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

Example 11 with AggregateCall

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

the class RelMdSize method averageColumnSizes.

public List<Double> averageColumnSizes(Aggregate rel, RelMetadataQuery mq) {
    final List<Double> inputColumnSizes = mq.getAverageColumnSizesNotNull(rel.getInput());
    final ImmutableList.Builder<Double> list = ImmutableList.builder();
    for (int key : rel.getGroupSet()) {
        list.add(inputColumnSizes.get(key));
    }
    for (AggregateCall aggregateCall : rel.getAggCallList()) {
        list.add(averageTypeValueSize(aggregateCall.type));
    }
    return list.build();
}

Example 12 with AggregateCall

use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rel.core.AggregateCall 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 13 with AggregateCall

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

the class AggregateExpandDistinctAggregatesRule method doRewrite.

/**
 * Converts all distinct aggregate calls to a given set of arguments.
 *
 * <p>This method is called several times, one for each set of arguments.
 * Each time it is called, it generates a JOIN to a new SELECT DISTINCT
 * relational expression, and modifies the set of top-level calls.
 *
 * @param aggregate Original aggregate
 * @param n         Ordinal of this in a join. {@code relBuilder} contains the
 *                  input relational expression (either the original
 *                  aggregate, the output from the previous call to this
 *                  method. {@code n} is 0 if we're converting the
 *                  first distinct aggregate in a query with no non-distinct
 *                  aggregates)
 * @param argList   Arguments to the distinct aggregate function
 * @param filterArg Argument that filters input to aggregate function, or -1
 * @param refs      Array of expressions which will be the projected by the
 *                  result of this rule. Those relating to this arg list will
 *                  be modified  @return Relational expression
 */
private void doRewrite(RelBuilder relBuilder, Aggregate aggregate, int n, List<Integer> argList, int filterArg, List<RexInputRef> refs) {
    final RexBuilder rexBuilder = aggregate.getCluster().getRexBuilder();
    final List<RelDataTypeField> leftFields;
    if (n == 0) {
        leftFields = null;
    } else {
        leftFields = relBuilder.peek().getRowType().getFieldList();
    }
    // Aggregate(
    // child,
    // {COUNT(DISTINCT 1), SUM(DISTINCT 1), SUM(2)})
    // 
    // becomes
    // 
    // Aggregate(
    // Join(
    // child,
    // Aggregate(child, < all columns > {}),
    // INNER,
    // <f2 = f5>))
    // 
    // E.g.
    // SELECT deptno, SUM(DISTINCT sal), COUNT(DISTINCT gender), MAX(age)
    // FROM Emps
    // GROUP BY deptno
    // 
    // becomes
    // 
    // SELECT e.deptno, adsal.sum_sal, adgender.count_gender, e.max_age
    // FROM (
    // SELECT deptno, MAX(age) as max_age
    // FROM Emps GROUP BY deptno) AS e
    // JOIN (
    // SELECT deptno, COUNT(gender) AS count_gender FROM (
    // SELECT DISTINCT deptno, gender FROM Emps) AS dgender
    // GROUP BY deptno) AS adgender
    // ON e.deptno = adgender.deptno
    // JOIN (
    // SELECT deptno, SUM(sal) AS sum_sal FROM (
    // SELECT DISTINCT deptno, sal FROM Emps) AS dsal
    // GROUP BY deptno) AS adsal
    // ON e.deptno = adsal.deptno
    // GROUP BY e.deptno
    // 
    // Note that if a query contains no non-distinct aggregates, then the
    // very first join/group by is omitted.  In the example above, if
    // MAX(age) is removed, then the sub-select of "e" is not needed, and
    // instead the two other group by's are joined to one another.
    // Project the columns of the GROUP BY plus the arguments
    // to the agg function.
    final Map<Integer, Integer> sourceOf = new HashMap<>();
    createSelectDistinct(relBuilder, aggregate, argList, filterArg, sourceOf);
    // Now compute the aggregate functions on top of the distinct dataset.
    // Each distinct agg becomes a non-distinct call to the corresponding
    // field from the right; for example,
    // "COUNT(DISTINCT e.sal)"
    // becomes
    // "COUNT(distinct_e.sal)".
    final List<AggregateCall> aggCallList = new ArrayList<>();
    final List<AggregateCall> aggCalls = aggregate.getAggCallList();
    final int groupAndIndicatorCount = aggregate.getGroupCount() + aggregate.getIndicatorCount();
    int i = groupAndIndicatorCount - 1;
    for (AggregateCall aggCall : aggCalls) {
        ++i;
        // COUNT(DISTINCT gender) or SUM(sal).
        if (!aggCall.isDistinct()) {
            continue;
        }
        if (!aggCall.getArgList().equals(argList)) {
            continue;
        }
        // Re-map arguments.
        final int argCount = aggCall.getArgList().size();
        final List<Integer> newArgs = new ArrayList<>(argCount);
        for (int j = 0; j < argCount; j++) {
            final Integer arg = aggCall.getArgList().get(j);
            newArgs.add(sourceOf.get(arg));
        }
        final int newFilterArg = aggCall.filterArg >= 0 ? sourceOf.get(aggCall.filterArg) : -1;
        final AggregateCall newAggCall = AggregateCall.create(aggCall.getAggregation(), false, aggCall.isApproximate(), newArgs, newFilterArg, aggCall.getType(), aggCall.getName());
        assert refs.get(i) == null;
        if (n == 0) {
            refs.set(i, new RexInputRef(groupAndIndicatorCount + aggCallList.size(), newAggCall.getType()));
        } else {
            refs.set(i, new RexInputRef(leftFields.size() + groupAndIndicatorCount + aggCallList.size(), newAggCall.getType()));
        }
        aggCallList.add(newAggCall);
    }
    final Map<Integer, Integer> map = new HashMap<>();
    for (Integer key : aggregate.getGroupSet()) {
        map.put(key, map.size());
    }
    final ImmutableBitSet newGroupSet = aggregate.getGroupSet().permute(map);
    assert newGroupSet.equals(ImmutableBitSet.range(aggregate.getGroupSet().cardinality()));
    ImmutableList<ImmutableBitSet> newGroupingSets = null;
    if (aggregate.indicator) {
        newGroupingSets = ImmutableBitSet.ORDERING.immutableSortedCopy(ImmutableBitSet.permute(aggregate.getGroupSets(), map));
    }
    relBuilder.push(aggregate.copy(aggregate.getTraitSet(), relBuilder.build(), aggregate.indicator, newGroupSet, newGroupingSets, aggCallList));
    // If there's no left child yet, no need to create the join
    if (n == 0) {
        return;
    }
    // Create the join condition. It is of the form
    // 'left.f0 = right.f0 and left.f1 = right.f1 and ...'
    // where {f0, f1, ...} are the GROUP BY fields.
    final List<RelDataTypeField> distinctFields = relBuilder.peek().getRowType().getFieldList();
    final List<RexNode> conditions = Lists.newArrayList();
    for (i = 0; i < groupAndIndicatorCount; ++i) {
        // null values form its own group
        // use "is not distinct from" so that the join condition
        // allows null values to match.
        conditions.add(rexBuilder.makeCall(SqlStdOperatorTable.IS_NOT_DISTINCT_FROM, RexInputRef.of(i, leftFields), new RexInputRef(leftFields.size() + i, distinctFields.get(i).getType())));
    }
    // Join in the new 'select distinct' relation.
    relBuilder.join(JoinRelType.INNER, conditions);
}

