Release aggregations earlier during reduce (elastic#124520)

Release each hit's aggregations before moving on to the next hit and unlink it from the shard result even earlier.
Also, do the aggregation-reduce earlier in the reduce steps to reduce average heap use over time.
To that effect, do not do the reduction in the search phase controller. This has the added benefit of removing any need for a fake aggs-reduce-context in scroll.

Author: original-brownbear

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

@@ -15,9 +15,12 @@
 import org.elasticsearch.action.search.SearchPhaseController.TopDocsStats;
 import org.elasticsearch.common.breaker.CircuitBreaker;
 import org.elasticsearch.common.breaker.CircuitBreakingException;
+import org.elasticsearch.common.collect.Iterators;
 import org.elasticsearch.common.io.stream.DelayableWriteable;
 import org.elasticsearch.common.lucene.search.TopDocsAndMaxScore;
 import org.elasticsearch.common.util.concurrent.AbstractRunnable;
+import org.elasticsearch.core.Releasable;
+import org.elasticsearch.core.Releasables;
 import org.elasticsearch.search.SearchPhaseResult;
 import org.elasticsearch.search.SearchService;
 import org.elasticsearch.search.SearchShardTarget;
@@ -31,6 +34,7 @@
 import java.util.ArrayList;
 import java.util.Collections;
 import java.util.Comparator;
+import java.util.Iterator;
 import java.util.List;
 import java.util.concurrent.Executor;
 import java.util.concurrent.atomic.AtomicReference;
@@ -174,14 +178,10 @@ public SearchPhaseController.ReducedQueryPhase reduce() throws Exception {
         this.mergeResult = null;
         final int resultSize = buffer.size() + (mergeResult == null ? 0 : 1);
         final List<TopDocs> topDocsList = hasTopDocs ? new ArrayList<>(resultSize) : null;
-        final List<DelayableWriteable<InternalAggregations>> aggsList = hasAggs ? new ArrayList<>(resultSize) : null;
         if (mergeResult != null) {
             if (topDocsList != null) {
                 topDocsList.add(mergeResult.reducedTopDocs);
             }
-            if (aggsList != null) {
-                aggsList.add(DelayableWriteable.referencing(mergeResult.reducedAggs));
-            }
         }
         for (QuerySearchResult result : buffer) {
             topDocsStats.add(result.topDocs(), result.searchTimedOut(), result.terminatedEarly());
@@ -190,34 +190,39 @@ public SearchPhaseController.ReducedQueryPhase reduce() throws Exception {
                 setShardIndex(topDocs.topDocs, result.getShardIndex());
                 topDocsList.add(topDocs.topDocs);
             }
-            if (aggsList != null) {
-                aggsList.add(result.getAggs());
-            }
         }
         SearchPhaseController.ReducedQueryPhase reducePhase;
         long breakerSize = circuitBreakerBytes;
+        final InternalAggregations aggs;
         try {
-            if (aggsList != null) {
+            if (hasAggs) {
                 // Add an estimate of the final reduce size
                 breakerSize = addEstimateAndMaybeBreak(estimateRamBytesUsedForReduce(breakerSize));
+                aggs = aggregate(
+                    buffer.iterator(),
+                    mergeResult,
+                    resultSize,
+                    performFinalReduce ? aggReduceContextBuilder.forFinalReduction() : aggReduceContextBuilder.forPartialReduction()
+                );
+            } else {
+                aggs = null;
             }
             reducePhase = SearchPhaseController.reducedQueryPhase(
                 results.asList(),
-                aggsList,
+                aggs,
                 topDocsList == null ? Collections.emptyList() : topDocsList,
                 topDocsStats,
                 numReducePhases,
                 false,
-                aggReduceContextBuilder,
-                queryPhaseRankCoordinatorContext,
-                performFinalReduce
+                queryPhaseRankCoordinatorContext
             );
+            buffer = null;
         } finally {
             releaseAggs(buffer);
         }
         if (hasAggs
             // reduced aggregations can be null if all shards failed
-            && reducePhase.aggregations() != null) {
+            && aggs != null) {
 
