Reapply "Switch percentiles implementation to MergingDigest (#18124)"

CHANGELOG.md

@@ -10,7 +10,6 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 - Add a mapper for context aware segments grouping criteria ([#19233](https://github.com/opensearch-project/OpenSearch/pull/19233))
 - Return full error for GRPC error response ([#19568](https://github.com/opensearch-project/OpenSearch/pull/19568))
 - Add pluggable gRPC interceptors with explicit ordering([#19005](https://github.com/opensearch-project/OpenSearch/pull/19005))
-
 - Add metrics for the merged segment warmer feature ([#18929](https://github.com/opensearch-project/OpenSearch/pull/18929))
 
 ### Changed
@@ -19,6 +18,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 - Omit maxScoreCollector in SimpleTopDocsCollectorContext when concurrent segment search enabled ([#19584](https://github.com/opensearch-project/OpenSearch/pull/19584))
 - Onboarding new maven snapshots publishing to s3 ([#19619](https://github.com/opensearch-project/OpenSearch/pull/19619))
 - Remove MultiCollectorWrapper and use MultiCollector in Lucene instead ([#19595](https://github.com/opensearch-project/OpenSearch/pull/19595))
+- Change implementation for `percentiles` aggregation for latency improvement ([#19648](https://github.com/opensearch-project/OpenSearch/pull/19648))
 
 ### Fixed
 - Fix Allocation and Rebalance Constraints of WeightFunction are incorrectly reset ([#19012](https://github.com/opensearch-project/OpenSearch/pull/19012))

gradle/libs.versions.toml

@@ -27,7 +27,7 @@ protobuf          = "3.25.8"
 jakarta_annotation = "1.3.5"
 google_http_client = "1.44.1"
 google_auth       = "1.29.0"
-tdigest           = "3.3"
+tdigest           = "3.3" # Warning: Before updating tdigest, ensure its serialization code for MergingDigest hasn't changed
 hdrhistogram      = "2.2.2"
 grpc              = "1.75.0"
 json_smart        = "2.5.2"

server/src/main/java/org/opensearch/search/aggregations/metrics/InternalMedianAbsoluteDeviation.java

@@ -43,6 +43,8 @@
 import java.util.Map;
 import java.util.Objects;
 
+import com.tdunning.math.stats.Centroid;
+
 /**
  * Implementation of median absolute deviation agg
  *
@@ -57,11 +59,14 @@ static double computeMedianAbsoluteDeviation(TDigestState valuesSketch) {
         } else {
             final double approximateMedian = valuesSketch.quantile(0.5);
             final TDigestState approximatedDeviationsSketch = new TDigestState(valuesSketch.compression());
-            valuesSketch.centroids().forEach(centroid -> {
+            for (Centroid centroid : valuesSketch.centroids()) {
                 final double deviation = Math.abs(approximateMedian - centroid.mean());
-                approximatedDeviationsSketch.add(deviation, centroid.count());
-            });
-
+                // Weighted add() isn't supported for faster MergingDigest implementation, so add iteratively instead. see
+                // https://github.com/tdunning/t-digest/issues/167
+                for (int i = 0; i < centroid.count(); i++) {
+                    approximatedDeviationsSketch.add(deviation);
+                }
+            }
             return approximatedDeviationsSketch.quantile(0.5);
         }
     }

server/src/main/java/org/opensearch/search/aggregations/metrics/TDigestState.java

@@ -31,21 +31,25 @@
 
 package org.opensearch.search.aggregations.metrics;
 
+import org.opensearch.Version;
 import org.opensearch.core.common.io.stream.StreamInput;
 import org.opensearch.core.common.io.stream.StreamOutput;
 
 import java.io.IOException;
+import java.nio.ByteBuffer;
 import java.util.Iterator;
+import java.util.List;
 
 import com.tdunning.math.stats.AVLTreeDigest;
 import com.tdunning.math.stats.Centroid;
+import com.tdunning.math.stats.MergingDigest;
 
 /**
  * Extension of {@link com.tdunning.math.stats.TDigest} with custom serialization.
  *
  * @opensearch.internal
  */
-public class TDigestState extends AVLTreeDigest {
+public class TDigestState extends MergingDigest {
 
     private final double compression;
 
@@ -54,28 +58,64 @@ public TDigestState(double compression) {
         this.compression = compression;
     }
 
+    private TDigestState(double compression, MergingDigest in) {
+        super(compression);
+        this.compression = compression;
+        this.add(List.of(in));
+    }
+
     @Override
     public double compression() {
         return compression;
     }
 
     public static void write(TDigestState state, StreamOutput out) throws IOException {
-        out.writeDouble(state.compression);
-        out.writeVInt(state.centroidCount());
-        for (Centroid centroid : state.centroids()) {
-            out.writeDouble(centroid.mean());
-            out.writeVLong(centroid.count());
+        if (out.getVersion().before(Version.V_3_4_0)) {
+            out.writeDouble(state.compression);
+            out.writeVInt(state.centroidCount());
+            for (Centroid centroid : state.centroids()) {
+                out.writeDouble(centroid.mean());
+                out.writeVLong(centroid.count());
+            }
+        } else {
+            int byteSize = state.byteSize();
+            out.writeVInt(byteSize);
+            ByteBuffer buf = ByteBuffer.allocate(byteSize);
+            state.asBytes(buf);
+            out.writeBytes(buf.array());
         }
     }
 
