12. Metrics: Spectator, Servo, and Atlas

When used together, Spectator/Servo and Atlas provide a near real-time operational insight platform.

Spectator and Servo are Netflix’s metrics collection libraries. Atlas is a Netflix metrics backend to manage dimensional time series data.

Servo served Netflix for several years and is still usable, but is gradually being phased out in favor of Spectator, which is only designed to work with Java 8. Spring Cloud Netflix provides support for both, but Java 8 based applications are encouraged to use Spectator.

12.1 Dimensional vs. Hierarchical Metrics

Spring Boot Actuator metrics are hierarchical and metrics are separated only by name. These names often follow a naming convention that embeds key/value attribute pairs (dimensions) into the name separated by periods. Consider the following metrics for two endpoints, root and star-star:

    "counter.status.200.root": 20,
    "counter.status.400.root": 3,
    "counter.status.200.star-star": 5,

The first metric gives us a normalized count of successful requests against the root endpoint per unit of time. But what if the system had 20 endpoints and you want to get a count of successful requests against all the endpoints? Some hierarchical metrics backends would allow you to specify a wild card such as counter.status.200.* that would read all 20 metrics and aggregate the results. Alternatively, you could provide a HandlerInterceptorAdapter that intercepts and records a metric like counter.status.200.all for all successful requests irrespective of the endpoint, but now you must write 20+1 different metrics. Similarly if you want to know the total number of successful requests for all endpoints in the service, you could specify a wild card such as counter.status.2*.*.

Even in the presence of wildcarding support on a hierarchical metrics backend, naming consistency can be difficult. Specifically the position of these tags in the name string can slip with time, breaking queries. For example, suppose we add an additional dimension to the hierarchical metrics above for HTTP method. Then counter.status.200.root becomes counter.status.200.method.get.root, etc. Our counter.status.200.* suddenly no longer has the same semantic meaning. Furthermore, if the new dimension is not applied uniformly across the codebase, certain queries may become impossible. This can quickly get out of hand.

Netflix metrics are tagged (a.k.a. dimensional). Each metric has a name, but this single named metric can contain multiple statistics and 'tag' key/value pairs that allows more querying flexibility. In fact, the statistics themselves are recorded in a special tag.

Recorded with Netflix Servo or Spectator, a timer for the root endpoint described above contains 4 statistics per status code, where the count statistic is identical to Spring Boot Actuator’s counter. In the event that we have encountered an HTTP 200 and 400 thus far, there will be 8 available data points:

    "root(status=200,stastic=count)": 20,
    "root(status=200,stastic=max)": 0.7265630630000001,
    "root(status=200,stastic=totalOfSquares)": 0.04759702862580789,
    "root(status=200,stastic=totalTime)": 0.2093076914666667,
    "root(status=400,stastic=count)": 1,
    "root(status=400,stastic=max)": 0,
    "root(status=400,stastic=totalOfSquares)": 0,
    "root(status=400,stastic=totalTime)": 0,

12.2 Default Metrics Collection

Without any additional dependencies or configuration, a Spring Cloud based service will autoconfigure a Servo MonitorRegistry and begin collecting metrics on every Spring MVC request. By default, a Servo timer with the name rest will be recorded for each MVC request which is tagged with:

  1. HTTP method
  2. HTTP status (e.g. 200, 400, 500)
  3. URI (or "root" if the URI is empty), sanitized for Atlas
  4. The exception class name, if the request handler threw an exception
  5. The caller, if a request header with a key matching netflix.metrics.rest.callerHeader is set on the request. There is no default key for netflix.metrics.rest.callerHeader. You must add it to your application properties if you wish to collect caller information.

Set the netflix.metrics.rest.metricName property to change the name of the metric from rest to a name you provide.

If Spring AOP is enabled and org.aspectj:aspectjweaver is present on your runtime classpath, Spring Cloud will also collect metrics on every client call made with RestTemplate. A Servo timer with the name of restclient will be recorded for each MVC request which is tagged with:

  1. HTTP method
  2. HTTP status (e.g. 200, 400, 500), "CLIENT_ERROR" if the response returned null, or "IO_ERROR" if an IOException occurred during the execution of the RestTemplate method
  3. URI, sanitized for Atlas
  4. Client name

Avoid using hardcoded url parameters within RestTemplate. When targeting dynamic endpoints use URL variables. This will avoid potential "GC Overhead Limit Reached" issues where ServoMonitorCache treats each url as a unique key.

