[ML] Inference duration and error metrics (#115876)

Add `es.inference.requests.time` metric around `infer` API.

As recommended by OTel spec, errors are determined by the
presence or absence of the `error.type` attribute in the metric.
"error.type" will be the http status code (as a string) if it is
available, otherwise it will be the name of the exception (e.g.
NullPointerException).

Additional notes:
- ApmInferenceStats is merged into InferenceStats. Originally we planned
  to have multiple implementations, but now we're only using APM.
- Request count is now always recorded, even when there are failures
  loading the endpoint configuration.
- Added a hook in streaming for cancel messages, so we can close the
  metrics when a user cancels the stream.

Author: prwhelan

docs/changelog/115876.yaml

@@ -0,0 +1,5 @@
+pr: 115876
+summary: Inference duration and error metrics
+area: Machine Learning
+type: enhancement
+issues: []

x-pack/plugin/inference/src/main/java/org/elasticsearch/xpack/inference/InferencePlugin.java

@@ -101,7 +101,6 @@
 import org.elasticsearch.xpack.inference.services.ibmwatsonx.IbmWatsonxService;
 import org.elasticsearch.xpack.inference.services.mistral.MistralService;
 import org.elasticsearch.xpack.inference.services.openai.OpenAiService;
-import org.elasticsearch.xpack.inference.telemetry.ApmInferenceStats;
 import org.elasticsearch.xpack.inference.telemetry.InferenceStats;
 
 import java.util.ArrayList;
@@ -239,7 +238,7 @@ public Collection<?> createComponents(PluginServices services) {
         shardBulkInferenceActionFilter.set(actionFilter);
 
         var meterRegistry = services.telemetryProvider().getMeterRegistry();
-        var stats = new PluginComponentBinding<>(InferenceStats.class, ApmInferenceStats.create(meterRegistry));
+        var stats = new PluginComponentBinding<>(InferenceStats.class, InferenceStats.create(meterRegistry));
 
         return List.of(modelRegistry, registry, httpClientManager, stats);
     }

x-pack/plugin/inference/src/main/java/org/elasticsearch/xpack/inference/action/TransportInferenceAction.java

@@ -7,12 +7,15 @@
 
 package org.elasticsearch.xpack.inference.action;
 
+import org.apache.logging.log4j.LogManager;
+import org.apache.logging.log4j.Logger;
 import org.elasticsearch.ElasticsearchStatusException;
 import org.elasticsearch.action.ActionListener;
 import org.elasticsearch.action.support.ActionFilters;
 import org.elasticsearch.action.support.HandledTransportAction;
 import org.elasticsearch.common.util.concurrent.EsExecutors;
 import org.elasticsearch.common.xcontent.ChunkedToXContent;
+import org.elasticsearch.core.Nullable;
 import org.elasticsearch.inference.InferenceService;
 import org.elasticsearch.inference.InferenceServiceRegistry;
 import org.elasticsearch.inference.InferenceServiceResults;
@@ -25,20 +28,22 @@
 import org.elasticsearch.transport.TransportService;
 import org.elasticsearch.xpack.core.inference.action.InferenceAction;
 import org.elasticsearch.xpack.inference.action.task.StreamingTaskManager;
+import org.elasticsearch.xpack.inference.common.DelegatingProcessor;
 import org.elasticsearch.xpack.inference.registry.ModelRegistry;
 import org.elasticsearch.xpack.inference.telemetry.InferenceStats;
+import org.elasticsearch.xpack.inference.telemetry.InferenceTimer;
 
-import java.util.Set;
 import java.util.stream.Collectors;
 
 import static org.elasticsearch.core.Strings.format;
+import static org.elasticsearch.xpack.inference.telemetry.InferenceStats.modelAttributes;
+import static org.elasticsearch.xpack.inference.telemetry.InferenceStats.responseAttributes;
 
 public class TransportInferenceAction extends HandledTransportAction<InferenceAction.Request, InferenceAction.Response> {
+    private static final Logger log = LogManager.getLogger(TransportInferenceAction.class);
     private static final String STREAMING_INFERENCE_TASK_TYPE = "streaming_inference";
     private static final String STREAMING_TASK_ACTION = "xpack/inference/streaming_inference[n]";
 
