Examples with JoinStatistics io.trino.spi.connector.JoinStatistics used on opensource projects

Example 1 with JoinStatistics

use of io.trino.spi.connector.JoinStatistics in project trino by trinodb.

the class PushJoinIntoTableScan method getJoinStatistics.

private JoinStatistics getJoinStatistics(JoinNode join, TableScanNode left, TableScanNode right, Context context) {
    return new JoinStatistics() {

        @Override
        public Optional<BasicRelationStatistics> getLeftStatistics() {
            return getBasicRelationStats(left, left.getOutputSymbols(), context);
        }

        @Override
        public Optional<BasicRelationStatistics> getRightStatistics() {
            return getBasicRelationStats(right, right.getOutputSymbols(), context);
        }

        @Override
        public Optional<BasicRelationStatistics> getJoinStatistics() {
            return getBasicRelationStats(join, join.getOutputSymbols(), context);
        }

        private Optional<BasicRelationStatistics> getBasicRelationStats(PlanNode node, List<Symbol> outputSymbols, Context context) {
            PlanNodeStatsEstimate stats = context.getStatsProvider().getStats(node);
            TypeProvider types = context.getSymbolAllocator().getTypes();
            double outputRowCount = stats.getOutputRowCount();
            double outputSize = stats.getOutputSizeInBytes(outputSymbols, types);
            if (isNaN(outputRowCount) || isNaN(outputSize)) {
                return Optional.empty();
            }
            return Optional.of(new BasicRelationStatistics((long) outputRowCount, (long) outputSize));
        }
    };
}

Example 2 with JoinStatistics

use of io.trino.spi.connector.JoinStatistics in project trino by trinodb.

the class PushJoinIntoTableScan method apply.

@Override
public Result apply(JoinNode joinNode, Captures captures, Context context) {
    if (joinNode.isCrossJoin()) {
        return Result.empty();
    }
    TableScanNode left = captures.get(LEFT_TABLE_SCAN);
    TableScanNode right = captures.get(RIGHT_TABLE_SCAN);
    verify(!left.isUpdateTarget() && !right.isUpdateTarget(), "Unexpected Join over for-update table scan");
    Expression effectiveFilter = getEffectiveFilter(joinNode);
    FilterSplitResult filterSplitResult = splitFilter(effectiveFilter, left.getOutputSymbols(), right.getOutputSymbols(), context);
    if (!filterSplitResult.getRemainingFilter().equals(BooleanLiteral.TRUE_LITERAL)) {
        // TODO add extra filter node above join
        return Result.empty();
    }
    if (left.getEnforcedConstraint().isNone() || right.getEnforcedConstraint().isNone()) {
        // enforced constraint harder below.
        return Result.empty();
    }
    Map<String, ColumnHandle> leftAssignments = left.getAssignments().entrySet().stream().collect(toImmutableMap(entry -> entry.getKey().getName(), Map.Entry::getValue));
    Map<String, ColumnHandle> rightAssignments = right.getAssignments().entrySet().stream().collect(toImmutableMap(entry -> entry.getKey().getName(), Map.Entry::getValue));
    /*
         * We are (lazily) computing estimated statistics for join node and left and right table
         * and passing those to connector via applyJoin.
         *
         * There are a couple reasons for this approach:
         * - the engine knows how to estimate join and connector may not
         * - the engine may have cached stats for the table scans (within context.getStatsProvider()), so can be able to provide information more inexpensively
         * - in the future, the engine may be able to provide stats for table scan even in case when connector no longer can (see https://github.com/trinodb/trino/issues/6998)
         * - the pushdown feasibility assessment logic may be different (or configured differently) for different connectors/catalogs.
         */
    JoinStatistics joinStatistics = getJoinStatistics(joinNode, left, right, context);
    Optional<JoinApplicationResult<TableHandle>> joinApplicationResult = metadata.applyJoin(context.getSession(), getJoinType(joinNode), left.getTable(), right.getTable(), filterSplitResult.getPushableConditions(), // TODO we could pass only subset of assignments here, those which are needed to resolve filterSplitResult.getPushableConditions
    leftAssignments, rightAssignments, joinStatistics);
    if (joinApplicationResult.isEmpty()) {
        return Result.empty();
    }
    TableHandle handle = joinApplicationResult.get().getTableHandle();
    Map<ColumnHandle, ColumnHandle> leftColumnHandlesMapping = joinApplicationResult.get().getLeftColumnHandles();
    Map<ColumnHandle, ColumnHandle> rightColumnHandlesMapping = joinApplicationResult.get().getRightColumnHandles();
    ImmutableMap.Builder<Symbol, ColumnHandle> assignmentsBuilder = ImmutableMap.builder();
    assignmentsBuilder.putAll(left.getAssignments().entrySet().stream().collect(toImmutableMap(Map.Entry::getKey, entry -> leftColumnHandlesMapping.get(entry.getValue()))));
    assignmentsBuilder.putAll(right.getAssignments().entrySet().stream().collect(toImmutableMap(Map.Entry::getKey, entry -> rightColumnHandlesMapping.get(entry.getValue()))));
    Map<Symbol, ColumnHandle> assignments = assignmentsBuilder.buildOrThrow();
    // convert enforced constraint
    JoinNode.Type joinType = joinNode.getType();
    TupleDomain<ColumnHandle> leftConstraint = deriveConstraint(left.getEnforcedConstraint(), leftColumnHandlesMapping, joinType == RIGHT || joinType == FULL);
    TupleDomain<ColumnHandle> rightConstraint = deriveConstraint(right.getEnforcedConstraint(), rightColumnHandlesMapping, joinType == LEFT || joinType == FULL);
    TupleDomain<ColumnHandle> newEnforcedConstraint = TupleDomain.withColumnDomains(ImmutableMap.<ColumnHandle, Domain>builder().putAll(leftConstraint.getDomains().orElseThrow()).putAll(rightConstraint.getDomains().orElseThrow()).buildOrThrow());
    return Result.ofPlanNode(new ProjectNode(context.getIdAllocator().getNextId(), new TableScanNode(joinNode.getId(), handle, ImmutableList.copyOf(assignments.keySet()), assignments, newEnforcedConstraint, deriveTableStatisticsForPushdown(context.getStatsProvider(), context.getSession(), joinApplicationResult.get().isPrecalculateStatistics(), joinNode), false, Optional.empty()), Assignments.identity(joinNode.getOutputSymbols())));
}