// recommended
String orderid = "1";
restTemplate.getForObject("http://testeurekabrixtonclient/orders/{orderid}", String.class, orderid)

// avoid
restTemplate.getForObject("http://testeurekabrixtonclient/orders/1", String.class)

12.3 Metrics Collection: Spectator

To enable Spectator metrics, include a dependency on spring-boot-starter-spectator:


In Spectator parlance, a meter is a named, typed, and tagged configuration and a metric represents the value of a given meter at a point in time. Spectator meters are created and controlled by a registry, which currently has several different implementations. Spectator provides 4 meter types: counter, timer, gauge, and distribution summary.

Spring Cloud Spectator integration configures an injectable com.netflix.spectator.api.Registry instance for you. Specifically, it configures a ServoRegistry instance in order to unify the collection of REST metrics and the exporting of metrics to the Atlas backend under a single Servo API. Practically, this means that your code may use a mixture of Servo monitors and Spectator meters and both will be scooped up by Spring Boot Actuator MetricReader instances and both will be shipped to the Atlas backend.

12.3.1 Spectator Counter

A counter is used to measure the rate at which some event is occurring.

// create a counter with a name and a set of tags
Counter counter = registry.counter("counterName", "tagKey1", "tagValue1", ...);
counter.increment(); // increment when an event occurs
counter.increment(10); // increment by a discrete amount

The counter records a single time-normalized statistic.

12.3.2 Spectator Timer

A timer is used to measure how long some event is taking. Spring Cloud automatically records timers for Spring MVC requests and conditionally RestTemplate requests, which can later be used to create dashboards for request related metrics like latency:

Figure 12.1. Request Latency


// create a timer with a name and a set of tags
Timer timer = registry.timer("timerName", "tagKey1", "tagValue1", ...);

// execute an operation and time it at the same time
T result = timer.record(() -> fooReturnsT());

// alternatively, if you must manually record the time
Long start = System.nanoTime();
T result = fooReturnsT();
timer.record(System.nanoTime() - start, TimeUnit.NANOSECONDS);

The timer simultaneously records 4 statistics: count, max, totalOfSquares, and totalTime. The count statistic will always match the single normalized value provided by a counter if you had called increment() once on the counter for each time you recorded a timing, so it is rarely necessary to count and time separately for a single operation.

For long running operations, Spectator provides a special LongTaskTimer.

12.3.3 Spectator Gauge

Gauges are used to determine some current value like the size of a queue or number of threads in a running state. Since gauges are sampled, they provide no information about how these values fluctuate between samples.

The normal use of a gauge involves registering the gauge once in initialization with an id, a reference to the object to be sampled, and a function to get or compute a numeric value based on the object. The reference to the object is passed in separately and the Spectator registry will keep a weak reference to the object. If the object is garbage collected, then Spectator will automatically drop the registration. See the note in Spectator’s documentation about potential memory leaks if this API is misused.

// the registry will automatically sample this gauge periodically
registry.gauge("gaugeName", pool, Pool::numberOfRunningThreads);

// manually sample a value in code at periodic intervals -- last resort!
registry.gauge("gaugeName", Arrays.asList("tagKey1", "tagValue1", ...), 1000);

12.3.4 Spectator Distribution Summaries

A distribution summary is used to track the distribution of events. It is similar to a timer, but more general in that the size does not have to be a period of time. For example, a distribution summary could be used to measure the payload sizes of requests hitting a server.

// the registry will automatically sample this gauge periodically
DistributionSummary ds = registry.distributionSummary("dsName", "tagKey1", "tagValue1", ...);

12.4 Metrics Collection: Servo


If your code is compiled on Java 8, please use Spectator instead of Servo as Spectator is destined to replace Servo entirely in the long term.

In Servo parlance, a monitor is a named, typed, and tagged configuration and a metric represents the value of a given monitor at a point in time. Servo monitors are logically equivalent to Spectator meters. Servo monitors are created and controlled by a MonitorRegistry. In spite of the above warning, Servo does have a wider array of monitor options than Spectator has meters.