-    private static final Set<Class<? extends InferenceService>> supportsStreaming = Set.of();
-
     private final ModelRegistry modelRegistry;
     private final InferenceServiceRegistry serviceRegistry;
     private final InferenceStats inferenceStats;
@@ -62,17 +67,22 @@ public TransportInferenceAction(
 
     @Override
     protected void doExecute(Task task, InferenceAction.Request request, ActionListener<InferenceAction.Response> listener) {
+        var timer = InferenceTimer.start();
 
-        ActionListener<UnparsedModel> getModelListener = listener.delegateFailureAndWrap((delegate, unparsedModel) -> {
+        var getModelListener = ActionListener.wrap((UnparsedModel unparsedModel) -> {
             var service = serviceRegistry.getService(unparsedModel.service());
             if (service.isEmpty()) {
-                listener.onFailure(unknownServiceException(unparsedModel.service(), request.getInferenceEntityId()));
+                var e = unknownServiceException(unparsedModel.service(), request.getInferenceEntityId());
+                recordMetrics(unparsedModel, timer, e);
+                listener.onFailure(e);
                 return;
             }
 
             if (request.getTaskType().isAnyOrSame(unparsedModel.taskType()) == false) {
                 // not the wildcard task type and not the model task type
-                listener.onFailure(incompatibleTaskTypeException(request.getTaskType(), unparsedModel.taskType()));
+                var e = incompatibleTaskTypeException(request.getTaskType(), unparsedModel.taskType());
+                recordMetrics(unparsedModel, timer, e);
+                listener.onFailure(e);
                 return;
             }
 
@@ -83,20 +93,69 @@ protected void doExecute(Task task, InferenceAction.Request request, ActionListe
                     unparsedModel.settings(),
                     unparsedModel.secrets()
                 );
-            inferOnService(model, request, service.get(), delegate);
+            inferOnServiceWithMetrics(model, request, service.get(), timer, listener);
+        }, e -> {
+            try {
+                inferenceStats.inferenceDuration().record(timer.elapsedMillis(), responseAttributes(e));
+            } catch (Exception metricsException) {
+                log.atDebug().withThrowable(metricsException).log("Failed to record metrics when the model is missing, dropping metrics");
+            }
+            listener.onFailure(e);
         });
 
         modelRegistry.getModelWithSecrets(request.getInferenceEntityId(), getModelListener);
     }
 
