[ML] Fork trained model stats work to thread pool threads (#99439)

The steps that assemble the trained model stats response
largely ran on the transport worker thread. Even though
none of the steps look massively expensive, it's likely
one of them is, because we've observed client lockups
(Kibana) while trained model stats is running. This change
moves the steps into the ML utility thread pool to avoid
any risk of blocking the transport thread.

---------

Co-authored-by: David Turner <david.turner@elastic.co>

x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/action/TransportGetTrainedModelsStatsAction.java

@@ -9,6 +9,7 @@ package org.elasticsearch.xpack.ml.action;
 import org.apache.logging.log4j.LogManager;
 import org.apache.logging.log4j.Logger;
 import org.elasticsearch.action.ActionListener;
+import org.elasticsearch.action.ActionRunnable;
 import org.elasticsearch.action.admin.cluster.node.stats.NodeStats;
 import org.elasticsearch.action.admin.cluster.node.stats.NodesStatsRequest;
 import org.elasticsearch.action.admin.cluster.node.stats.NodesStatsRequestParameters;
@@ -17,7 +18,8 @@ import org.elasticsearch.action.admin.cluster.node.stats.TransportNodesStatsActi
 import org.elasticsearch.action.search.SearchRequest;
 import org.elasticsearch.action.search.TransportSearchAction;
 import org.elasticsearch.action.support.ActionFilters;
-import org.elasticsearch.action.support.HandledTransportAction;
+import org.elasticsearch.action.support.SubscribableListener;
+import org.elasticsearch.action.support.TransportAction;
 import org.elasticsearch.client.internal.Client;
 import org.elasticsearch.cluster.ClusterState;
 import org.elasticsearch.cluster.service.ClusterService;
@@ -26,8 +28,6 @@ import org.elasticsearch.common.document.DocumentField;
 import org.elasticsearch.common.inject.Inject;
 import org.elasticsearch.common.metrics.CounterMetric;
 import org.elasticsearch.common.util.Maps;
-import org.elasticsearch.common.util.concurrent.EsExecutors;
-import org.elasticsearch.common.util.concurrent.ListenableFuture;
 import org.elasticsearch.common.util.set.Sets;
 import org.elasticsearch.core.Tuple;
 import org.elasticsearch.index.query.QueryBuilder;
@@ -37,6 +37,7 @@ import org.elasticsearch.search.SearchHit;
 import org.elasticsearch.search.sort.SortOrder;
 import org.elasticsearch.tasks.Task;
 import org.elasticsearch.tasks.TaskId;
+import org.elasticsearch.threadpool.ThreadPool;
 import org.elasticsearch.transport.TransportService;
 import org.elasticsearch.xpack.core.action.util.ExpandedIdsMatcher;
 import org.elasticsearch.xpack.core.ml.action.GetDeploymentStatsAction;
@@ -53,6 +54,7 @@ import org.elasticsearch.xpack.core.ml.inference.persistence.InferenceIndexConst
 import org.elasticsearch.xpack.core.ml.inference.trainedmodel.InferenceStats;
 import org.elasticsearch.xpack.core.ml.inference.trainedmodel.TrainedModelSizeStats;
 import org.elasticsearch.xpack.core.ml.utils.TransportVersionUtils;
+import org.elasticsearch.xpack.ml.MachineLearning;
 import org.elasticsearch.xpack.ml.inference.persistence.TrainedModelDefinitionDoc;
 import org.elasticsearch.xpack.ml.inference.persistence.TrainedModelProvider;
 
@@ -65,6 +67,7 @@ import java.util.LinkedHashMap;
 import java.util.List;
 import java.util.Map;
 import java.util.Set;
+import java.util.concurrent.Executor;
 import java.util.function.Function;
 import java.util.stream.Collectors;
 import java.util.stream.Stream;
@@ -73,7 +76,7 @@ import static org.elasticsearch.xpack.core.ClientHelper.ML_ORIGIN;
 import static org.elasticsearch.xpack.core.ClientHelper.executeAsyncWithOrigin;
 import static org.elasticsearch.xpack.ml.utils.InferenceProcessorInfoExtractor.pipelineIdsByResource;
 
