Add prometheus metrics to serving

infra/charts/feast/charts/feast-serving/templates/deployment.yaml

@@ -9,6 +9,13 @@ metadata:
     chart: {{ .Chart.Name }}-{{ .Chart.Version | replace "+" "_" }}
     release: {{ .Release.Name }}
     heritage: {{ .Release.Service }}
+  annotations:
+    {{- if .Values.prometheus.enabled }}
+    {{ $config := index .Values "application.yaml" }}
+    prometheus.io/path: /metrics
+    prometheus.io/port: "{{ $config.server.port }}"
+    prometheus.io/scrape: "true"
+    {{- end }}
 spec:
   replicas: {{ .Values.replicaCount }}
   selector:

infra/charts/feast/charts/feast-serving/values.yaml

@@ -63,6 +63,8 @@ application.yaml:
   grpc:
     port: 6566
     enable-reflection: true
+  server:
+    port: 8080
   spring:
     main:
       web-application-type: none
@@ -177,6 +179,9 @@ ingress:
   # - host: chart-example.local
   #   port: http
 
+prometheus:
+  enabled: true
+
 resources: {}
   # We usually recommend not to specify default resources and to leave this as a conscious
   # choice for the user. This also increases chances charts run on environments with little

serving/pom.xml

@@ -168,6 +168,30 @@
       <version>0.31.0</version>
     </dependency>
 
+    <!-- The client -->
+    <dependency>
+      <groupId>io.prometheus</groupId>
+      <artifactId>simpleclient</artifactId>
+      <version>0.8.0</version>
+    </dependency>
+    <!-- Hotspot JVM metrics-->
+    <dependency>
+      <groupId>io.prometheus</groupId>
+      <artifactId>simpleclient_hotspot</artifactId>
+      <version>0.8.0</version>
+    </dependency>
+    <!-- Exposition HTTPServer-->
+    <dependency>
+      <groupId>io.prometheus</groupId>
+      <artifactId>simpleclient_servlet</artifactId>
+      <version>0.8.0</version>
+    </dependency>
+    <dependency>
+      <groupId>io.prometheus</groupId>
+      <artifactId>simpleclient_spring_boot</artifactId>
+      <version>0.8.0</version>
+    </dependency>
+
     <!-- Google Cloud -->
     <dependency>
       <groupId>com.google.cloud</groupId>

serving/src/main/java/feast/serving/configuration/InstrumentationConfig.java

@@ -3,19 +3,29 @@
 import feast.serving.FeastProperties;
 import io.opentracing.Tracer;
 import io.opentracing.noop.NoopTracerFactory;
+import io.prometheus.client.hotspot.DefaultExports;
+import io.prometheus.client.exporter.MetricsServlet;
 import org.springframework.beans.factory.annotation.Autowired;
+import org.springframework.boot.web.servlet.ServletRegistrationBean;
 import org.springframework.context.annotation.Bean;
 import org.springframework.context.annotation.Configuration;
 
 @Configuration
 public class InstrumentationConfig {
+
   private FeastProperties feastProperties;
 
   @Autowired
   public InstrumentationConfig(FeastProperties feastProperties) {
     this.feastProperties = feastProperties;
   }
 
+  @Bean
+  public ServletRegistrationBean servletRegistrationBean() {
+    DefaultExports.initialize();
+    return new ServletRegistrationBean(new MetricsServlet(), "/metrics");
+  }
+
   @Bean
   public Tracer tracer() {
     if (!feastProperties.getTracing().isEnabled()) {

serving/src/main/java/feast/serving/service/BigQueryServingService.java

@@ -1,6 +1,8 @@
 package feast.serving.service;
 
 import static feast.serving.util.BigQueryUtil.getTimestampLimitQuery;
+import static feast.serving.util.Metrics.requestCount;
+import static feast.serving.util.Metrics.requestLatency;
 
 import com.google.cloud.bigquery.BigQuery;
 import com.google.cloud.bigquery.BigQueryException;
@@ -37,6 +39,7 @@
 import feast.serving.ServingAPIProto.JobType;
 import feast.serving.util.BigQueryUtil;
 import io.grpc.Status;
+import io.prometheus.client.Histogram.Timer;
 import java.io.IOException;
 import java.util.ArrayList;
 import java.util.List;
@@ -100,11 +103,13 @@ public GetOnlineFeaturesResponse getOnlineFeatures(GetOnlineFeaturesRequest getF
    */
   @Override
   public GetBatchFeaturesResponse getBatchFeatures(GetBatchFeaturesRequest getFeaturesRequest) {
-
+    Timer getBatchFeaturesTimer = requestLatency.labels("getBatchFeatures").startTimer();
     List<FeatureSetSpec> featureSetSpecs =
         getFeaturesRequest.getFeatureSetsList().stream()
-            .map(featureSet ->
-                specService.getFeatureSet(featureSet.getName(), featureSet.getVersion())
+            .map(featureSet -> {
+                  requestCount.labels(featureSet.getName()).inc();
+                  return specService.getFeatureSet(featureSet.getName(), featureSet.getVersion());
+                }
             )
             .collect(Collectors.toList());
 
