair = Iterables.getOnlyElement(argLists);
        final RelBuilder relBuilder = call.builder();
        convertMonopole(relBuilder, aggregate, pair.left, pair.right);
        call.transformTo(relBuilder.build());
        return;
    }
    if (useGroupingSets) {
        rewriteUsingGroupingSets(call, aggregate, argLists);
        return;
    }
    // we can generate multi-phase aggregates
    if (// one distinct aggregate
    distinctCount == 1 && // no filter
    filterCount == 0 && // sum/min/max/count in non-distinct aggregate
    unsupportedAggCount == 0 && nonDistinctCount > 0) {
        // one or more non-distinct aggregates
        final RelBuilder relBuilder = call.builder();
        convertSingletonDistinct(relBuilder, aggregate, argLists);
        call.transformTo(relBuilder.build());
        return;
    }
    // Create a list of the expressions which will yield the final result.
    // Initially, the expressions point to the input field.
    final List<RelDataTypeField> aggFields = aggregate.getRowType().getFieldList();
    final List<RexInputRef> refs = new ArrayList<>();
    final List<String> fieldNames = aggregate.getRowType().getFieldNames();
    final ImmutableBitSet groupSet = aggregate.getGroupSet();
    final int groupAndIndicatorCount = aggregate.getGroupCount() + aggregate.getIndicatorCount();
    for (int i : Util.range(groupAndIndicatorCount)) {
        refs.add(RexInputRef.of(i, aggFields));
    }
    // Aggregate the original relation, including any non-distinct aggregates.
    final List<AggregateCall> newAggCallList = new ArrayList<>();
    int i = -1;
    for (AggregateCall aggCall : aggregate.getAggCallList()) {
        ++i;
        if (aggCall.isDistinct()) {
            refs.add(null);
            continue;
        }
        refs.add(new RexInputRef(groupAndIndicatorCount + newAggCallList.size(), aggFields.get(groupAndIndicatorCount + i).getType()));
        newAggCallList.add(aggCall);
    }
    // In the case where there are no non-distinct aggregates (regardless of
    // whether there are group bys), there's no need to generate the
    // extra aggregate and join.
    final RelBuilder relBuilder = call.builder();
    relBuilder.push(aggregate.getInput());
    int n = 0;
    if (!newAggCallList.isEmpty()) {
        final RelBuilder.GroupKey groupKey = relBuilder.groupKey(groupSet, aggregate.indicator, aggregate.getGroupSets());
        relBuilder.aggregate(groupKey, newAggCallList);
        ++n;
    }
    // set of operands.
    for (Pair<List<Integer>, Integer> argList : argLists) {
        doRewrite(relBuilder, aggregate, n++, argList.left, argList.right, refs);
    }
    relBuilder.project(refs, fieldNames);
    call.transformTo(relBuilder.build());
}

Example 3 with Aggregate

use of org.apache.calcite.rel.core.Aggregate in project hive by apache.

the class HiveExpandDistinctAggregatesRule method convert.

/**
 * Converts an aggregate relational expression that contains only
 * count(distinct) to grouping sets with count. For example select
 * count(distinct department_id), count(distinct gender), count(distinct
 * education_level) from employee; can be transformed to
 * select
 * count(case when i=1 and department_id is not null then 1 else null end) as c0,
 * count(case when i=2 and gender is not null then 1 else null end) as c1,
 * count(case when i=4 and education_level is not null then 1 else null end) as c2
 * from (select
 * grouping__id as i, department_id, gender, education_level from employee
 * group by department_id, gender, education_level grouping sets
 * (department_id, gender, education_level))subq;
 * @throws CalciteSemanticException
 */
private RelNode convert(Aggregate aggregate, List<List<Integer>> argList, List<Integer> sourceOfForCountDistinct) throws CalciteSemanticException {
    // we use this map to map the position of argList to the position of grouping set
    Map<Integer, Integer> map = new HashMap<>();
    List<List<Integer>> cleanArgList = new ArrayList<>();
    final Aggregate groupingSets = createGroupingSets(aggregate, argList, cleanArgList, map, sourceOfForCountDistinct);
    return createCount(groupingSets, argList, cleanArgList, map, sourceOfForCountDistinct);
}

Example 4 with Aggregate

use of org.apache.calcite.rel.core.Aggregate in project calcite by apache.

the class AggregateExpandDistinctAggregatesRule method onMatch.

