Do not evaluate shard_size and shard_min_doc_count at segment slice l…

…evel (#9085) (#9211)

* Use BucketCountThresholds in InternalTerms and InternalAggregations and do not apply shard level thresholds at slice level for Concurrent Segment Search



* Addressing comments



* Re-introduce shardSize member to InternalMultiTerms and InternalMappedTerms



* Introduce LocalBucketCountThresholds for local size and min_doc_count values



---------

Signed-off-by: Jay Deng <jayd0104@gmail.com>

CHANGELOG.md

@@ -55,6 +55,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 - Make MultiBucketConsumerService thread safe to use across slices during search ([#9047](https://github.com/opensearch-project/OpenSearch/pull/9047))
 - Removed blocking wait in TransportGetSnapshotsAction which was exhausting generic threadpool ([#8377](https://github.com/opensearch-project/OpenSearch/pull/8377))
 - Adds support for tracing runnable scenarios ([#8831](https://github.com/opensearch-project/OpenSearch/pull/8831))
+- Change shard_size and shard_min_doc_count evaluation to happen in shard level reduce phase ([#9085](https://github.com/opensearch-project/OpenSearch/pull/9085))
 
 ### Deprecated
 

benchmarks/src/main/java/org/opensearch/benchmark/search/aggregations/TermsReduceBenchmark.java

@@ -57,6 +57,7 @@
 import org.opensearch.search.aggregations.InternalAggregations;
 import org.opensearch.search.aggregations.MultiBucketConsumerService;
 import org.opensearch.search.aggregations.bucket.terms.StringTerms;
+import org.opensearch.search.aggregations.bucket.terms.TermsAggregator;
 import org.opensearch.search.aggregations.pipeline.PipelineAggregator;
 import org.opensearch.search.builder.SearchSourceBuilder;
 import org.opensearch.search.query.QuerySearchResult;
@@ -170,15 +171,14 @@ private StringTerms newTerms(Random rand, BytesRef[] dict, boolean withNested) {
                 "terms",
                 BucketOrder.key(true),
                 BucketOrder.count(false),
-                topNSize,
-                1,
                 Collections.emptyMap(),
                 DocValueFormat.RAW,
                 numShards,
                 true,
                 0,
                 buckets,
-                0
+                0,
+                new TermsAggregator.BucketCountThresholds(1, 0, topNSize, numShards)
             );
         }
 

benchmarks/src/main/java/org/opensearch/benchmark/search/aggregations/bucket/terms/StringTermsSerializationBenchmark.java

@@ -51,6 +51,7 @@
 import org.openjdk.jmh.annotations.Setup;
 import org.openjdk.jmh.annotations.State;
 import org.openjdk.jmh.annotations.Warmup;
+import org.opensearch.search.aggregations.bucket.terms.TermsAggregator;
 
 import java.util.ArrayList;
 import java.util.List;
@@ -86,15 +87,14 @@ private StringTerms newTerms(boolean withNested) {
             "test",
             BucketOrder.key(true),
             BucketOrder.key(true),
-            buckets,
-            1,
             null,
             DocValueFormat.RAW,
             buckets,
             false,
             100000,
             resultBuckets,
-            0
+            0,
+            new TermsAggregator.BucketCountThresholds(1, 0, buckets, buckets)
         );
     }
 

server/src/main/java/org/opensearch/search/aggregations/AggregationCollectorManager.java

@@ -26,7 +26,7 @@
  * aggregation operators
  */
 class AggregationCollectorManager implements CollectorManager<Collector, ReduceableSearchResult> {
-    private final SearchContext context;
+    protected final SearchContext context;
     private final CheckedFunction<SearchContext, List<Aggregator>, IOException> aggProvider;
     private final String collectorReason;
 