@@ -233,6 +238,7 @@ public GetBatchFeaturesResponse getBatchFeatures(GetBatchFeaturesRequest getFeat
         })
         .start();
 
+    getBatchFeaturesTimer.observeDuration();
     return GetBatchFeaturesResponse.newBuilder().setJob(feastJob).build();
   }
 

serving/src/main/java/feast/serving/service/CachedSpecService.java

@@ -16,6 +16,7 @@
 import feast.core.StoreProto.Store.Subscription;
 import feast.serving.exception.SpecRetrievalException;
 import io.grpc.StatusRuntimeException;
+import io.prometheus.client.Gauge;
 import java.io.IOException;
 import java.nio.file.Files;
 import java.nio.file.Path;
@@ -40,6 +41,15 @@ public class CachedSpecService {
   private final LoadingCache<String, FeatureSetSpec> featureSetSpecCache;
   private Store store;
 
+  private static Gauge featureSetsCount = Gauge.build().name("feature_set_count")
+      .subsystem("feast_serving")
+      .help("number of feature sets served by this instance")
+      .register();
+  private static Gauge cacheLastUpdated = Gauge.build().name("cache_last_updated")
+      .subsystem("feast_serving")
+      .help("epoch time of the last time the cache was updated")
+      .register();
+
   public CachedSpecService(CoreSpecService coreService, Path configPath) {
     this.configPath = configPath;
     this.coreService = coreService;
@@ -102,6 +112,9 @@ public void populateCache() {
     this.store = updateStore(readConfig(configPath));
     Map<String, FeatureSetSpec> featureSetSpecMap = getFeatureSetSpecMap();
     featureSetSpecCache.putAll(featureSetSpecMap);
+
+    featureSetsCount.set(featureSetSpecCache.size());
+    cacheLastUpdated.set(System.currentTimeMillis());
   }
 
   public void scheduledPopulateCache() {

serving/src/main/java/feast/serving/service/RedisServingService.java

@@ -16,6 +16,11 @@
 
 package feast.serving.service;
 
+import static feast.serving.util.Metrics.missingKeyCount;
+import static feast.serving.util.Metrics.requestLatency;
+import static feast.serving.util.Metrics.requestCount;
+import static feast.serving.util.Metrics.staleKeyCount;
+
 import com.google.common.collect.Maps;
 import com.google.protobuf.AbstractMessageLite;
 import com.google.protobuf.Duration;
@@ -41,6 +46,7 @@
 import io.grpc.Status;
 import io.opentracing.Scope;
 import io.opentracing.Tracer;
+import io.prometheus.client.Histogram.Timer;
 import java.util.List;
 import java.util.Map;
 import java.util.stream.Collectors;
@@ -61,7 +67,9 @@ public RedisServingService(JedisPool jedisPool, CachedSpecService specService, T
     this.tracer = tracer;
   }
 
-  /** {@inheritDoc} */
+  /**
+   * {@inheritDoc}
+   */
   @Override
   public GetFeastServingInfoResponse getFeastServingInfo(
       GetFeastServingInfoRequest getFeastServingInfoRequest) {
@@ -70,10 +78,13 @@ public GetFeastServingInfoResponse getFeastServingInfo(
         .build();
   }
 
-  /** {@inheritDoc} */
+  /**
+   * {@inheritDoc}
+   */
   @Override
   public GetOnlineFeaturesResponse getOnlineFeatures(GetOnlineFeaturesRequest request) {
     try (Scope scope = tracer.buildSpan("Redis-getOnlineFeatures").startActive(true)) {
+      Timer getOnlineFeaturesTimer = requestLatency.labels("getOnlineFeatures").startTimer();
       GetOnlineFeaturesResponse.Builder getOnlineFeaturesResponseBuilder =
           GetOnlineFeaturesResponse.newBuilder();
 
@@ -114,6 +125,7 @@ public GetOnlineFeaturesResponse getOnlineFeatures(GetOnlineFeaturesRequest requ
           featureValuesMap.values().stream()
               .map(m -> FieldValues.newBuilder().putAllFields(m).build())
               .collect(Collectors.toList());
+      getOnlineFeaturesTimer.observeDuration();
       return getOnlineFeaturesResponseBuilder.addAllFieldValues(fieldValues).build();
     }
   }
@@ -166,9 +178,11 @@ private RedisKey makeRedisKey(
     for (int i = 0; i < featureSetEntityNames.size(); i++) {
       String entityName = featureSetEntityNames.get(i);
 