-    private void inferOnService(
+    private void recordMetrics(UnparsedModel model, InferenceTimer timer, @Nullable Throwable t) {
+        try {
+            inferenceStats.inferenceDuration().record(timer.elapsedMillis(), responseAttributes(model, t));
+        } catch (Exception e) {
+            log.atDebug().withThrowable(e).log("Failed to record metrics with an unparsed model, dropping metrics");
+        }
+    }
+
+    private void inferOnServiceWithMetrics(
         Model model,
         InferenceAction.Request request,
         InferenceService service,
+        InferenceTimer timer,
         ActionListener<InferenceAction.Response> listener
+    ) {
+        inferenceStats.requestCount().incrementBy(1, modelAttributes(model));
+        inferOnService(model, request, service, ActionListener.wrap(inferenceResults -> {
+            if (request.isStreaming()) {
+                var taskProcessor = streamingTaskManager.<ChunkedToXContent>create(STREAMING_INFERENCE_TASK_TYPE, STREAMING_TASK_ACTION);
+                inferenceResults.publisher().subscribe(taskProcessor);
+
+                var instrumentedStream = new PublisherWithMetrics(timer, model);
+                taskProcessor.subscribe(instrumentedStream);
+
+                listener.onResponse(new InferenceAction.Response(inferenceResults, instrumentedStream));
+            } else {
+                recordMetrics(model, timer, null);
+                listener.onResponse(new InferenceAction.Response(inferenceResults));
+            }
+        }, e -> {
+            recordMetrics(model, timer, e);
+            listener.onFailure(e);
+        }));
+    }
+
+    private void recordMetrics(Model model, InferenceTimer timer, @Nullable Throwable t) {
+        try {
+            inferenceStats.inferenceDuration().record(timer.elapsedMillis(), responseAttributes(model, t));
+        } catch (Exception e) {
+            log.atDebug().withThrowable(e).log("Failed to record metrics with a parsed model, dropping metrics");
+        }
+    }
+
+    private void inferOnService(
+        Model model,
+        InferenceAction.Request request,
+        InferenceService service,
+        ActionListener<InferenceServiceResults> listener
     ) {
         if (request.isStreaming() == false || service.canStream(request.getTaskType())) {
-            inferenceStats.incrementRequestCount(model);
             service.infer(
                 model,
                 request.getQuery(),
@@ -105,7 +164,7 @@ private void inferOnService(
                 request.getTaskSettings(),
                 request.getInputType(),
                 request.getInferenceTimeout(),
-                createListener(request, listener)
+                listener
             );
         } else {
             listener.onFailure(unsupportedStreamingTaskException(request, service));
@@ -133,20 +192,6 @@ private ElasticsearchStatusException unsupportedStreamingTaskException(Inference
         }
     }
 
-    private ActionListener<InferenceServiceResults> createListener(
-        InferenceAction.Request request,
-        ActionListener<InferenceAction.Response> listener
-    ) {
-        if (request.isStreaming()) {
-            return listener.delegateFailureAndWrap((l, inferenceResults) -> {
-                var taskProcessor = streamingTaskManager.<ChunkedToXContent>create(STREAMING_INFERENCE_TASK_TYPE, STREAMING_TASK_ACTION);
-                inferenceResults.publisher().subscribe(taskProcessor);
-                l.onResponse(new InferenceAction.Response(inferenceResults, taskProcessor));
-            });
-        }
-        return listener.delegateFailureAndWrap((l, inferenceResults) -> l.onResponse(new InferenceAction.Response(inferenceResults)));
-    }
-
     private static ElasticsearchStatusException unknownServiceException(String service, String inferenceId) {
         return new ElasticsearchStatusException("Unknown service [{}] for model [{}]. ", RestStatus.BAD_REQUEST, service, inferenceId);
     }
@@ -160,4 +205,37 @@ private static ElasticsearchStatusException incompatibleTaskTypeException(TaskTy
         );
     }
 
+    private class PublisherWithMetrics extends DelegatingProcessor<ChunkedToXContent, ChunkedToXContent> {
+        private final InferenceTimer timer;
+        private final Model model;
+
+        private PublisherWithMetrics(InferenceTimer timer, Model model) {
+            this.timer = timer;
+            this.model = model;
+        }
+
+        @Override
+        protected void next(ChunkedToXContent item) {
+            downstream().onNext(item);
+        }
+
+        @Override
+        public void onError(Throwable throwable) {
+            recordMetrics(model, timer, throwable);
+            super.onError(throwable);
+        }
+
+        @Override
+        protected void onCancel() {
+            recordMetrics(model, timer, null);
+            super.onCancel();
+        }
+
+        @Override
+        public void onComplete() {
+            recordMetrics(model, timer, null);
+            super.onComplete();
+        }
+    }
+
 }

x-pack/plugin/inference/src/main/java/org/elasticsearch/xpack/inference/common/DelegatingProcessor.java

@@ -61,11 +61,14 @@ public void request(long n) {
             public void cancel() {
                 if (isClosed.compareAndSet(false, true) && upstream != null) {
                     upstream.cancel();
+                    onCancel();
                 }
             }
         };
     }
 