-public class TransportGetTrainedModelsStatsAction extends HandledTransportAction<
+public class TransportGetTrainedModelsStatsAction extends TransportAction<
     GetTrainedModelsStatsAction.Request,
     GetTrainedModelsStatsAction.Response> {
 
@@ -82,25 +85,22 @@ public class TransportGetTrainedModelsStatsAction extends HandledTransportAction
     private final Client client;
     private final ClusterService clusterService;
     private final TrainedModelProvider trainedModelProvider;
+    private final Executor executor;
 
     @Inject
     public TransportGetTrainedModelsStatsAction(
         TransportService transportService,
         ActionFilters actionFilters,
         ClusterService clusterService,
+        ThreadPool threadPool,
         TrainedModelProvider trainedModelProvider,
         Client client
     ) {
-        super(
-            GetTrainedModelsStatsAction.NAME,
-            transportService,
-            actionFilters,
-            GetTrainedModelsStatsAction.Request::new,
-            EsExecutors.DIRECT_EXECUTOR_SERVICE
-        );
+        super(GetTrainedModelsStatsAction.NAME, actionFilters, transportService.getTaskManager());
         this.client = client;
         this.clusterService = clusterService;
         this.trainedModelProvider = trainedModelProvider;
+        this.executor = threadPool.executor(MachineLearning.UTILITY_THREAD_POOL_NAME);
     }
 
     @Override
@@ -108,6 +108,15 @@ public class TransportGetTrainedModelsStatsAction extends HandledTransportAction
         Task task,
         GetTrainedModelsStatsAction.Request request,
         ActionListener<GetTrainedModelsStatsAction.Response> listener
+    ) {
+        // workaround for https://github.com/elastic/elasticsearch/issues/97916 - TODO remove this when we can
+        executor.execute(ActionRunnable.wrap(listener, l -> doExecuteForked(task, request, l)));
+    }
+
+    protected void doExecuteForked(
+        Task task,
+        GetTrainedModelsStatsAction.Request request,
+        ActionListener<GetTrainedModelsStatsAction.Response> listener
     ) {
         final TaskId parentTaskId = new TaskId(clusterService.localNode().getId(), task.getId());
         final ModelAliasMetadata modelAliasMetadata = ModelAliasMetadata.fromState(clusterService.state());
@@ -116,101 +125,108 @@ public class TransportGetTrainedModelsStatsAction extends HandledTransportAction
 
         GetTrainedModelsStatsAction.Response.Builder responseBuilder = new GetTrainedModelsStatsAction.Response.Builder();
 
-        ListenableFuture<Map<String, TrainedModelSizeStats>> modelSizeStatsListener = new ListenableFuture<>();
-        modelSizeStatsListener.addListener(listener.delegateFailureAndWrap((l, modelSizeStatsByModelId) -> {
-            responseBuilder.setModelSizeStatsByModelId(modelSizeStatsByModelId);
-            l.onResponse(
-                responseBuilder.build(modelToDeployments(responseBuilder.getExpandedModelIdsWithAliases().keySet(), assignmentMetadata))
-            );
-        }));
-
-        ListenableFuture<GetDeploymentStatsAction.Response> deploymentStatsListener = new ListenableFuture<>();
-        deploymentStatsListener.addListener(listener.delegateFailureAndWrap((delegate, deploymentStats) -> {
-            // deployment stats for each matching deployment
-            // not necessarily for all models
-            responseBuilder.setDeploymentStatsByDeploymentId(
-                deploymentStats.getStats()
-                    .results()
+        SubscribableListener
+
+            .<Tuple<Long, Map<String, Set<String>>>>newForked(l -> {
+                // When the request resource is a deployment find the model used in that deployment for the model stats
+                final var idExpression = addModelsUsedInMatchingDeployments(request.getResourceId(), assignmentMetadata);
+
+                logger.debug("Expanded models/deployment Ids request [{}]", idExpression);
+
+                // the request id may contain deployment ids
+                // It is not an error if these don't match a model id but
+                // they need to be included in case the deployment id is also
+                // a model id. Hence, the `matchedDeploymentIds` parameter
+                trainedModelProvider.expandIds(
+                    idExpression,
+                    request.isAllowNoResources(),
+                    request.getPageParams(),
+                    Collections.emptySet(),
+                    modelAliasMetadata,
+                    parentTaskId,
+                    matchedDeploymentIds,
+                    l
+                );
+            })
+            .andThenAccept(tuple -> responseBuilder.setExpandedModelIdsWithAliases(tuple.v2()).setTotalModelCount(tuple.v1()))
+
+            .<NodesStatsResponse>andThen(
+                (l, ignored) -> executeAsyncWithOrigin(
+                    client,
+                    ML_ORIGIN,
+                    TransportNodesStatsAction.TYPE,
+                    nodeStatsRequest(clusterService.state(), parentTaskId),
+                    l
+                )
+            )
+            .<List<InferenceStats>>andThen(executor, null, (l, nodesStatsResponse) -> {
+                // find all pipelines whether using the model id, alias or deployment id.
+                Set<String> allPossiblePipelineReferences = responseBuilder.getExpandedModelIdsWithAliases()
+                    .entrySet()
                     .stream()
-                    .collect(Collectors.toMap(AssignmentStats::getDeploymentId, Function.identity()))
-            );
+                    .flatMap(entry -> Stream.concat(entry.getValue().stream(), Stream.of(entry.getKey())))
+                    .collect(Collectors.toSet());
+                allPossiblePipelineReferences.addAll(matchedDeploymentIds);
 