-      if (!fieldsMap.containsKey(entityName)){
+      if (!fieldsMap.containsKey(entityName)) {
         throw Status.INVALID_ARGUMENT
-            .withDescription(String.format("Entity row fields \"%s\" does not contain required entity field \"%s\"", fieldsMap.keySet().toString(), entityName))
+            .withDescription(String
+                .format("Entity row fields \"%s\" does not contain required entity field \"%s\"",
+                    fieldsMap.keySet().toString(), entityName))
             .asRuntimeException();
       }
 
@@ -186,33 +200,45 @@ private void sendAndProcessMultiGet(
       throws InvalidProtocolBufferException {
 
     List<byte[]> jedisResps = sendMultiGet(redisKeys);
-
+    Timer processResponseTimer = requestLatency.labels("processResponse")
+        .startTimer();
     try (Scope scope = tracer.buildSpan("Redis-processResponse").startActive(true)) {
       String featureSetId =
           String.format("%s:%d", featureSetRequest.getName(), featureSetRequest.getVersion());
+
       Map<String, Value> nullValues =
           featureSetRequest.getFeatureNamesList().stream()
               .collect(
                   Collectors.toMap(
                       name -> featureSetId + ":" + name, name -> Value.newBuilder().build()));
+
       for (int i = 0; i < jedisResps.size(); i++) {
         EntityRow entityRow = entityRows.get(i);
         Map<String, Value> featureValues = featureValuesMap.get(entityRow);
+
         byte[] jedisResponse = jedisResps.get(i);
         if (jedisResponse == null) {
+          missingKeyCount.labels(featureSetRequest.getName()).inc();
           featureValues.putAll(nullValues);
           continue;
         }
+
         FeatureRow featureRow = FeatureRow.parseFrom(jedisResponse);
+
         boolean stale = isStale(featureSetRequest, entityRow, featureRow);
         if (stale) {
+          staleKeyCount.labels(featureSetRequest.getName()).inc();
           featureValues.putAll(nullValues);
           continue;
         }
+
+        requestCount.labels(featureSetRequest.getName()).inc();
         featureRow.getFieldsList().stream()
             .filter(f -> featureSetRequest.getFeatureNamesList().contains(f.getName()))
             .forEach(f -> featureValues.put(featureSetId + ":" + f.getName(), f.getValue()));
       }
+    } finally {
+      processResponseTimer.observeDuration();
     }
   }
 
@@ -237,6 +263,7 @@ private boolean isStale(
    */
   private List<byte[]> sendMultiGet(List<RedisKey> keys) {
     try (Scope scope = tracer.buildSpan("Redis-sendMultiGet").startActive(true)) {
+      Timer sendMultiGetTimer = requestLatency.labels("sendMultiGet").startTimer();
       try (Jedis jedis = jedisPool.getResource()) {
         byte[][] binaryKeys =
             keys.stream()
@@ -249,6 +276,8 @@ private List<byte[]> sendMultiGet(List<RedisKey> keys) {
             .withDescription("Unable to retrieve feature from Redis")
             .withCause(e)
             .asRuntimeException();
+      } finally {
+        sendMultiGetTimer.observeDuration();
       }
     }
   }

serving/src/main/java/feast/serving/util/Metrics.java

@@ -0,0 +1,37 @@
+package feast.serving.util;
+
+import io.prometheus.client.Counter;
+import io.prometheus.client.Histogram;
+import io.prometheus.client.Summary;
+
+public class Metrics {
+
+  public static final Histogram requestLatency = Histogram.build()
+      .buckets(2, 4, 6, 8, 10, 15, 20, 25, 30, 35, 50)
+      .name("request_latency_ms")
+      .subsystem("feast_serving")
+      .help("Request latency in milliseconds.")
+      .labelNames("method")
+      .register();
+
+  public static final Counter requestCount = Counter.build()
+      .name("request_feature_count")
+      .subsystem("feast_serving")
+      .help("number of feature rows requested")
+      .labelNames("feature_set_name")
+      .register();
+
+  public static final Counter missingKeyCount = Counter.build()
+      .name("missing_feature_count")
+      .subsystem("feast_serving")
+      .help("number requested feature rows that were not found")
+      .labelNames("feature_set_name")
+      .register();
+
+  public static final Counter staleKeyCount = Counter.build()
+      .name("stale_feature_count")
+      .subsystem("feast_serving")
+      .help("number requested feature rows that were stale")
+      .labelNames("feature_set_name")
+      .register();
+}

serving/src/main/resources/application.yml

@@ -55,6 +55,6 @@ grpc:
 
 server:
   # The port number on which the Tomcat webserver that serves REST API endpoints should listen
-  # It is set by default to 8080 so it does not conflict with Tomcat webserver on Feast Core
+  # It is set by default to 8081 so it does not conflict with Tomcat webserver on Feast Core
   # if both Feast Core and Serving are running on the same machine
   port: ${SERVER_PORT:8081}