@@ -68,18 +68,11 @@ public ReduceableSearchResult reduce(Collection<Collector> collectors) throws IO
             internals,
             context.request().source().aggregations()::buildPipelineTree
         );
-        // Reduce the aggregations across slices before sending to the coordinator. We will perform shard level reduce iff multiple slices
-        // were created to execute this request and it used concurrent segment search path
-        // TODO: Add the check for flag that the request was executed using concurrent search
-        if (collectors.size() > 1) {
-            // using topLevelReduce here as PipelineTreeSource needs to be sent to coordinator in older release of OpenSearch. The actual
-            // evaluation of pipeline aggregation though happens on the coordinator during final reduction phase
-            return new AggregationReduceableSearchResult(
-                InternalAggregations.topLevelReduce(Collections.singletonList(internalAggregations), context.partialOnShard())
-            );
-        } else {
-            return new AggregationReduceableSearchResult(internalAggregations);
-        }
+        return buildAggregationResult(internalAggregations);
+    }
+
+    protected AggregationReduceableSearchResult buildAggregationResult(InternalAggregations internalAggregations) {
+        return new AggregationReduceableSearchResult(internalAggregations);
     }
 
     static Collector createCollector(SearchContext context, List<Aggregator> collectors, String reason) throws IOException {

server/src/main/java/org/opensearch/search/aggregations/GlobalAggCollectorManager.java

@@ -14,6 +14,7 @@
 import org.opensearch.search.profile.query.CollectorResult;
 
 import java.io.IOException;
+import java.util.Collections;
 import java.util.Objects;
 
 /**
@@ -38,4 +39,15 @@ public Collector newCollector() throws IOException {
             return super.newCollector();
         }
     }
+
+    @Override
+    protected AggregationReduceableSearchResult buildAggregationResult(InternalAggregations internalAggregations) {
+        // Reduce the aggregations across slices before sending to the coordinator. We will perform shard level reduce as long as any slices
+        // were created so that we can apply shard level bucket count thresholds in the reduce phase.
+        return new AggregationReduceableSearchResult(
+            // using topLevelReduce here as PipelineTreeSource needs to be sent to coordinator in older release of OpenSearch. The actual
+            // evaluation of pipeline aggregation though happens on the coordinator during final reduction phase
+            InternalAggregations.topLevelReduce(Collections.singletonList(internalAggregations), context.partialOnShard())
+        );
+    }
 }

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

@@ -41,6 +41,8 @@
 import org.opensearch.core.xcontent.XContentBuilder;
 import org.opensearch.rest.action.search.RestSearchAction;
 import org.opensearch.script.ScriptService;
+import org.opensearch.search.aggregations.bucket.LocalBucketCountThresholds;
+import org.opensearch.search.aggregations.bucket.terms.TermsAggregator;
 import org.opensearch.search.aggregations.pipeline.PipelineAggregator;
 import org.opensearch.search.aggregations.pipeline.PipelineAggregator.PipelineTree;
 import org.opensearch.search.aggregations.support.AggregationPath;
@@ -161,6 +163,18 @@ public boolean isSliceLevel() {
             return this.isSliceLevel;
         }
 
+        /**
+         * For slice level partial reduce we will apply shard level `shard_size` and `shard_min_doc_count` limits
+         * whereas for coordinator level partial reduce it will use top level `size` and `min_doc_count`
+         */
+        public LocalBucketCountThresholds asLocalBucketCountThresholds(TermsAggregator.BucketCountThresholds bucketCountThresholds) {
+            if (isSliceLevel()) {
+                return new LocalBucketCountThresholds(bucketCountThresholds.getShardMinDocCount(), bucketCountThresholds.getShardSize());
+            } else {
+                return new LocalBucketCountThresholds(bucketCountThresholds.getMinDocCount(), bucketCountThresholds.getRequiredSize());
+            }
+        }
+
         public BigArrays bigArrays() {
             return bigArrays;
         }

server/src/main/java/org/opensearch/search/aggregations/NonGlobalAggCollectorManager.java

@@ -14,6 +14,7 @@
 import org.opensearch.search.profile.query.CollectorResult;
 
 import java.io.IOException;
+import java.util.Collections;
 import java.util.Objects;
 
 /**
@@ -38,4 +39,15 @@ public Collector newCollector() throws IOException {
             return super.newCollector();
         }
     }
+
+    @Override
+    protected AggregationReduceableSearchResult buildAggregationResult(InternalAggregations internalAggregations) {
+        // Reduce the aggregations across slices before sending to the coordinator. We will perform shard level reduce as long as any slices
+        // were created so that we can apply shard level bucket count thresholds in the reduce phase.
+        return new AggregationReduceableSearchResult(
+            // using topLevelReduce here as PipelineTreeSource needs to be sent to coordinator in older release of OpenSearch. The actual
+            // evaluation of pipeline aggregation though happens on the coordinator during final reduction phase
+            InternalAggregations.topLevelReduce(Collections.singletonList(internalAggregations), context.partialOnShard())
+        );
+    }
 }

server/src/main/java/org/opensearch/search/aggregations/bucket/LocalBucketCountThresholds.java