-            int numberOfAllocations = deploymentStats.getStats().results().stream().mapToInt(AssignmentStats::getNumberOfAllocations).sum();
-            modelSizeStats(
-                responseBuilder.getExpandedModelIdsWithAliases(),
-                request.isAllowNoResources(),
-                parentTaskId,
-                modelSizeStatsListener,
-                numberOfAllocations
-            );
-        }));
-
-        ListenableFuture<List<InferenceStats>> inferenceStatsListener = new ListenableFuture<>();
-        // inference stats are per model and are only
-        // persisted for boosted tree models
-        inferenceStatsListener.addListener(listener.delegateFailureAndWrap((l, inferenceStats) -> {
-            responseBuilder.setInferenceStatsByModelId(
-                inferenceStats.stream().collect(Collectors.toMap(InferenceStats::getModelId, Function.identity()))
-            );
-            getDeploymentStats(client, request.getResourceId(), parentTaskId, assignmentMetadata, deploymentStatsListener);
-        }));
-
-        ListenableFuture<NodesStatsResponse> nodesStatsListener = new ListenableFuture<>();
-        nodesStatsListener.addListener(listener.delegateFailureAndWrap((delegate, nodesStatsResponse) -> {
-            // find all pipelines whether using the model id,
-            // alias or deployment id.
-            Set<String> allPossiblePipelineReferences = responseBuilder.getExpandedModelIdsWithAliases()
-                .entrySet()
-                .stream()
-                .flatMap(entry -> Stream.concat(entry.getValue().stream(), Stream.of(entry.getKey())))
-                .collect(Collectors.toSet());
-            allPossiblePipelineReferences.addAll(matchedDeploymentIds);
-
-            Map<String, Set<String>> pipelineIdsByResource = pipelineIdsByResource(clusterService.state(), allPossiblePipelineReferences);
-            Map<String, IngestStats> modelIdIngestStats = inferenceIngestStatsByModelId(
-                nodesStatsResponse,
-                modelAliasMetadata,
-                pipelineIdsByResource
-            );
-            responseBuilder.setIngestStatsByModelId(modelIdIngestStats);
-            trainedModelProvider.getInferenceStats(
-                responseBuilder.getExpandedModelIdsWithAliases().keySet().toArray(new String[0]),
-                parentTaskId,
-                inferenceStatsListener
-            );
-        }));
-
-        ListenableFuture<Tuple<Long, Map<String, Set<String>>>> idsListener = new ListenableFuture<>();
-        idsListener.addListener(listener.delegateFailureAndWrap((delegate, tuple) -> {
-            responseBuilder.setExpandedModelIdsWithAliases(tuple.v2()).setTotalModelCount(tuple.v1());
-            executeAsyncWithOrigin(
-                client,
-                ML_ORIGIN,
-                TransportNodesStatsAction.TYPE,
-                nodeStatsRequest(clusterService.state(), parentTaskId),
-                nodesStatsListener
-            );
-        }));
-
-        // When the request resource is a deployment find the
-        // model used in that deployment for the model stats
-        String idExpression = addModelsUsedInMatchingDeployments(request.getResourceId(), assignmentMetadata);
-        logger.debug("Expanded models/deployment Ids request [{}]", idExpression);
-
-        // the request id may contain deployment ids
-        // It is not an error if these don't match a model id but
-        // they need to be included in case the deployment id is also
-        // a model id. Hence, the `matchedDeploymentIds` parameter
-        trainedModelProvider.expandIds(
-            idExpression,
-            request.isAllowNoResources(),
-            request.getPageParams(),
-            Collections.emptySet(),
-            modelAliasMetadata,
-            parentTaskId,
-            matchedDeploymentIds,
