Examples with ResultMergeQueryRunner org.apache.druid.query.ResultMergeQueryRunner used on opensource projects

Example 1 with ResultMergeQueryRunner

use of org.apache.druid.query.ResultMergeQueryRunner in project druid by druid-io.

the class GroupByStrategyV2 method mergeResults.

@Override
public Sequence<ResultRow> mergeResults(final QueryRunner<ResultRow> baseRunner, final GroupByQuery query, final ResponseContext responseContext) {
    // Merge streams using ResultMergeQueryRunner, then apply postaggregators, then apply limit (which may
    // involve materialization)
    final ResultMergeQueryRunner<ResultRow> mergingQueryRunner = new ResultMergeQueryRunner<>(baseRunner, this::createResultComparator, this::createMergeFn);
    // Set up downstream context.
    final ImmutableMap.Builder<String, Object> context = ImmutableMap.builder();
    context.put("finalize", false);
    context.put(GroupByQueryConfig.CTX_KEY_STRATEGY, GroupByStrategySelector.STRATEGY_V2);
    context.put(CTX_KEY_OUTERMOST, false);
    Granularity granularity = query.getGranularity();
    List<DimensionSpec> dimensionSpecs = query.getDimensions();
    // the CTX_TIMESTAMP_RESULT_FIELD is set in DruidQuery.java
    final String timestampResultField = query.getContextValue(GroupByQuery.CTX_TIMESTAMP_RESULT_FIELD);
    final boolean hasTimestampResultField = (timestampResultField != null && !timestampResultField.isEmpty()) && query.getContextBoolean(CTX_KEY_OUTERMOST, true) && !query.isApplyLimitPushDown();
    int timestampResultFieldIndex = 0;
    if (hasTimestampResultField) {
        // sql like "group by city_id,time_floor(__time to day)",
        // the original translated query is granularity=all and dimensions:[d0, d1]
        // the better plan is granularity=day and dimensions:[d0]
        // but the ResultRow structure is changed from [d0, d1] to [__time, d0]
        // this structure should be fixed as [d0, d1] (actually it is [d0, __time]) before postAggs are called.
        // 
        // the above is the general idea of this optimization.
        // but from coding perspective, the granularity=all and "d0" dimension are referenced by many places,
        // eg: subtotals, having, grouping set, post agg,
        // there would be many many places need to be fixed if "d0" dimension is removed from query.dimensions
        // and the same to the granularity change.
        // so from easier coding perspective, this optimization is coded as groupby engine-level inner process change.
        // the most part of codes are in GroupByStrategyV2 about the process change between broker and compute node.
        // the basic logic like nested queries and subtotals are kept unchanged,
        // they will still see the granularity=all and the "d0" dimension.
        // 
        // the tradeoff is that GroupByStrategyV2 behaviors differently according to the query contexts set in DruidQuery
        // in another word,
        // the query generated by "explain plan for select ..." doesn't match to the native query ACTUALLY being executed,
        // the granularity and dimensions are slightly different.
        // now, part of the query plan logic is handled in GroupByStrategyV2, not only in DruidQuery.toGroupByQuery()
        final Granularity timestampResultFieldGranularity = query.getContextValue(GroupByQuery.CTX_TIMESTAMP_RESULT_FIELD_GRANULARITY);
        dimensionSpecs = query.getDimensions().stream().filter(dimensionSpec -> !dimensionSpec.getOutputName().equals(timestampResultField)).collect(Collectors.toList());
        granularity = timestampResultFieldGranularity;
        // when timestampResultField is the last dimension, should set sortByDimsFirst=true,
        // otherwise the downstream is sorted by row's timestamp first which makes the final ordering not as expected
        timestampResultFieldIndex = query.getContextValue(GroupByQuery.CTX_TIMESTAMP_RESULT_FIELD_INDEX);
        if (!query.getContextSortByDimsFirst() && timestampResultFieldIndex == query.getDimensions().size() - 1) {
            context.put(GroupByQuery.CTX_KEY_SORT_BY_DIMS_FIRST, true);
        }