@@ -0,0 +1,36 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * The OpenSearch Contributors require contributions made to
+ * this file be licensed under the Apache-2.0 license or a
+ * compatible open source license.
+ */
+
+package org.opensearch.search.aggregations.bucket;
+
+import org.opensearch.search.aggregations.bucket.terms.TermsAggregator;
+
+/**
+ * BucketCountThresholds type that holds the local (either shard level or request level) bucket count thresholds in minDocCount and requireSize fields.
+ * Similar to {@link TermsAggregator.BucketCountThresholds} however only provides getters for the local members and no setters.
+ *
+ * @opensearch.internal
+ */
+public class LocalBucketCountThresholds {
+
+    private final long minDocCount;
+    private final int requiredSize;
+
+    public LocalBucketCountThresholds(long localminDocCount, int localRequiredSize) {
+        this.minDocCount = localminDocCount;
+        this.requiredSize = localRequiredSize;
+    }
+
+    public int getRequiredSize() {
+        return requiredSize;
+    }
+
+    public long getMinDocCount() {
+        return minDocCount;
+    }
+}

server/src/main/java/org/opensearch/search/aggregations/bucket/terms/AbstractStringTermsAggregator.java

@@ -76,15 +76,14 @@ protected StringTerms buildEmptyTermsAggregation() {
             name,
             order,
             order,
-            bucketCountThresholds.getRequiredSize(),
-            bucketCountThresholds.getMinDocCount(),
             metadata(),
             format,
             bucketCountThresholds.getShardSize(),
             showTermDocCountError,
             0,
             emptyList(),
-            0
+            0,
+            bucketCountThresholds
         );
     }
 
@@ -95,14 +94,13 @@ protected SignificantStringTerms buildEmptySignificantTermsAggregation(long subs
         int supersetSize = topReader.numDocs();
         return new SignificantStringTerms(
             name,
-            bucketCountThresholds.getRequiredSize(),
-            bucketCountThresholds.getMinDocCount(),
             metadata(),
             format,
             subsetSize,
             supersetSize,
             significanceHeuristic,
-            emptyList()
+            emptyList(),
+            bucketCountThresholds
         );
     }
 }

server/src/main/java/org/opensearch/search/aggregations/bucket/terms/DoubleTerms.java

@@ -130,29 +130,27 @@ public DoubleTerms(
         String name,
         BucketOrder reduceOrder,
         BucketOrder order,
-        int requiredSize,
-        long minDocCount,
         Map<String, Object> metadata,
         DocValueFormat format,
         int shardSize,
         boolean showTermDocCountError,
         long otherDocCount,
         List<Bucket> buckets,
-        long docCountError
+        long docCountError,
+        TermsAggregator.BucketCountThresholds bucketCountThresholds
     ) {
         super(
             name,
             reduceOrder,
             order,
-            requiredSize,
-            minDocCount,
             metadata,
             format,
             shardSize,
             showTermDocCountError,
             otherDocCount,
             buckets,
-            docCountError
+            docCountError,
+            bucketCountThresholds
         );
     }
 
@@ -174,15 +172,14 @@ public DoubleTerms create(List<Bucket> buckets) {
             name,
             reduceOrder,
             order,
-            requiredSize,
-            minDocCount,
             metadata,
             format,
             shardSize,
             showTermDocCountError,
             otherDocCount,
             buckets,
-            docCountError
+            docCountError,
+            bucketCountThresholds
         );
     }
 
@@ -204,15 +201,14 @@ protected DoubleTerms create(String name, List<Bucket> buckets, BucketOrder redu
             name,
             reduceOrder,
             order,
-            requiredSize,
-            minDocCount,
             getMetadata(),
             format,
             shardSize,
             showTermDocCountError,
             otherDocCount,
             buckets,
-            docCountError
+            docCountError,
+            bucketCountThresholds
         );
     }
 

server/src/main/java/org/opensearch/search/aggregations/bucket/terms/GlobalOrdinalsStringTermsAggregator.java

