Decouple pipeline reductions from final agg reduction (#45796)

Historically only two things happened in the final reduction:
empty buckets were filled, and pipeline aggs were reduced (since it
was the final reduction, this was safe).  Usage of the final reduction
is growing however.  Auto-date-histo might need to perform
many reductions on final-reduce to merge down buckets, CCS
may need to side-step the final reduction if sending to a
different cluster, etc

Having pipelines generate their output in the final reduce was
convenient, but is becoming increasingly difficult to manage
as the rest of the agg framework advances.

This commit decouples pipeline aggs from the final reduction by
introducing a new "top level" reduce, which should be called
at the beginning of the reduce cycle (e.g. from the SearchPhaseController).
This will only reduce pipeline aggs on the final reduce after
the non-pipeline agg tree has been fully reduced.

By separating pipeline reduction into their own set of methods,
aggregations are free to use the final reduction for whatever
purpose without worrying about generating pipeline results
which are non-reducible

Author: polyfractal

modules/aggs-matrix-stats/src/main/java/org/elasticsearch/search/aggregations/matrix/stats/InternalMatrixStats.java

@@ -233,7 +233,7 @@ public Object getProperty(List<String> path) {
     }
 
     @Override
-    public InternalAggregation doReduce(List<InternalAggregation> aggregations, ReduceContext reduceContext) {
+    public InternalAggregation reduce(List<InternalAggregation> aggregations, ReduceContext reduceContext) {
         // merge stats across all shards
         List<InternalAggregation> aggs = new ArrayList<>(aggregations);
         aggs.removeIf(p -> ((InternalMatrixStats)p).stats == null);

server/src/main/java/org/elasticsearch/action/search/SearchPhaseController.java

@@ -487,7 +487,7 @@ private ReducedQueryPhase reducedQueryPhase(Collection<? extends SearchPhaseResu
         }
         ReduceContext reduceContext = reduceContextFunction.apply(performFinalReduce);
         final InternalAggregations aggregations = aggregationsList.isEmpty() ? null :
-            InternalAggregations.reduce(aggregationsList, reduceContext);
+            InternalAggregations.topLevelReduce(aggregationsList, reduceContext);
         final SearchProfileShardResults shardResults = profileResults.isEmpty() ? null : new SearchProfileShardResults(profileResults);
         final SortedTopDocs sortedTopDocs = sortDocs(isScrollRequest, queryResults, bufferedTopDocs, topDocsStats, from, size,
             reducedCompletionSuggestions);
@@ -625,7 +625,7 @@ private synchronized void consumeInternal(QuerySearchResult querySearchResult) {
             if (index == bufferSize) {
                 if (hasAggs) {
                     ReduceContext reduceContext = controller.reduceContextFunction.apply(false);
-                    InternalAggregations reducedAggs = InternalAggregations.reduce(Arrays.asList(aggsBuffer), reduceContext);
+                    InternalAggregations reducedAggs = InternalAggregations.topLevelReduce(Arrays.asList(aggsBuffer), reduceContext);
                     Arrays.fill(aggsBuffer, null);
                     aggsBuffer[0] = reducedAggs;
                 }

server/src/main/java/org/elasticsearch/action/search/SearchResponseMerger.java

@@ -196,7 +196,7 @@ SearchResponse getMergedResponse(Clusters clusters) {
         SearchHits mergedSearchHits = topDocsToSearchHits(topDocs, topDocsStats);
         setSuggestShardIndex(shards, groupedSuggestions);
         Suggest suggest = groupedSuggestions.isEmpty() ? null : new Suggest(Suggest.reduce(groupedSuggestions));
-        InternalAggregations reducedAggs = InternalAggregations.reduce(aggs, reduceContextFunction.apply(true));
+        InternalAggregations reducedAggs = InternalAggregations.topLevelReduce(aggs, reduceContextFunction.apply(true));
         ShardSearchFailure[] shardFailures = failures.toArray(ShardSearchFailure.EMPTY_ARRAY);
         SearchProfileShardResults profileShardResults = profileResults.isEmpty() ? null : new SearchProfileShardResults(profileResults);
         //make failures ordering consistent between ordinary search and CCS by looking at the shard they come from

server/src/main/java/org/elasticsearch/search/aggregations/InternalAggregation.java

