+ * After testing this, the agg serialized size is around 5MB. + * The CB is set to 7MB, as the reduction is expected to add an extra 50% overhead. + *
+ */ + public void testCBTrippingOnReduction() throws IOException { + createIndex(); + addDocs(100, 100, 100); + + // Some leaks (Check ESTestCase#loggedLeaks) aren't logged unless we run the test twice. + // So we run it multiple times to ensure everything gets collected before the final test checks. + for (int i = 0; i < 10; i++) { + assertCBTrip( + () -> internalCluster().coordOnlyNodeClient() + .prepareSearch("index") + .setSize(0) + .addAggregation( + composite( + "composite", + List.of( + new TermsValuesSourceBuilder("integer").field("integer"), + new TermsValuesSourceBuilder("long").field("long") + ) + ).size(5000).subAggregation(topHits("top_hits").size(10)) + ) + .setBatchedReduceSize(randomIntBetween(2, 5)), + e -> { + var completeException = ExceptionsHelper.stackTrace(e); + // If a shard fails, we can't check reduction + assumeTrue(completeException, e.shardFailures().length == 0); + assertThat(e.getCause(), instanceOf(CircuitBreakingException.class)); + assertThat(completeException, containsString("QueryPhaseResultConsumer.reduce")); + } + ); + } + } + + public void assertCBTrip(Supplier+ * Similar to {@link #testBatchedEstimateSizeTooBig()}, but this tests the extra size + *
+ */ + public void testBatchedEstimateSizeTooBigAfterDeserialization() throws Exception { + SearchRequest searchRequest = new SearchRequest("index"); + searchRequest.source(new SearchSourceBuilder().aggregation(new SumAggregationBuilder("sum"))); + + var aggCount = randomIntBetween(1, 10); + var circuitBreakerLimit = ByteSizeValue.ofMb(256); + var circuitBreaker = newLimitedBreaker(circuitBreakerLimit); + // Less than the CB, but more after the 1.5x + long aggregationEstimatedSize = (long) (circuitBreakerLimit.getBytes() * 0.75 / aggCount); + long totalAggregationsEstimatedSize = aggregationEstimatedSize * aggCount; + + try ( + QueryPhaseResultConsumer queryPhaseResultConsumer = new QueryPhaseResultConsumer( + searchRequest, + executor, + circuitBreaker, + searchPhaseController, + () -> false, + new SearchProgressListener() { + }, + 10, + e -> {} + ) + ) { + for (int i = 0; i < aggCount; i++) { + // Add a dummy merge result with a high estimated size + var mergeResult = new QueryPhaseResultConsumer.MergeResult(List.of(), null, new DelegatingDelayableWriteable<>(() -> { + fail("This shouldn't be called"); + return null; + }), aggregationEstimatedSize); + queryPhaseResultConsumer.addBatchedPartialResult(new SearchPhaseController.TopDocsStats(0), mergeResult); + } + + try { + queryPhaseResultConsumer.reduce(); + fail("Expecting a circuit breaking exception to be thrown"); + } catch (CircuitBreakingException e) { + assertThat(circuitBreaker.getUsed(), greaterThanOrEqualTo(aggregationEstimatedSize)); + // A final +0.5x is added to account for the serialized->deserialized extra size + assertThat(e.getBytesWanted(), equalTo(Math.round(totalAggregationsEstimatedSize * 0.5))); + } + } + } + + /** + * DelayableWriteable that delegates expansion to a supplier. + */ + private static class DelegatingDelayableWriteable* A test cluster creates a set of nodes in the background before the test starts. The number of nodes in the cluster is * determined at random and can change across tests. The {@link ClusterScope} allows configuring the initial number of nodes - * that are created before the tests start. + * that are created before the tests start. More information about node configurations and settings in {@link InternalTestCluster}. *
* {@literal @}NodeScope(scope=Scope.SUITE, numDataNodes=3)
* public class SomeIT extends ESIntegTestCase {
@@ -2149,7 +2149,7 @@ public enum Scope {
/**
* Indicates whether the cluster can have dedicated master nodes. If {@code false} means data nodes will serve as master nodes
- * and there will be no dedicated master (and data) nodes. Default is {@code false} which means
+ * and there will be no dedicated master (and data) nodes. Default is {@code true} which means
* dedicated master nodes will be randomly used.
*/
boolean supportsDedicatedMasters() default true;
diff --git a/test/framework/src/main/java/org/elasticsearch/test/InternalTestCluster.java b/test/framework/src/main/java/org/elasticsearch/test/InternalTestCluster.java
index 453cddd504364..dbf88cee69778 100644
--- a/test/framework/src/main/java/org/elasticsearch/test/InternalTestCluster.java
+++ b/test/framework/src/main/java/org/elasticsearch/test/InternalTestCluster.java
@@ -176,6 +176,20 @@
* to the async nature of Elasticsearch in combination with randomized testing. Once Threads and asynchronous calls
* are involved reproducibility is very limited. This class should only be used through {@link ESIntegTestCase}.
*
+ * {@link NodeConfigurationSource#nodeSettings} ordinals
+ * The supplied node ordinal for this node settings getter is generated as follows:
+ *
+ * -
+ * The first indices correspond to the dedicated master nodes, enabled with the {@code randomlyAddDedicatedMasters} setting.
+ * Its amount can't currently be configured.
+ *
+ * -
+ * Then, the data nodes, controlled by the @{code minNumDataNodes} and @{code maxNumDataNodes} settings.
+ *
+ * -
+ * Finally, the coordinating-only, configured with the {@code numClientNodes} setting.
+ *
+ *
*/
public final class InternalTestCluster extends TestCluster {