@@ -57,6 +57,7 @@
 import org.opensearch.search.aggregations.InternalOrder;
 import org.opensearch.search.aggregations.LeafBucketCollector;
 import org.opensearch.search.aggregations.LeafBucketCollectorBase;
+import org.opensearch.search.aggregations.bucket.LocalBucketCountThresholds;
 import org.opensearch.search.aggregations.bucket.terms.SignificanceLookup.BackgroundFrequencyForBytes;
 import org.opensearch.search.aggregations.bucket.terms.heuristic.SignificanceHeuristic;
 import org.opensearch.search.aggregations.support.ValuesSource;
@@ -603,6 +604,7 @@ abstract class ResultStrategy<
         TB extends InternalMultiBucketAggregation.InternalBucket> implements Releasable {
 
         private InternalAggregation[] buildAggregations(long[] owningBucketOrds) throws IOException {
+            LocalBucketCountThresholds localBucketCountThresholds = context.asLocalBucketCountThresholds(bucketCountThresholds);
             if (valueCount == 0) { // no context in this reader
                 InternalAggregation[] results = new InternalAggregation[owningBucketOrds.length];
                 for (int ordIdx = 0; ordIdx < owningBucketOrds.length; ordIdx++) {
@@ -615,11 +617,11 @@ private InternalAggregation[] buildAggregations(long[] owningBucketOrds) throws
             long[] otherDocCount = new long[owningBucketOrds.length];
             for (int ordIdx = 0; ordIdx < owningBucketOrds.length; ordIdx++) {
                 final int size;
-                if (bucketCountThresholds.getMinDocCount() == 0) {
+                if (localBucketCountThresholds.getMinDocCount() == 0) {
                     // if minDocCount == 0 then we can end up with more buckets then maxBucketOrd() returns
-                    size = (int) Math.min(valueCount, bucketCountThresholds.getShardSize());
+                    size = (int) Math.min(valueCount, localBucketCountThresholds.getRequiredSize());
                 } else {
-                    size = (int) Math.min(maxBucketOrd(), bucketCountThresholds.getShardSize());
+                    size = (int) Math.min(maxBucketOrd(), localBucketCountThresholds.getRequiredSize());
                 }
                 PriorityQueue<TB> ordered = buildPriorityQueue(size);
                 final int finalOrdIdx = ordIdx;
@@ -630,7 +632,7 @@ private InternalAggregation[] buildAggregations(long[] owningBucketOrds) throws
                     @Override
                     public void accept(long globalOrd, long bucketOrd, long docCount) throws IOException {
                         otherDocCount[finalOrdIdx] += docCount;
-                        if (docCount >= bucketCountThresholds.getShardMinDocCount()) {
+                        if (docCount >= localBucketCountThresholds.getMinDocCount()) {
                             if (spare == null) {
                                 spare = buildEmptyTemporaryBucket();
                             }
@@ -799,15 +801,14 @@ StringTerms buildResult(long owningBucketOrd, long otherDocCount, StringTerms.Bu
                 name,
                 reduceOrder,
                 order,
-                bucketCountThresholds.getRequiredSize(),
-                bucketCountThresholds.getMinDocCount(),
                 metadata(),
                 format,
                 bucketCountThresholds.getShardSize(),
                 showTermDocCountError,
                 otherDocCount,
                 Arrays.asList(topBuckets),
-                0
+                0,
+                bucketCountThresholds
             );
         }
 
@@ -924,14 +925,13 @@ void buildSubAggs(SignificantStringTerms.Bucket[][] topBucketsPreOrd) throws IOE
         SignificantStringTerms buildResult(long owningBucketOrd, long otherDocCount, SignificantStringTerms.Bucket[] topBuckets) {
             return new SignificantStringTerms(
                 name,
-                bucketCountThresholds.getRequiredSize(),
-                bucketCountThresholds.getMinDocCount(),
                 metadata(),
                 format,
                 subsetSize(owningBucketOrd),
                 supersetSize,
                 significanceHeuristic,
-                Arrays.asList(topBuckets)
+                Arrays.asList(topBuckets),
+                bucketCountThresholds
             );
         }