Metrics

Nowadays, there are a lot of complex distributed systems, and every day their number is increasing. When developing such systems, it is important to strive for the highest possible performance. But how can we assess the performance of a system? There are special metrics for this.

Metrics are numerical data that are calculated or aggregated over a period of time. Metrics give us an idea of the current and historical state of the system. They can be used for statistical analysis, to predict the future behavior of the system.

There are many types of metrics that can be used for monitoring the performance of distributed systems. For example, latency, bandwidth, error, saturation, traffic, etc. For each distributed system, we need to choose a list of metrics that suit it exactly.

For convenient graphical display and analysis of metrics, we should use monitoring systems such as Grafana, Prometheus. These systems allow us to create dashboards and display specific metrics over a given period of time. For each distributed system, it is necessary to be able to store metrics into monitoring systems for their further analysis.

Ubiquitous Metrics - The metrics abstraction used by Proto.Actor

Abstractions for metrics with implementation for APM providers.

NuGet packages

Package	What for
`Ubiquitous.Metrics`	Core library
`Ubiquitous.Metrics.Prometheus`	Prometheus support
`Ubiquitous.Metrics.Dogstatsd`	Datadog StatsD support
`Ubiquitous.Metrics.MicrosoftLog`	Log measurements

Why?

Metrics in OpenTelemetry for .NET is still out of sight, and we have a need to expose metrics to different vendors.

In order to avoid rewriting the whole measurement when switching from one vendor to another, we decided to create a small abstraction layer, which exposes similar metrics in a uniform fashion.

Then, we can configure a specific provider to expose the measurements properly. By default, measurements will still be performed, even when no provider is configures, using the no-op provider. It works great for tests, where you won’t need to care about configuring, for example, Prometheus, to prevent the test from crashing.

It should be trivial also to create a provider for a benchmarking tool and collect performance evidence when running the tests.

Essentials

The library provides three metric types:

Counter
Gauge
Histogram

Having histograms support essentially limits the number of providers we could support, but we accepted this trade-off due to the high value of histogram metrics.

Getting started

In your application, you start by creating an instance of the Metrics class with one or more configuration providers. Configuration providers implement specifics for a given observability product (Prometheus, Datadog, etc).

You might want to isolate the initial but to a separate function, which you then call from your Startup:

public static class Measurements {
    public static void ConfigureMetrics(string environment) {
        var metrics = Metrics.CreateUsing(
            new PrometheusConfigurator(
                new Label("app", "myAwesomeApp"),
                new Label("environment", environment)
            )
        );
        MyAwesomeAppMetrics.Configure(metrics);
    }
}

In the snipped above, we added the default labels app and environment, which will be added to all the measurements implicitly.

Then, in the MyAwesomeAppMetrics.Configure function, you need to configure all the metrics needed for your application:

public static class MyAwesomeMetrics {
    public static void Configure(Metrics metrics) {
        CurrentUsers    = metrics.CreateGauge("current_users", "The number of users on the site");
        HttpApiErrors   = metrics.CreateCount("api_errors", "Number of requests, which failed");
        HttpApiResponse = metrics.CreateHistogram("http_response_time", "HTTP request processing time");
    }

    public static IGaugeMetric CurrentUsers { get; private set; } = null!;

    public static IHistogramMetric HttpApiResponse { get; private set; } = null!;

    public static ICountMetric HttpApiErrors { get; private set; } = null!;
}

As you probably want to see measurements per resource, you can add labels, identifying the resource, which is being observed:

HttpApiResponse = metrics.CreateHistogram(
    "http_response_time",
    "HTTP request processing time",
    "http_resource",
    "http_method"
);

Then, when you observe the execution time, you need to supply the label value:

app.Use((context, next) => Metrics.Measure(next,
    MyAwesomeMetrics.HttpApiResponse,
    MyAwesomeMetrics.HttpApiErrors,
    new [] {
        context.Request.Path.Value,
        context.Request.Method
    })
);

In this simple HTTP middleware, we use the Measure function, which calls the specified action, wrapped in the time measurement. The metric histogram will get the observation, where the time is measured in seconds.

Benefits

Tests

In your application, you’d need to configure the metrics instance with a proper provider, so you get your application properly measured.

