This project contains modifications to the dropwizard metrics project.

The primary additions includes a replacement for MetricRegistry allowing for metric names containing tags through MetricId.

The following are the interfaces and classes that has to be used from this package in order for MetricId to be used.

You will find these types in com.spotify.metrics.core.

• SemanticMetricRegistry — Replacement for MetricRegistry.
• MetricId — Replacement for string-based metric names.
• SemanticMetricFilter — Replacement for MetricFilter.
• SemanticMetricRegistryListener — Replacement for MetricRegistryListener.
• SemanticMetricSet — Replacement for MetricSet.

Care must be taken not to use the upstream MetricRegistry because it does not support the use of MetricId. To ease this, all of the replacing classes follow the Semantic* naming convention.

As an effect of this, pre-existing plugins for codahale metrics will not work.

Add a dependency in maven.

<dependency>
  <groupId>com.spotify.metrics</groupId>
  <artifactId>semantic-metrics-core</artifactId>
  <version>${semantic-metrics.version}</version>
</dependency>

This project provide the following set of plugins;

• com.spotify.metrics.ffwd A reporter into FastForward.
• com.spotify.metrics.jvm Ported MetricSet's for internal java statistics.

See and run examples.

If you find yourself in a situation where you create many instances of this class (i.e. when reporting metrics) and profiling/benchmarks show a significant amount of time spent constructing MetricId instances, considering making use of a MetricIdCache

The following is an example integrating with Guava.

// GuavaCache.java

public final class GuavaCache<T> implements MetricIdCache.Cache<T> {
    final Cache<T, MetricId> cache = CacheBuilder.newBuilder().expireAfterAccess(6, TimeUnit.HOURS)
            .build();

    private final MetricIdCache.Loader<T> loader;

    public GuavaCache(Loader<T> loader) {
        this.loader = loader;
    }

    @Override
    public MetricId get(final MetricId base, final T key) throws ExecutionException {
        return cache.get(key, new Callable<MetricId>() {
            @Override
            public MetricId call() throws Exception {
                return loader.load(base, key);
            }
        });
    }

    @Override
    public void invalidate(T key) {
        cache.invalidate(key);
    }

    @Override
    public void invalidateAll() {
        cache.invalidateAll();
    }

    public static MetricIdCache.Any setup() {
        return MetricIdCache.builder().cacheBuilder(new MetricIdCache.CacheBuilder() {
            @Override
            public <T> MetricIdCache.Cache<T> build(final Loader<T> loader) {
                return new GuavaCache<T>(loader);
            }
        });
    }
}

// MyApplicationStatistics.java

public class MyApplicationStatistics() {
    private final MetricIdCache.Typed<String> endpoint = GuavaCache.setup()
        .loader(new MetricIdCache.Loader<String>() {
            @Override
            public MetricId load(MetricId base, String endpoint) {
                return base.tagged("endpoint", endpoint);
            }
        });

    private final MetricIdCache<String> requests = endpoint
        .metricId(MetricId.build().tagged("what", "endpoint-requests", "unit", "request"))
        .build();

    private final MetricIdCache<String> errors = endpoint
        .metricId(MetricId.build().tagged("what", "endpoint-errors", "unit", "error"))
        .build();

    private final SemanticMetricRegistry registry;

    public MyApplicationStatistics(SemanticMetricRegistry registry) {
        this.registry = registry;
    }

    public void reportRequest(String endpoint) {
        registry.meter(requests.get(endpoint)).mark();
    }

    public void reportError(String endpoint) {
        registry.meter(errors.get(endpoint)).mark();
    }
}

Avoid performing deep integration of semantic-metrics into your library or application. This will prevent you, and third parties, from integrating your code with different metric collectors.

As an alternative you should build a tree of interfaces that your application uses to report metrics (e.g. my-service-statistics), and use these to build an implementation using semantic metrics (my-service-semantic-statistics).

This pattern greatly simplifies integrating your application with more than one metric collector, or ditching semantic-metrics when it becomes superseded by something better.

At configuration time your application can decide which implementation to use by simply providing an instance of the statistics API which suits their requirements.