             // Update the circuit breaker to replace the estimation with the serialized size of the newly reduced result
             long finalSize = DelayableWriteable.getSerializedSize(reducePhase.aggregations()) - breakerSize;
@@ -249,17 +254,7 @@ private MergeResult partialReduce(
         toConsume.sort(RESULT_COMPARATOR);
 
         final TopDocs newTopDocs;
-        final InternalAggregations newAggs;
-        final List<DelayableWriteable<InternalAggregations>> aggsList;
         final int resultSetSize = toConsume.size() + (lastMerge != null ? 1 : 0);
-        if (hasAggs) {
-            aggsList = new ArrayList<>(resultSetSize);
-            if (lastMerge != null) {
-                aggsList.add(DelayableWriteable.referencing(lastMerge.reducedAggs));
-            }
-        } else {
-            aggsList = null;
-        }
         List<TopDocs> topDocsList;
         if (hasTopDocs) {
             topDocsList = new ArrayList<>(resultSetSize);
@@ -269,14 +264,12 @@ private MergeResult partialReduce(
         } else {
             topDocsList = null;
         }
+        final InternalAggregations newAggs;
         try {
             for (QuerySearchResult result : toConsume) {
                 topDocsStats.add(result.topDocs(), result.searchTimedOut(), result.terminatedEarly());
                 SearchShardTarget target = result.getSearchShardTarget();
                 processedShards.add(new SearchShard(target.getClusterAlias(), target.getShardId()));
-                if (aggsList != null) {
-                    aggsList.add(result.getAggs());
-                }
                 if (topDocsList != null) {
                     TopDocsAndMaxScore topDocs = result.consumeTopDocs();
                     setShardIndex(topDocs.topDocs, result.getShardIndex());
@@ -285,9 +278,10 @@ private MergeResult partialReduce(
             }
             // we have to merge here in the same way we collect on a shard
             newTopDocs = topDocsList == null ? null : mergeTopDocs(topDocsList, topNSize, 0);
-            newAggs = aggsList == null
-                ? null
-                : InternalAggregations.topLevelReduceDelayable(aggsList, aggReduceContextBuilder.forPartialReduction());
+            newAggs = hasAggs
+                ? aggregate(toConsume.iterator(), lastMerge, resultSetSize, aggReduceContextBuilder.forPartialReduction())
+                : null;
+            toConsume = null;
         } finally {
             releaseAggs(toConsume);
         }
@@ -302,6 +296,45 @@ private MergeResult partialReduce(
         return new MergeResult(processedShards, newTopDocs, newAggs, newAggs != null ? DelayableWriteable.getSerializedSize(newAggs) : 0);
     }
 
+    private static InternalAggregations aggregate(
+        Iterator<QuerySearchResult> toConsume,
+        MergeResult lastMerge,
+        int resultSetSize,
+        AggregationReduceContext reduceContext
+    ) {
+        interface ReleasableIterator extends Iterator<InternalAggregations>, Releasable {}
+        try (var aggsIter = new ReleasableIterator() {
+
+            private Releasable toRelease;
+
+            @Override
+            public void close() {
+                Releasables.close(toRelease);
+            }
+
+            @Override
+            public boolean hasNext() {
+                return toConsume.hasNext();
+            }
+
+            @Override
+            public InternalAggregations next() {
+                var res = toConsume.next().consumeAggs();
+                Releasables.close(toRelease);
+                toRelease = res;
+                return res.expand();
+            }
+        }) {
+            return InternalAggregations.topLevelReduce(
+                lastMerge == null ? aggsIter : Iterators.concat(Iterators.single(lastMerge.reducedAggs), aggsIter),
+                resultSetSize,
+                reduceContext
+            );
+        } finally {
+            toConsume.forEachRemaining(QuerySearchResult::releaseAggs);
+        }
+    }
+
     public int getNumReducePhases() {
         return numReducePhases;
     }
@@ -517,8 +550,10 @@ public void onFailure(Exception exc) {
     }
 
     private static void releaseAggs(List<QuerySearchResult> toConsume) {
-        for (QuerySearchResult result : toConsume) {
-            result.releaseAggs();
+        if (toConsume != null) {
+            for (QuerySearchResult result : toConsume) {
+                result.releaseAggs();
+            }
         }
     }
 