     public static TDigestState read(StreamInput in) throws IOException {
-        double compression = in.readDouble();
-        TDigestState state = new TDigestState(compression);
-        int n = in.readVInt();
-        for (int i = 0; i < n; i++) {
-            state.add(in.readDouble(), in.readVInt());
+        if (in.getVersion().before(Version.V_3_4_0)) {
+            // In older versions TDigestState was based on AVLTreeDigest. Load centroids into this class, then add it to MergingDigest.
+            double compression = in.readDouble();
+            AVLTreeDigest treeDigest = new AVLTreeDigest(compression);
+            int n = in.readVInt();
+            if (n > 0) {
+                for (int i = 0; i < n; i++) {
+                    treeDigest.add(in.readDouble(), in.readVInt());
+                }
+                TDigestState state = new TDigestState(compression);
+                state.add(List.of(treeDigest));
+                return state;
+            }
+            return new TDigestState(compression);
+        } else {
+            // For MergingDigest, adding the original centroids in ascending order to a new, empty MergingDigest isn't guaranteed
+            // to produce a MergingDigest whose centroids are exactly equal to the originals.
+            // So, use the library's serialization code to ensure we get the exact same centroids, allowing us to compare with equals().
+            // The AVLTreeDigest had the same limitation for equals() where it was only guaranteed to return true if the other object was
+            // produced by de/serializing the object, so this should be fine.
+            int byteSize = in.readVInt();
+            byte[] bytes = new byte[byteSize];
+            in.readBytes(bytes, 0, byteSize);
+            MergingDigest mergingDigest = MergingDigest.fromBytes(ByteBuffer.wrap(bytes));
+            if (mergingDigest.centroids().isEmpty()) {
+                return new TDigestState(mergingDigest.compression());
+            }
+            return new TDigestState(mergingDigest.compression(), mergingDigest);
         }
-        return state;
     }
 
     @Override

server/src/test/java/org/opensearch/search/aggregations/metrics/InternalTDigestPercentilesRanksTests.java

@@ -54,7 +54,7 @@ protected InternalTDigestPercentileRanks createTestInstance(
         Arrays.stream(values).forEach(state::add);
 
         // the number of centroids is defined as <= the number of samples inserted
-        assertTrue(state.centroidCount() <= values.length);
+        assertTrue(state.centroids().size() <= values.length);
         return new InternalTDigestPercentileRanks(name, percents, state, keyed, format, metadata);
     }
 
@@ -66,7 +66,7 @@ protected void assertReduced(InternalTDigestPercentileRanks reduced, List<Intern
         double max = Double.NEGATIVE_INFINITY;
         long totalCount = 0;
         for (InternalTDigestPercentileRanks ranks : inputs) {
-            if (ranks.state.centroidCount() == 0) {
+            if (ranks.state.centroids().isEmpty()) {
                 // quantiles would return NaN
                 continue;
             }

server/src/test/java/org/opensearch/search/aggregations/metrics/InternalTDigestPercentilesTests.java

@@ -54,7 +54,7 @@ protected InternalTDigestPercentiles createTestInstance(
         Arrays.stream(values).forEach(state::add);
 
         // the number of centroids is defined as <= the number of samples inserted
-        assertTrue(state.centroidCount() <= values.length);
+        assertTrue(state.centroids().size() <= values.length);
         return new InternalTDigestPercentiles(name, percents, state, keyed, format, metadata);
     }
 

server/src/test/java/org/opensearch/search/aggregations/metrics/TDigestPercentilesAggregatorTests.java

@@ -104,7 +104,7 @@ public void testSomeMatchesSortedNumericDocValues() throws IOException {
             iw.addDocument(singleton(new SortedNumericDocValuesField("number", 0)));
         }, tdigest -> {
             assertEquals(7L, tdigest.state.size());
-            assertEquals(7L, tdigest.state.centroidCount());
+            assertEquals(7L, tdigest.state.centroids().size());
             assertEquals(5.0d, tdigest.percentile(75), 0.0d);
             assertEquals("5.0", tdigest.percentileAsString(75));
             assertEquals(3.0d, tdigest.percentile(71), 0.0d);
@@ -128,7 +128,7 @@ public void testSomeMatchesNumericDocValues() throws IOException {
             iw.addDocument(singleton(new NumericDocValuesField("number", 0)));
         }, tdigest -> {
             assertEquals(tdigest.state.size(), 7L);
-            assertTrue(tdigest.state.centroidCount() <= 7L);
+            assertTrue(tdigest.state.centroids().size() <= 7L);
             assertEquals(8.0d, tdigest.percentile(100), 0.0d);
             assertEquals("8.0", tdigest.percentileAsString(100));
             assertEquals(8.0d, tdigest.percentile(88), 0.0d);
@@ -156,7 +156,7 @@ public void testQueryFiltering() throws IOException {
 
         testCase(LongPoint.newRangeQuery("row", 1, 4), docs, tdigest -> {
             assertEquals(4L, tdigest.state.size());
-            assertEquals(4L, tdigest.state.centroidCount());
+            assertEquals(4L, tdigest.state.centroids().size());
             assertEquals(2.0d, tdigest.percentile(100), 0.0d);
             assertEquals(1.0d, tdigest.percentile(50), 0.0d);
             assertEquals(1.0d, tdigest.percentile(25), 0.0d);
@@ -165,7 +165,7 @@ public void testQueryFiltering() throws IOException {
 
         testCase(LongPoint.newRangeQuery("row", 100, 110), docs, tdigest -> {
             assertEquals(0L, tdigest.state.size());
-            assertEquals(0L, tdigest.state.centroidCount());
+            assertEquals(0L, tdigest.state.centroids().size());
             assertFalse(AggregationInspectionHelper.hasValue(tdigest));
         });
     }