-            idsListener
-        );
+                Map<String, Set<String>> pipelineIdsByResource = pipelineIdsByResource(
+                    clusterService.state(),
+                    allPossiblePipelineReferences
+                );
+                Map<String, IngestStats> modelIdIngestStats = inferenceIngestStatsByModelId(
+                    nodesStatsResponse,
+                    modelAliasMetadata,
+                    pipelineIdsByResource
+                );
+                responseBuilder.setIngestStatsByModelId(modelIdIngestStats);
+                trainedModelProvider.getInferenceStats(
+                    responseBuilder.getExpandedModelIdsWithAliases().keySet().toArray(new String[0]),
+                    parentTaskId,
+                    l
+                );
+            })
+            .andThenAccept(
+                // inference stats are per model and are only persisted for boosted tree models
+                inferenceStats -> responseBuilder.setInferenceStatsByModelId(
+                    inferenceStats.stream().collect(Collectors.toMap(InferenceStats::getModelId, Function.identity()))
+                )
+            )
+
+            .<GetDeploymentStatsAction.Response>andThen(
+                executor,
+                null,
+                (l, ignored) -> getDeploymentStats(client, request.getResourceId(), parentTaskId, assignmentMetadata, l)
+            )
+            .andThenApply(deploymentStats -> {
+                // deployment stats for each matching deployment not necessarily for all models
+                responseBuilder.setDeploymentStatsByDeploymentId(
+                    deploymentStats.getStats()
+                        .results()
+                        .stream()
+                        .collect(Collectors.toMap(AssignmentStats::getDeploymentId, Function.identity()))
+                );
+                return deploymentStats.getStats().results().stream().mapToInt(AssignmentStats::getNumberOfAllocations).sum();
+            })
+
+            .<Map<String, TrainedModelSizeStats>>andThen(
+                executor,
+                null,
+                (l, numberOfAllocations) -> modelSizeStats(
+                    responseBuilder.getExpandedModelIdsWithAliases(),
+                    request.isAllowNoResources(),
+                    parentTaskId,
+                    l,
+                    numberOfAllocations
+                )
+            )
+            .andThenAccept(responseBuilder::setModelSizeStatsByModelId)
+
+            .andThenApply(
+                ignored -> responseBuilder.build(
+                    modelToDeployments(responseBuilder.getExpandedModelIdsWithAliases().keySet(), assignmentMetadata)
+                )
+            )
+
+            .addListener(listener, executor, null);
     }
 
     static String addModelsUsedInMatchingDeployments(String idExpression, TrainedModelAssignmentMetadata assignmentMetadata) {

x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/utils/InferenceProcessorInfoExtractor.java

@@ -13,6 +13,7 @@ import org.elasticsearch.cluster.metadata.Metadata;
 import org.elasticsearch.ingest.ConfigurationUtils;
 import org.elasticsearch.ingest.IngestMetadata;
 import org.elasticsearch.ingest.Pipeline;
+import org.elasticsearch.transport.Transports;
 
 import java.util.HashMap;
 import java.util.LinkedHashSet;
@@ -75,6 +76,7 @@ public final class InferenceProcessorInfoExtractor {
      */
     @SuppressWarnings("unchecked")
     public static Map<String, Set<String>> pipelineIdsByResource(ClusterState state, Set<String> ids) {
+        assert Transports.assertNotTransportThread("non-trivial nested loops over cluster state structures");
         Map<String, Set<String>> pipelineIdsByModelIds = new HashMap<>();
         Metadata metadata = state.metadata();
         if (metadata == null) {