Spring Cloud integration configures an injectable com.netflix.servo.MonitorRegistry instance for you. Once you have created the appropriate Monitor type in Servo, the process of recording data is wholly similar to Spectator.

12.4.1 Creating Servo Monitors

If you are using the Servo MonitorRegistry instance provided by Spring Cloud (specifically, an instance of DefaultMonitorRegistry), Servo provides convenience classes for retrieving counters and timers. These convenience classes ensure that only one Monitor is registered for each unique combination of name and tags.

To manually create a Monitor type in Servo, especially for the more exotic monitor types for which convenience methods are not provided, instantiate the appropriate type by providing a MonitorConfig instance:

MonitorConfig config = MonitorConfig.builder("timerName").withTag("tagKey1", "tagValue1").build();

// somewhere we should cache this Monitor by MonitorConfig
Timer timer = new BasicTimer(config);

12.5 Metrics Backend: Atlas

Atlas was developed by Netflix to manage dimensional time series data for near real-time operational insight. Atlas features in-memory data storage, allowing it to gather and report very large numbers of metrics, very quickly.

Atlas captures operational intelligence. Whereas business intelligence is data gathered for analyzing trends over time, operational intelligence provides a picture of what is currently happening within a system.

Spring Cloud provides a spring-cloud-starter-atlas that has all the dependencies you need. Then just annotate your Spring Boot application with @EnableAtlas and provide a location for your running Atlas server with the netflix.atlas.uri property.

12.5.1 Global tags

Spring Cloud enables you to add tags to every metric sent to the Atlas backend. Global tags can be used to separate metrics by application name, environment, region, etc.

Each bean implementing AtlasTagProvider will contribute to the global tag list:

AtlasTagProvider atlasCommonTags(
    @Value("${spring.application.name}") String appName) {
  return () -> Collections.singletonMap("app", appName);

12.5.2 Using Atlas

To bootstrap a in-memory standalone Atlas instance:

$ curl -LO https://github.com/Netflix/atlas/releases/download/v1.4.2/atlas-1.4.2-standalone.jar
$ java -jar atlas-1.4.2-standalone.jar

An Atlas standalone node running on an r3.2xlarge (61GB RAM) can handle roughly 2 million metrics per minute for a given 6 hour window.

Once running and you have collected a handful of metrics, verify that your setup is correct by listing tags on the Atlas server:

$ curl http://ATLAS/api/v1/tags

After executing several requests against your service, you can gather some very basic information on the request latency of every request by pasting the following url in your browser: http://ATLAS/api/v1/graph?q=name,rest,:eq,:avg

The Atlas wiki contains a compilation of sample queries for various scenarios.

Make sure to check out the alerting philosophy and docs on using double exponential smoothing to generate dynamic alert thresholds.

12.6 Retrying Failed Requests

Spring Cloud Netflix offers a variety of ways to make HTTP requests. You can use a load balanced RestTemplate, Ribbon, or Feign. No matter how you choose to your HTTP requests, there is always a chance the request may fail. When a request fails you may want to have the request retried automatically. To accomplish this when using Sping Cloud Netflix you need to include Spring Retry on your application’s classpath. When Spring Retry is present load balanced RestTemplates, Feign, and Zuul will automatically retry any failed requests (assuming you configuration allows it to).

12.6.1 Configuration

Anytime Ribbon is used with Spring Retry you can control the retry functionality by configuring certain Ribbon properties. The properties you can use are client.ribbon.MaxAutoRetries, client.ribbon.MaxAutoRetriesNextServer, and client.ribbon.OkToRetryOnAllOperations. See the Ribbon documentation for a description of what there properties do.

In addition you may want to retry requests when certain status codes are returned in the response. You can list the response codes you would like the Ribbon client to retry using the property clientName.ribbon.retryableStatusCodes. For example

    retryableStatusCodes: 404,502

You can also create a bean of type LoadBalancedRetryPolicy and implement the retryableStatusCode method to determine whether you want to retry a request given the status code.

12.6.2 Zuul

You can turn off Zuul’s retry functionality by setting zuul.retryable to false. You can also disable retry functionality on route by route basis by setting zuul.routes.routename.retryable to false.