However, you might want to avoid this in your test project. There, you can use the NoMetrics configuration provider. When instantiating the Metrics instance without any provider supplied, the NoMetrics provider will be used.

The Metrics.Instance static member will implicitly create an instance of the Metrics class with NoMetrics provider.

public class MyTestFixture {
    static MyTextFixture() => MyAwesomeAppMetrics.Configure(Metrics.Instance);

    // here are my other setups
}

Environment-specific

Another scenario would be that your organisation is using Datadog APM to measure apps in production, but when running locally, you’d still like to measure and you don’t have the Datadog agent running on your dev machine.

In this case, you can run Prometheus and Grafana locally in Docker and use different configuration providers, based on the environment name.

public static class Measurements {
    public static void ConfigureMetrics(string environment) {
        IMetricsProvider configProvider = environment == "Development"
            ? new PrometheusConfigurator()
            : new StatsdConfigurator(new [] {
                  new Label("app", "myAwesomeApp"),
                  new Label("environment", environment)
              });
        );
        MyAwesomeAppMetrics.Configure(Metrics.CreateUsing(configProvider));
    }
}

Libraries

It’s probably not the best idea for a library author to couple measurements to a specific product.

Using this library, you can be vendor-agnostic when collecting the metrics and give your users the ability to use the provider they want.

Vendors support

Currently, the library supports exposing metrics for:

Prometheus (using prometheus-net)
Datadog Statsd (using Dogstatsd)

In addition, we provide a few “fake” configurators, which either do nothing, or allow you to look at the measurements locally.

NoMetrics: very fast, good for libraries as it’s the default one
Microsoft Logging: logs measurements using Microsoft logging framework
InMemory: WIP

Built in metrics in Proto.Actor

List all the metric names and based on module: https://github.com/asynkron/protoactor-dotnet/issues/948

Proto.Actor Metrics

Name	Type	Labels
protoactor_threadpool_latency_duration_seconds	Histogram	id, address
protoactor_deadletter_count	Count	id, address, messagetype
protoactor_actor_spawn_count	Count	id, address, actortype
protoactor_actor_stopped_count	Count	id, address, actortype
protoactor_actor_restarted_count	Count	id, address, actortype
protoactor_actor_failure_count	Count	id, address, actortype
protoactor_actor_mailbox_length	Gauge	id, address, actortype
protoactor_actor_messagereceive_duration_seconds	Histogram	id, address, actortype, messagetype
protoactor_future_started_count	Count	id, address
protoactor_future_timedout_count	Count	id, address
protoactor_future_completed_count	Count	id, address

Proto.Remote Metrics

Name	Type	Labels
protoremote_message_serialize_count	Count	id, address, messagetype
protoremote_message_deserialize_count	Count	id, address, messagetype
protoremote_kind_count	Count	id, address
protoremote_spawn_count	Count	id, address, kind
protoremote_endpoint_connected_count	Count	id, address, destinationaddress
protoremote_endpoint_disconnected_count	Count	id, address, destinationaddress

Proto.Cluster Metrics

Name	Type	Labels
protocluster_virtualactors	Gauge	id, address, clusterkind
protocluster_virtualactor_spawn_duration_seconds	Histogram	id, address, clusterkind
protocluster_virtualactor_requestasync_duration_seconds	Histogram	id, address, clusterkind, messagetype, pidsource
protocluster_virtualactor_requestasync_retry_count	Count	id, address, clusterkind, messagetype
protocluster_topology_events	Gauge	id, address, membershiphashcode
protocluster_resolve_pid_duration_seconds	Histogram	id, address, clusterkind

Running monitoring systems using Docker

In this section, we will examine how to run Prometheus and Grafana locally using Docker. This is very easy to do, just run one Docker command.

Running Prometheus

First, you need to download the Docker image using command:

docker pull prom/prometheus

In order to run Prometheus on Docker use the following command

docker run -p 9090:9090 prom/prometheus

This command starts Prometheus with initial configuration on port 9090. Read more about running Prometheus here.

Running Grafana

Select the version of Grafana you need and download the corresponding Docker image using the command:

docker pull grafana/grafana:<version>

To run Grafana use the following command

docker run -d -p 3000:3000 grafana/grafana

Read more about running Grafana using Docker here.