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

@@ -20,11 +20,11 @@
 import org.apache.lucene.search.TotalHits;
 import org.apache.lucene.search.TotalHits.Relation;
 import org.elasticsearch.common.breaker.CircuitBreaker;
-import org.elasticsearch.common.io.stream.DelayableWriteable;
 import org.elasticsearch.common.lucene.Lucene;
 import org.elasticsearch.common.lucene.search.TopDocsAndMaxScore;
 import org.elasticsearch.common.util.Maps;
 import org.elasticsearch.common.util.concurrent.AtomicArray;
+import org.elasticsearch.core.Nullable;
 import org.elasticsearch.index.fielddata.IndexFieldData;
 import org.elasticsearch.lucene.grouping.TopFieldGroups;
 import org.elasticsearch.search.DocValueFormat;
@@ -402,22 +402,20 @@ private static SearchHits getHits(
     /**
      * Reduces the given query results and consumes all aggregations and profile results.
      * @param queryResults a list of non-null query shard results
-     * @param bufferedAggs a list of pre-collected aggregations.
+     * @param reducedAggs already reduced aggregations
      * @param bufferedTopDocs a list of pre-collected top docs.
      * @param numReducePhases the number of non-final reduce phases applied to the query results.
      * @see QuerySearchResult#getAggs()
      * @see QuerySearchResult#consumeProfileResult()
      */
     static ReducedQueryPhase reducedQueryPhase(
         Collection<? extends SearchPhaseResult> queryResults,
-        List<DelayableWriteable<InternalAggregations>> bufferedAggs,
+        @Nullable InternalAggregations reducedAggs,
         List<TopDocs> bufferedTopDocs,
         TopDocsStats topDocsStats,
         int numReducePhases,
         boolean isScrollRequest,
-        AggregationReduceContext.Builder aggReduceContextBuilder,
-        QueryPhaseRankCoordinatorContext queryPhaseRankCoordinatorContext,
-        boolean performFinalReduce
+        QueryPhaseRankCoordinatorContext queryPhaseRankCoordinatorContext
     ) {
         assert numReducePhases >= 0 : "num reduce phases must be >= 0 but was: " + numReducePhases;
         numReducePhases++; // increment for this phase
@@ -521,12 +519,7 @@ static ReducedQueryPhase reducedQueryPhase(
             topDocsStats.timedOut,
             topDocsStats.terminatedEarly,
             reducedSuggest,
-            bufferedAggs == null
-                ? null
-                : InternalAggregations.topLevelReduceDelayable(
-                    bufferedAggs,
-                    performFinalReduce ? aggReduceContextBuilder.forFinalReduction() : aggReduceContextBuilder.forPartialReduction()
-                ),
+            reducedAggs,
             profileShardResults.isEmpty() ? null : new SearchProfileResultsBuilder(profileShardResults),
             sortedTopDocs,
             sortValueFormats,

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

@@ -20,7 +20,6 @@
 import org.elasticsearch.core.Nullable;
 import org.elasticsearch.search.SearchPhaseResult;
 import org.elasticsearch.search.SearchShardTarget;
-import org.elasticsearch.search.aggregations.AggregationReduceContext;
 import org.elasticsearch.search.internal.InternalScrollSearchRequest;
 import org.elasticsearch.search.internal.SearchContext;
 import org.elasticsearch.search.internal.ShardSearchContextId;
@@ -313,17 +312,6 @@ protected Transport.Connection getConnection(String clusterAlias, DiscoveryNode
      * @param queryResults a list of non-null query shard results
      */
     protected static SearchPhaseController.ReducedQueryPhase reducedScrollQueryPhase(Collection<? extends SearchPhaseResult> queryResults) {
-        AggregationReduceContext.Builder aggReduceContextBuilder = new AggregationReduceContext.Builder() {
-            @Override
-            public AggregationReduceContext forPartialReduction() {
-                throw new UnsupportedOperationException("Scroll requests don't have aggs");
-            }
-
-            @Override
-            public AggregationReduceContext forFinalReduction() {
-                throw new UnsupportedOperationException("Scroll requests don't have aggs");
-            }
-        };
         final SearchPhaseController.TopDocsStats topDocsStats = new SearchPhaseController.TopDocsStats(
             SearchContext.TRACK_TOTAL_HITS_ACCURATE
         );
@@ -339,16 +327,6 @@ public AggregationReduceContext forFinalReduction() {
                 topDocs.add(td.topDocs);
             }
         }
-        return SearchPhaseController.reducedQueryPhase(
-            queryResults,
-            null,
-            topDocs,
-            topDocsStats,
-            0,
-            true,
-            aggReduceContextBuilder,
-            null,
-            true
-        );
+        return SearchPhaseController.reducedQueryPhase(queryResults, null, topDocs, topDocsStats, 0, true, null);
     }
 }