Build an interface describing all the things that your application reports.

public interface MyApplicationStatistics {
    /**
     * Report that a single request has been received by the application.
     */
    void reportRequest();
}

Provide a semantic-metrics implementation.

public class SemanticMyApplicationStatistics implements MyApplicationStatistics {
    private final SemanticMetricRegistry registry;

    private final Meter request;

    public SemanticMyApplicationStatistics(SemanticMetricRegistry registry) {
        this.registry = registry;
        this.request = registry.meter(MetricId.build().tagged(
            "what", "requests", "unit", "request"));
    }

    @Override
    public void reportRequest() {
        request.mark();
    }
}

Now a user of your framework/application can do something like the following to bootstrap your application.

public class Entry {
    public static void main(String[] argv) {
        final SemanticMetricRegistry registry = new SemanticMetricRegistry();
        final MyApplicationStatistics statistics = new SemanticMyApplicationStatistics(registry);
        /* your application */
        final MyApplication app = MyApplication.builder().statistics(statistics).build();

        final FastForwardReporter reporter = FastForwardReporter.forRegistry(registry).build()

        reporter.start();
        app.start();

        app.join();
        reporter.stopWithFlush();
        System.exit(0);
    }
}

There are different metric types that can be used depending on what it is that we want to measure, e.g., queue length, or request time, etc.

A gauge is an instantaneous measurement of a value. For example if we want to measure the number of pending jobs in a queue.

registry.register(metric.tagged("what", "job-queue-length"), new Gauge<Integer>() {
    @Override
    public Integer getValue() {
        // fetch the queue length the way you like
        final int queueLength = 10;
        // obviously this is gonna keep reporting 10, but you know ;)
        return queueLength;
    }
});

In addition to the tags that are specified (e.g., "what" in this example), FfwdReporter adds the following tags to each Gauge data point:

A counter is just a gauge for an AtomicLong instance. You can increment or decrement its value.

For example we want a more efficient way of measuring the pending job in a queue.

final Counter counter = registry.counter(metric.tagged("what", "job-count"));
// Somewhere in your code where you are adding new jobs to the queue you increment the counter as well
counter.inc();
// Somewhere in your code the job is going to be removed from the queue you decrement the counter
counter.dec();

In addition to the tags that are specified (e.g., "what" in this example), FfwdReporter adds the following tags to each Counter data point:

A meter measures the rate of events over time (e.g., "requests per second"). In addition to the mean rate, meters also track 1- and 5-minute moving averages.

For example we have an endpoint that we want to measure how frequent we receive requests for it.

Meter meter = registry.meter(metric.tagged("what", "incoming-requests").tagged("endpoint", "/v1/list"));
// Now a request comes and it's time to mark the meter
meter.mark();

In addition to the tags that are specified (e.g., "what" and "endpoint" in this example), FfwdReporter adds the following tags to each Meter data point:

NOTE: Meter also reports the meter counter value to allow platforms to derive rates using the monotonically increasing count instead of only aggregating the rate computed by the meter itself. It is useful for applications to be able to report both count and rate using a meter.

A deriving meter takes the derivative of a value that is expected to be monotonically increasing.

A typical use case is to get the rate of change of a counter of the total number of events.

This implementation ignores updates that decrease the counter value. The rationale is that the counter is expected to be monotonically increasing between infrequent resets (when a process has been restarted, for example). Thus, negative values should only happen on restart and should be safe to discard.

DerivingMeter derivingMeter = registry.derivingMeter(metric.tagged("what", "incoming-requests").tagged("endpoint", "/v1/list"));
derivingMeter.mark();

In addition to the tags that are specified (e.g., "what" and "endpoint" in this example), FfwdReporter adds the following tags to each Meter data point:

A histogram measures the statistical distribution of values in a stream of data. It measures minimum, maximum, mean, median, standard deviation, as well as 75th and 99th percentiles.

For example this histogram will measure the size of responses in bytes.

Histogram histogram = registry.histogram(metric.tagged("what", "response-size").tagged("endpoint", "/v1/content"));
// fetch the size of the response
final long responseSize = getResponseSize(response);
histogram.update(responseSize);

In addition to the tags that are specified (e.g., "what" and "endpoint" in this example), FfwdReporter adds the following tags to each Histogram data point:

Note that added custom percentiles will show up in the stat tag.

HistogramWithTtl changes the behavior of the default codahale histogram when update rate is low. If the update rate goes below a certain threshold for a certain time, all samples that have been received during that time are used instead of the random sample that is used in the default histogram implementation. When update rates are above the threshold, the default implementation is used.

What problem does it solve?