@@ -126,23 +126,25 @@ public String getName() {
         return name;
     }
 
+    /**
+     * Creates the output from all pipeline aggs that this aggregation is associated with.  Should only
+     * be called after all aggregations have been fully reduced
+     */
+    public InternalAggregation reducePipelines(InternalAggregation reducedAggs, ReduceContext reduceContext) {
+        assert reduceContext.isFinalReduce();
+        for (PipelineAggregator pipelineAggregator : pipelineAggregators) {
+            reducedAggs = pipelineAggregator.reduce(reducedAggs, reduceContext);
+        }
+        return reducedAggs;
+    }
+
     /**
      * Reduces the given aggregations to a single one and returns it. In <b>most</b> cases, the assumption will be the all given
      * aggregations are of the same type (the same type as this aggregation). For best efficiency, when implementing,
      * try reusing an existing instance (typically the first in the given list) to save on redundant object
      * construction.
      */
-    public final InternalAggregation reduce(List<InternalAggregation> aggregations, ReduceContext reduceContext) {
-        InternalAggregation aggResult = doReduce(aggregations, reduceContext);
-        if (reduceContext.isFinalReduce()) {
-            for (PipelineAggregator pipelineAggregator : pipelineAggregators) {
-                aggResult = pipelineAggregator.reduce(aggResult, reduceContext);
-            }
-        }
-        return aggResult;
-    }
-
-    public abstract InternalAggregation doReduce(List<InternalAggregation> aggregations, ReduceContext reduceContext);
+    public abstract InternalAggregation reduce(List<InternalAggregation> aggregations, ReduceContext reduceContext);
 
     /**
      * Return true if this aggregation is mapped, and can lead a reduction.  If this agg returns

server/src/main/java/org/elasticsearch/search/aggregations/InternalAggregations.java

@@ -33,6 +33,7 @@
 import java.util.List;
 import java.util.Map;
 import java.util.Objects;
+import java.util.stream.Collectors;
 
 /**
  * An internal implementation of {@link Aggregations}.
@@ -91,10 +92,47 @@ public List<SiblingPipelineAggregator> getTopLevelPipelineAggregators() {
         return topLevelPipelineAggregators;
     }
 
+    @SuppressWarnings("unchecked")
+    private List<InternalAggregation> getInternalAggregations() {
+        return (List<InternalAggregation>) aggregations;
+    }
+
+    /**
+     * Begin the reduction process.  This should be the entry point for the "first" reduction, e.g. called by
+     * SearchPhaseController or anywhere else that wants to initiate a reduction.  It _should not_ be called
+     * as an intermediate reduction step (e.g. in the middle of an aggregation tree).
+     *
+     * This method first reduces the aggregations, and if it is the final reduce, then reduce the pipeline
+     * aggregations (both embedded parent/sibling as well as top-level sibling pipelines)
+     */
+    public static InternalAggregations topLevelReduce(List<InternalAggregations> aggregationsList, ReduceContext context) {
+        InternalAggregations reduced = reduce(aggregationsList, context);
+        if (reduced == null) {
+            return null;
+        }
+
+        if (context.isFinalReduce()) {
+            List<InternalAggregation> reducedInternalAggs = reduced.getInternalAggregations();
+            reducedInternalAggs = reducedInternalAggs.stream()
+                .map(agg -> agg.reducePipelines(agg, context))
+                .collect(Collectors.toList());
+
+            List<SiblingPipelineAggregator> topLevelPipelineAggregators = aggregationsList.get(0).getTopLevelPipelineAggregators();
+            for (SiblingPipelineAggregator pipelineAggregator : topLevelPipelineAggregators) {
+                InternalAggregation newAgg
+                    = pipelineAggregator.doReduce(new InternalAggregations(reducedInternalAggs), context);
+                reducedInternalAggs.add(newAgg);
+            }
+            return new InternalAggregations(reducedInternalAggs);
+        }
+        return reduced;
+    }
+
     /**
      * Reduces the given list of aggregations as well as the top-level pipeline aggregators extracted from the first
      * {@link InternalAggregations} object found in the list.
-     * Note that top-level pipeline aggregators are reduced only as part of the final reduction phase, otherwise they are left untouched.
+     * Note that pipeline aggregations _are not_ reduced by this method.  Pipelines are handled
+     * separately by {@link InternalAggregations#topLevelReduce(List, ReduceContext)}
      */
     public static InternalAggregations reduce(List<InternalAggregations> aggregationsList, ReduceContext context) {
         if (aggregationsList.isEmpty()) {
@@ -123,13 +161,6 @@ public static InternalAggregations reduce(List<InternalAggregations> aggregation
             reducedAggregations.add(first.reduce(aggregations, context));
         }
 
-        if (context.isFinalReduce()) {
-            for (SiblingPipelineAggregator pipelineAggregator : topLevelPipelineAggregators) {
-                InternalAggregation newAgg = pipelineAggregator.doReduce(new InternalAggregations(reducedAggregations), context);
-                reducedAggregations.add(newAgg);
-            }
-            return new InternalAggregations(reducedAggregations);
-        }
         return new InternalAggregations(reducedAggregations, topLevelPipelineAggregators);
     }
 }

server/src/main/java/org/elasticsearch/search/aggregations/InternalMultiBucketAggregation.java