// ~ Methods ----------------------------------------------------------------
public void onMatch(RelOptRuleCall call) {
    final Aggregate aggregate = call.rel(0);
    if (!aggregate.containsDistinctCall()) {
        return;
    }
    // Find all of the agg expressions. We use a LinkedHashSet to ensure determinism.
    // find all aggregate calls without distinct
    int nonDistinctAggCallCount = 0;
    int filterCount = 0;
    int unsupportedNonDistinctAggCallCount = 0;
    final Set<Pair<List<Integer>, Integer>> argLists = new LinkedHashSet<>();
    for (AggregateCall aggCall : aggregate.getAggCallList()) {
        if (aggCall.filterArg >= 0) {
            ++filterCount;
        }
        if (!aggCall.isDistinct()) {
            ++nonDistinctAggCallCount;
            final SqlKind aggCallKind = aggCall.getAggregation().getKind();
            // We only support COUNT/SUM/MIN/MAX for the "single" count distinct optimization
            switch(aggCallKind) {
                case COUNT:
                case SUM:
                case SUM0:
                case MIN:
                case MAX:
                    break;
                default:
                    ++unsupportedNonDistinctAggCallCount;
            }
        } else {
            argLists.add(Pair.of(aggCall.getArgList(), aggCall.filterArg));
        }
    }
    final int distinctAggCallCount = aggregate.getAggCallList().size() - nonDistinctAggCallCount;
    Preconditions.checkState(argLists.size() > 0, "containsDistinctCall lied");
    // arguments then we can use a more efficient form.
    if (nonDistinctAggCallCount == 0 && argLists.size() == 1 && aggregate.getGroupType() == Group.SIMPLE) {
        final Pair<List<Integer>, Integer> pair = Iterables.getOnlyElement(argLists);
        final RelBuilder relBuilder = call.builder();
        convertMonopole(relBuilder, aggregate, pair.left, pair.right);
        call.transformTo(relBuilder.build());
        return;
    }
    if (useGroupingSets) {
        rewriteUsingGroupingSets(call, aggregate);
        return;
    }
    // we can generate multi-phase aggregates
    if (// one distinct aggregate
    distinctAggCallCount == 1 && // no filter
    filterCount == 0 && // sum/min/max/count in non-distinct aggregate
    unsupportedNonDistinctAggCallCount == 0 && nonDistinctAggCallCount > 0) {
        // one or more non-distinct aggregates
        final RelBuilder relBuilder = call.builder();
        convertSingletonDistinct(relBuilder, aggregate, argLists);
        call.transformTo(relBuilder.build());
        return;
    }
    // Create a list of the expressions which will yield the final result.
    // Initially, the expressions point to the input field.
    final List<RelDataTypeField> aggFields = aggregate.getRowType().getFieldList();
    final List<RexInputRef> refs = new ArrayList<>();
    final List<String> fieldNames = aggregate.getRowType().getFieldNames();
    final ImmutableBitSet groupSet = aggregate.getGroupSet();
    final int groupAndIndicatorCount = aggregate.getGroupCount() + aggregate.getIndicatorCount();
    for (int i : Util.range(groupAndIndicatorCount)) {
        refs.add(RexInputRef.of(i, aggFields));
    }
    // Aggregate the original relation, including any non-distinct aggregates.
    final List<AggregateCall> newAggCallList = new ArrayList<>();
    int i = -1;
    for (AggregateCall aggCall : aggregate.getAggCallList()) {
        ++i;
        if (aggCall.isDistinct()) {
            refs.add(null);
            continue;
        }
        refs.add(new RexInputRef(groupAndIndicatorCount + newAggCallList.size(), aggFields.get(groupAndIndicatorCount + i).getType()));
        newAggCallList.add(aggCall);
    }
    // In the case where there are no non-distinct aggregates (regardless of
    // whether there are group bys), there's no need to generate the
    // extra aggregate and join.
    final RelBuilder relBuilder = call.builder();
    relBuilder.push(aggregate.getInput());
    int n = 0;
    if (!newAggCallList.isEmpty()) {
        final RelBuilder.GroupKey groupKey = relBuilder.groupKey(groupSet, aggregate.getGroupSets());
        relBuilder.aggregate(groupKey, newAggCallList);
        ++n;
    }
    // set of operands.
    for (Pair<List<Integer>, Integer> argList : argLists) {
        doRewrite(relBuilder, aggregate, n++, argList.left, argList.right, refs);
    }
    relBuilder.project(refs, fieldNames);
    call.transformTo(relBuilder.build());
}

Example 5 with Aggregate

use of org.apache.calcite.rel.core.Aggregate 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());
}