+    protected void onCancel() {}
+
     @Override
     public void onSubscribe(Flow.Subscription subscription) {
         if (upstream != null) {

x-pack/plugin/inference/src/main/java/org/elasticsearch/xpack/inference/telemetry/ApmInferenceStats.java

@@ -7,15 +7,87 @@
 
 package org.elasticsearch.xpack.inference.telemetry;
 
+import org.elasticsearch.ElasticsearchStatusException;
+import org.elasticsearch.core.Nullable;
 import org.elasticsearch.inference.Model;
+import org.elasticsearch.inference.UnparsedModel;
+import org.elasticsearch.telemetry.metric.LongCounter;
+import org.elasticsearch.telemetry.metric.LongHistogram;
+import org.elasticsearch.telemetry.metric.MeterRegistry;
 
-public interface InferenceStats {
+import java.util.Map;
+import java.util.Objects;
+import java.util.stream.Collectors;
+import java.util.stream.Stream;
 
-    /**
-     * Increment the counter for a particular value in a thread safe manner.
-     * @param model the model to increment request count for
-     */
-    void incrementRequestCount(Model model);
+import static java.util.Map.entry;
+import static java.util.stream.Stream.concat;
 
-    InferenceStats NOOP = model -> {};
+public record InferenceStats(LongCounter requestCount, LongHistogram inferenceDuration) {
+
+    public InferenceStats {
+        Objects.requireNonNull(requestCount);
+        Objects.requireNonNull(inferenceDuration);
+    }
+
+    public static InferenceStats create(MeterRegistry meterRegistry) {
+        return new InferenceStats(
+            meterRegistry.registerLongCounter(
+                "es.inference.requests.count.total",
+                "Inference API request counts for a particular service, task type, model ID",
+                "operations"
+            ),
+            meterRegistry.registerLongHistogram(
+                "es.inference.requests.time",
+                "Inference API request counts for a particular service, task type, model ID",
+                "ms"
+            )
+        );
+    }
+
+    public static Map<String, Object> modelAttributes(Model model) {
+        return toMap(modelAttributeEntries(model));
+    }
+
+    private static Stream<Map.Entry<String, Object>> modelAttributeEntries(Model model) {
+        var stream = Stream.<Map.Entry<String, Object>>builder()
+            .add(entry("service", model.getConfigurations().getService()))
+            .add(entry("task_type", model.getTaskType().toString()));
+        if (model.getServiceSettings().modelId() != null) {
+            stream.add(entry("model_id", model.getServiceSettings().modelId()));
+        }
+        return stream.build();
+    }
+
+    private static Map<String, Object> toMap(Stream<Map.Entry<String, Object>> stream) {
+        return stream.collect(Collectors.toMap(Map.Entry::getKey, Map.Entry::getValue));
+    }
+
+    public static Map<String, Object> responseAttributes(Model model, @Nullable Throwable t) {
+        return toMap(concat(modelAttributeEntries(model), errorAttributes(t)));
+    }
+
+    public static Map<String, Object> responseAttributes(UnparsedModel model, @Nullable Throwable t) {
+        var unknownModelAttributes = Stream.<Map.Entry<String, Object>>builder()
+            .add(entry("service", model.service()))
+            .add(entry("task_type", model.taskType().toString()))
+            .build();
+
+        return toMap(concat(unknownModelAttributes, errorAttributes(t)));
+    }
+
+    public static Map<String, Object> responseAttributes(@Nullable Throwable t) {
+        return toMap(errorAttributes(t));
+    }
+
+    private static Stream<Map.Entry<String, Object>> errorAttributes(@Nullable Throwable t) {
+        return switch (t) {
+            case null -> Stream.of(entry("status_code", 200));
+            case ElasticsearchStatusException ese -> Stream.<Map.Entry<String, Object>>builder()
+                .add(entry("status_code", ese.status().getStatus()))
+                .add(entry("error.type", String.valueOf(ese.status().getStatus())))
+                .build();
+            default -> Stream.of(entry("error.type", t.getClass().getSimpleName()));
+        };
+    }
 }