The default histogram implementation uses a random sampling algorithm with exponentially decaying probabilities over time. This works well if update rates are approximately 10 requests per second or above. When rates go below that, the metrics, especially p99 and above tends to flatline because the values are not replaced often enough. We solve this by using a different implementation whenever the update rate goes below 10 RPS. This gives much more dynamic percentile measurements during low update rates. When update rates go above the threshold we switch to the default implementation.

This was authored by Johan Buratti.

Distributions are no longer supported. The code to create them and the Heroic code to query them still exists, however they are being retired and no further adoption should occur.

Heroic is being retired in favor of OpenSource alternatives, and this distribution implementation will not be portable to the future TSDB/query interface. Since only a few services with a few metrics had experimented with distributions, the choice was made to halt adoption now, to reduce the pain of conversion to a proper histrogram later.

For historical refrence only

DO NOT USE

Distribution is a simple interface that allows users to record measurements to compute rank statistics on data distribution, not just a local source.

Every implementation should produce a serialized data sketch in a byteBuffer as this metric point value.

Unlike traditional histograms, distribution doesn't require a predefined percentile value. Data recorded can be used upstream to compute any percentile.

Distribution doesn't require any binning configuration. Just get an instance through SemanticMetricBuilder and record data.

Distribution is a good choice if you care about percentile accuracy in a distributed environment and you want to rely on P99 to set SLOs.

For example, this distribution will measure the size of messages in bytes.

Distribution distribution = registry.distribution(metric.tagged("what", "distribution-message-size", "unit", Units.BYTE));
// fetch the size of the message
int size = getMessageSize(response);
distribution.record(size);

In addition to the tags that are specified (e.g., "what" and "unit" in this example), FfwdReporter adds the following tags to each Histogram data point:

What problem does it solve?

• Accurate Aggregated Histogram Data

This can record data and send data sketches. A sketch of a dataset is a small data structure that lets you approximate certain characteristics of the original dataset. Sketches are used to compute rank based statistics such as percentile. Sketches are mergeable and can be used to compute any percentile on the entire data distribution.

• Support Sophisticated Data-point Values

With distributions we are able to support sophisticated data point values, such as the Open-census metric distribution.

Authored by Adele Okoubo.

A timer measures both the rate that a particular piece of code is called and the distribution of its duration.

For example we want to measure the rate and handling duration of incoming requests.

Timer timer = registry.timer(metric.tagged("what", "incoming-request-time").tagged("endpoint", "/v1/get_stuff"));
// Do this before starting to do the thing. This creates a measurement context object that you can pass around.
final Context context = timer.time();
doStuff();
// Tell the context that it's done. This will register the duration and counts one occurrence.
context.stop();

In addition to the tags that are specified (e.g., "what" and "endpoint" in this example), FfwdReporter adds the following tags to each Timer data point:

NOTE: Timer is really just a combination of a Histogram and a Meter, so apart from the tags above, combination of both Histogram and Meter tags will be included.

When dealing with thousands of similar timeseries over thousands of hosts, classification becomes a big issue.

Classical systems organize metric names as strings, containing a lot of information about the metric in question.

You will often see things like webserver.host.example.com.df.used./.

The same metric expressed as a set of tags could look like.

{"role": "webserver", "host": "host.example.com", "what": "disk-used",
 "mountpoint": "/"}

This system of classification from the host machine greatly simplifies any metrics pipeline. When transported with a stable serialization method (like JSON) it does not matter if we add additional tags, or decide to change the order in which the timeseries happens to be designated.

We can also easily index this timeseries by its tag using a system like ElasticSearch and ask it interesting questions about which timeseries are available.

If used with a metrics backend that supports efficient aggregation and filtering across tags you gain a flexible and intionistic pipeline that is powerful and agnostic about what it sends, all the way from the service being monitored to your metrics GUI.

This project adheres to the Open Code of Conduct. By participating, you are expected to honor this code.

1. Fork semantic-metrics from github and clone your fork.
2. Hack.
3. Push the branch back to GitHub.
4. Send a pull request to our upstream repo.

Releasing is done via the maven-release-plugin and nexus-staging-plugin which are configured via the release profile. Deploys are staged in oss.sonatype.org before being deployed to Maven Central. Check out the maven-release-plugin docs and the nexus-staging-plugin docs for more information.

To release, first run:

mvn -P release release:prepare

You will be prompted for the release version and the next development version. On success, follow with:

mvn -P release release:perform

When you have finished these steps, please "Draft a new release" in Github and list the included PRs (aside from changes to documentation).