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

@@ -10,7 +10,6 @@
 
 import org.apache.lucene.util.SetOnce;
 import org.elasticsearch.common.ParsingException;
-import org.elasticsearch.common.io.stream.DelayableWriteable;
 import org.elasticsearch.common.io.stream.StreamInput;
 import org.elasticsearch.common.io.stream.StreamOutput;
 import org.elasticsearch.common.io.stream.Writeable;
@@ -26,7 +25,6 @@
 import org.elasticsearch.xcontent.XContentParser;
 
 import java.io.IOException;
-import java.util.AbstractList;
 import java.util.ArrayList;
 import java.util.Iterator;
 import java.util.List;
@@ -206,44 +204,22 @@ public SortValue sortValue(AggregationPath.PathElement head, Iterator<Aggregatio
     }
 
     /**
-     * Equivalent to {@link #topLevelReduce(List, AggregationReduceContext)} but it takes a list of
-     * {@link DelayableWriteable}. The object will be expanded once via {@link DelayableWriteable#expand()}
-     * but it is the responsibility of the caller to release those releasables.
+     * Equivalent to {@link #topLevelReduce(List, AggregationReduceContext)} but it takes an iterator and a count.
      */
-    public static InternalAggregations topLevelReduceDelayable(
-        List<DelayableWriteable<InternalAggregations>> delayableAggregations,
-        AggregationReduceContext context
-    ) {
-        final List<InternalAggregations> aggregations = new AbstractList<>() {
-            @Override
-            public InternalAggregations get(int index) {
-                return delayableAggregations.get(index).expand();
-            }
-
-            @Override
-            public int size() {
-                return delayableAggregations.size();
-            }
-        };
-        return topLevelReduce(aggregations, context);
+    public static InternalAggregations topLevelReduce(Iterator<InternalAggregations> aggs, int count, AggregationReduceContext context) {
+        if (count == 0) {
+            return null;
+        }
+        return maybeExecuteFinalReduce(context, count == 1 ? reduce(aggs.next(), context) : reduce(aggs, count, context));
     }
 
-    /**
-     * 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, AggregationReduceContext context) {
-        InternalAggregations reduced = reduce(aggregationsList, context);
+    private static InternalAggregations maybeExecuteFinalReduce(AggregationReduceContext context, InternalAggregations reduced) {
         if (reduced == null) {
             return null;
         }
         if (context.isFinalReduce()) {
-            List<InternalAggregation> reducedInternalAggs = reduced.getInternalAggregations();
-            reducedInternalAggs = reducedInternalAggs.stream()
+            List<InternalAggregation> reducedInternalAggs = reduced.getInternalAggregations()
+                .stream()
                 .map(agg -> agg.reducePipelines(agg, context, context.pipelineTreeRoot().subTree(agg.getName())))
                 .collect(Collectors.toCollection(ArrayList::new));
 
@@ -257,6 +233,18 @@ public static InternalAggregations topLevelReduce(List<InternalAggregations> agg
         return reduced;
     }
 
+    /**
+     * 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, AggregationReduceContext context) {
+        return maybeExecuteFinalReduce(context, reduce(aggregationsList, context));
+    }
+
     /**
      * 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.
@@ -280,6 +268,16 @@ public static InternalAggregations reduce(List<InternalAggregations> aggregation
         }
     }
 
+    private static InternalAggregations reduce(Iterator<InternalAggregations> aggsIterator, int count, AggregationReduceContext context) {
+        // general case
+        var first = aggsIterator.next();
+        try (AggregatorsReducer reducer = new AggregatorsReducer(first, context, count)) {
+            reducer.accept(first);
+            aggsIterator.forEachRemaining(reducer::accept);
+            return reducer.get();
+        }
+    }
+
     public static InternalAggregations reduce(InternalAggregations aggregations, AggregationReduceContext context) {
         final List<InternalAggregation> internalAggregations = aggregations.asList();
         int size = internalAggregations.size();