Example 14 with AggregateCall

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

the class AggregateExpandDistinctAggregatesRule method rewriteAggCalls.

private static void rewriteAggCalls(List<AggregateCall> newAggCalls, List<Integer> argList, Map<Integer, Integer> sourceOf) {
    // "COUNT(DISTINCT e.sal)" becomes   "COUNT(distinct_e.sal)".
    for (int i = 0; i < newAggCalls.size(); i++) {
        final AggregateCall aggCall = newAggCalls.get(i);
        // COUNT(DISTINCT gender) or SUM(sal).
        if (!aggCall.isDistinct()) {
            continue;
        }
        if (!aggCall.getArgList().equals(argList)) {
            continue;
        }
        // Re-map arguments.
        final int argCount = aggCall.getArgList().size();
        final List<Integer> newArgs = new ArrayList<>(argCount);
        for (int j = 0; j < argCount; j++) {
            final Integer arg = aggCall.getArgList().get(j);
            newArgs.add(sourceOf.get(arg));
        }
        final AggregateCall newAggCall = AggregateCall.create(aggCall.getAggregation(), false, aggCall.isApproximate(), newArgs, -1, aggCall.getType(), aggCall.getName());
        newAggCalls.set(i, newAggCall);
    }
}

Example 15 with AggregateCall

use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rel.core.AggregateCall 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());
}