@@ -26,6 +26,7 @@
 import org.elasticsearch.search.aggregations.pipeline.PipelineAggregator;
 
 import java.io.IOException;
+import java.util.ArrayList;
 import java.util.List;
 import java.util.Map;
 
@@ -73,7 +74,7 @@ protected InternalMultiBucketAggregation(StreamInput in) throws IOException {
     protected abstract B reduceBucket(List<B> buckets, ReduceContext context);
 
     @Override
-    public abstract List<? extends InternalBucket> getBuckets();
+    public abstract List<B> getBuckets();
 
     @Override
     public Object getProperty(List<String> path) {
@@ -141,6 +142,30 @@ public static int countInnerBucket(Aggregation agg) {
         return size;
     }
 
+    /**
+     * Unlike {@link InternalAggregation#reducePipelines(InternalAggregation, ReduceContext)}, a multi-bucket
+     * agg needs to first reduce the buckets (and their parent pipelines) before allowing sibling pipelines
+     * to materialize
+     */
+    @Override
+    public final InternalAggregation reducePipelines(InternalAggregation reducedAggs, ReduceContext reduceContext) {
+        assert reduceContext.isFinalReduce();
+        List<B> materializedBuckets = reducePipelineBuckets(reduceContext);
+        return super.reducePipelines(create(materializedBuckets), reduceContext);
+    }
+
+    private List<B> reducePipelineBuckets(ReduceContext reduceContext) {
+        List<B> reducedBuckets = new ArrayList<>();
+        for (B bucket : getBuckets()) {
+            List<InternalAggregation> aggs = new ArrayList<>();
+            for (Aggregation agg : bucket.getAggregations()) {
+                aggs.add(((InternalAggregation)agg).reducePipelines((InternalAggregation)agg, reduceContext));
+            }
+            reducedBuckets.add(createBucket(new InternalAggregations(aggs), bucket));
+        }
+        return reducedBuckets;
+    }
+
     public abstract static class InternalBucket implements Bucket, Writeable {
 
         public Object getProperty(String containingAggName, List<String> path) {

server/src/main/java/org/elasticsearch/search/aggregations/bucket/InternalSingleBucketAggregation.java

@@ -97,7 +97,7 @@ public InternalSingleBucketAggregation create(InternalAggregations subAggregatio
     protected abstract InternalSingleBucketAggregation newAggregation(String name, long docCount, InternalAggregations subAggregations);
 
     @Override
-    public InternalAggregation doReduce(List<InternalAggregation> aggregations, ReduceContext reduceContext) {
+    public InternalAggregation reduce(List<InternalAggregation> aggregations, ReduceContext reduceContext) {
         long docCount = 0L;
         List<InternalAggregations> subAggregationsList = new ArrayList<>(aggregations.size());
         for (InternalAggregation aggregation : aggregations) {

server/src/main/java/org/elasticsearch/search/aggregations/bucket/adjacency/InternalAdjacencyMatrix.java

@@ -181,7 +181,7 @@ public InternalBucket getBucketByKey(String key) {
     }
 
     @Override
-    public InternalAggregation doReduce(List<InternalAggregation> aggregations, ReduceContext reduceContext) {
+    public InternalAggregation reduce(List<InternalAggregation> aggregations, ReduceContext reduceContext) {
         Map<String, List<InternalBucket>> bucketsMap = new HashMap<>();
         for (InternalAggregation aggregation : aggregations) {
             InternalAdjacencyMatrix filters = (InternalAdjacencyMatrix) aggregation;

server/src/main/java/org/elasticsearch/search/aggregations/bucket/composite/InternalComposite.java

@@ -143,7 +143,7 @@ int[] getReverseMuls() {
     }
 
     @Override
-    public InternalAggregation doReduce(List<InternalAggregation> aggregations, ReduceContext reduceContext) {
+    public InternalAggregation reduce(List<InternalAggregation> aggregations, ReduceContext reduceContext) {
         PriorityQueue<BucketIterator> pq = new PriorityQueue<>(aggregations.size());
         for (InternalAggregation agg : aggregations) {
             InternalComposite sortedAgg = (InternalComposite) agg;

server/src/main/java/org/elasticsearch/search/aggregations/bucket/filter/InternalFilters.java

@@ -189,7 +189,7 @@ public InternalBucket getBucketByKey(String key) {
     }
 
     @Override
-    public InternalAggregation doReduce(List<InternalAggregation> aggregations, ReduceContext reduceContext) {
+    public InternalAggregation reduce(List<InternalAggregation> aggregations, ReduceContext reduceContext) {
         List<List<InternalBucket>> bucketsList = null;
         for (InternalAggregation aggregation : aggregations) {
             InternalFilters filters = (InternalFilters) aggregation;