        // it is actually equals to sortByDimsFirst=false
        if (query.getContextSortByDimsFirst() && timestampResultFieldIndex == 0) {
            context.put(GroupByQuery.CTX_KEY_SORT_BY_DIMS_FIRST, false);
        }
    // when hasTimestampResultField=true and timestampResultField is neither first nor last dimension,
    // the DefaultLimitSpec will always do the reordering
    }
    final int timestampResultFieldIndexInOriginalDimensions = timestampResultFieldIndex;
    if (query.getUniversalTimestamp() != null && !hasTimestampResultField) {
        // universalTimestamp works only when granularity is all
        // hasTimestampResultField works only when granularity is all
        // fudgeTimestamp should not be used when hasTimestampResultField=true due to the row's actual timestamp is used
        context.put(CTX_KEY_FUDGE_TIMESTAMP, String.valueOf(query.getUniversalTimestamp().getMillis()));
    }
    // The having spec shouldn't be passed down, so we need to convey the existing limit push down status
    context.put(GroupByQueryConfig.CTX_KEY_APPLY_LIMIT_PUSH_DOWN, query.isApplyLimitPushDown());
    // Always request array result rows when passing the query downstream.
    context.put(GroupByQueryConfig.CTX_KEY_ARRAY_RESULT_ROWS, true);
    final GroupByQuery newQuery = new GroupByQuery(query.getDataSource(), query.getQuerySegmentSpec(), query.getVirtualColumns(), query.getDimFilter(), granularity, dimensionSpecs, query.getAggregatorSpecs(), // Don't apply postaggregators on compute nodes
    ImmutableList.of(), // Don't do "having" clause until the end of this method.
    null, // higher-up).
    query.isApplyLimitPushDown() ? ((DefaultLimitSpec) query.getLimitSpec()).withOffsetToLimit() : null, query.getSubtotalsSpec(), query.getContext()).withOverriddenContext(context.build());
    final Sequence<ResultRow> mergedResults = mergingQueryRunner.run(QueryPlus.wrap(newQuery), responseContext);
    if (!query.getContextBoolean(CTX_KEY_OUTERMOST, true) || query.getContextBoolean(GroupByQueryConfig.CTX_KEY_EXECUTING_NESTED_QUERY, false)) {
        return mergedResults;
    } else if (query.getPostAggregatorSpecs().isEmpty()) {
        if (!hasTimestampResultField) {
            return mergedResults;
        }
        return Sequences.map(mergedResults, row -> {
            final ResultRow resultRow = ResultRow.create(query.getResultRowSizeWithoutPostAggregators());
            moveOrReplicateTimestampInRow(query, timestampResultFieldIndexInOriginalDimensions, row, resultRow);
            return resultRow;
        });
    } else {
        return Sequences.map(mergedResults, row -> {
            // This function's purpose is to apply PostAggregators.
            final ResultRow rowWithPostAggregations = ResultRow.create(query.getResultRowSizeWithPostAggregators());
            // Copy everything that comes before the postaggregations.
            if (hasTimestampResultField) {
                moveOrReplicateTimestampInRow(query, timestampResultFieldIndexInOriginalDimensions, row, rowWithPostAggregations);
            } else {
                for (int i = 0; i < query.getResultRowPostAggregatorStart(); i++) {
                    rowWithPostAggregations.set(i, row.get(i));
                }
            }
            // Compute postaggregations. We need to do this with a result-row map because PostAggregator.compute
            // expects a map. Some further design adjustment may eliminate the need for it, and speed up this function.
            final Map<String, Object> mapForPostAggregationComputation = rowWithPostAggregations.toMap(query);
            for (int i = 0; i < query.getPostAggregatorSpecs().size(); i++) {
                final PostAggregator postAggregator = query.getPostAggregatorSpecs().get(i);
                final Object value = postAggregator.compute(mapForPostAggregationComputation);
                rowWithPostAggregations.set(query.getResultRowPostAggregatorStart() + i, value);
                mapForPostAggregationComputation.put(postAggregator.getName(), value);
            }
            return rowWithPostAggregations;
        });
    }
}