Instrumenting Phoenix with Telemetry Part II: Telemetry Metrics + Reporters

29 Apr 2020 · by Sophie DeBenedetto in General

Table Of Contents

In this series, we’re instrumenting a Phoenix app and sending metrics to StatsD with the help of Elixir and Erlang’s Telemetry offerings.

Intro

In Part I of this series, we learned why observability is important and introduced Erlang’s Telemetry library. We used it to hand-roll some instrumentation for our Phoenix app, but it left us with some additional problems to solve. In this post, we’ll use Elixir’s Telemetry.Metrics and TelemetryMetricsStatsd libraries to define and send metrics to StatsD for a given Telemetry event.

Recap

In our previous post, we added some Telemetry instrumentation to our Phoenix app, Quantum. You can review the final code from our previous article here. To recap, we established a Telemetry event, [:phoenix, :request], that we attached to a handler module, Quantum.Telemetry.Metrics. We executed this event from just one controller action–the new action of the UserController.

From that controller action, we execute the Telemetry event with a measurement map that includes the duration of the web request along with the request conn:

# lib/quantum_web/controllers/user_controller.ex
def new(conn, _params) do
  start = System.monotonic_time()
  changeset = Accounts.change_user(%User{})
  :telemetry.execute([:phoenix, :request], %{duration: System.monotonic_time() - start}, conn)
  render(conn, "new.html", changeset: changeset)
end

We handle this event in our handler module, Quantum.Telemetry.Metric, with the handle_event/4 callback function. In this function we use the event data, including the duration and information in the conn to, send a set of metrics to StatsD with the help of the Statix Elixir StatsD client library:

# lib/quantum/telemetry/metrics.ex
defmodule Quantum.Telemetry.Metrics do
  require Logger
  alias Quantum.Telemetry.StatsdReporter

  def handle_event([:phoenix, :request], %{duration: dur}, metadata, _config) do
    StatsdReporter.increment("phoenix.request.success", 1)
    StatsdReporter.timing("phoenix.request.success", dur)
  end
end

What’s Wrong With This?

Telemetry made it easy for us to emit an event and operate on it, but our current usage of the Telemetry library leaves a lot to be desired.

One drawback of our current approach is that it leaves us on the hook for Telemetry event handling and metrics reporting. We had to define our own custom event handling module, manually attach that module to the given Telemetry event and define the handler’s callback function.

In order for that callback function to report metrics to StatsD for a given event, we had to create our own custom module that uses the Statix library and write code that formats the metric to send to StatsD for a given Telemetry event. The mental overhead of translating Telemetry event data into the appropriate StatsD metric is costly, and that effort will have to be undertaken for every new Telemetry event we execute and handle.

We Need Help

Wouldn’t it be great if we didn’t have to define our own handler modules or metric reporting logic? If only there was some way to simply list the Telemetry events we care about and have them automatically reported to StatsD as the correctly formatted metric…

This is exactly where the Telemetry.Metrics and TelemetryMetricsStatsd libraries come in!

Introducing Telemetry.Metrics and TelemetryMetricsStatsd

The Telemetry.Metrics library provides a common interface for defining metrics based on Telemetry events. It allows us declare the set of Telemetry events that we want to handle and specify which metrics to construct for these events. It also allows us to specify an out-of-the-box reporter with which to handle and report our events to third-parties.

This means we don’t have to define our own handler modules and functions and we don’t have to write any code responsible for reporting metrics for events to common third-party tools like StatsD. We’ll report our metrics to StatsD with the TelemetryMetricsStatsd reporting library, but Elixir’s Telemetry family of libraries also includes a reporter for Prometheus, or you can roll your own.

In the previous post, we added code to execute the following Telemetry event from the new action of our UserController:

:telemetry.execute([:phoenix, :request], %{duration: System.monotonic_time() - start}, conn)

Now, instead of handling this event with our custom handler and Statix reporter, will use Telemetry.Metrics and the TelemetryMetricsStatsd reporter to do all of the work for us!

How It Works

Before we start writing code, let’s walk through how Telemetry.Metrics and the TelemetryMetricsStatsd reporter work together with Erlang’s Telemetry library to handle Telemetry events.

The Telemetry.Metrics library is responsible for specifying which Telemetry events we want to handle as metrics. It defines the list of events we care about and specifies which events should be sent to StatsD as which type of metric (for example, counter, timing, gauge etc.). It gives this list of events-as-metrics to the Telemetry reporting client, TelemetryMetricsStatsd.

The TelemetryMetricsStatsd library is responsible for taking that list of events and attaching its own event handler module, TelemetryMetricsStatsd.EventHandler to each event via a call to :telemetry.attach/4. Recall from our first post that :telemetry/attach/4 stores events and their associated handlers in an ETS table.

Later, when a Telemetry event is executed via a call to :telemetry.execute/3, Telemetry looks up the event handler, TelemetryMetricsStatsd.EventHandler, for the given event in the ETS table and invokes it. The event handler module will format the event, metadata and any associated tags as the appropriate StatsD metric and send the resulting metric to StatsD over UDP.

Most of this happens under the hood. We are only on the hook for defining a Telemetry.Metrics module and listing the Telemetry events we want to handle as which type of metric. That’s it!

Getting Started

You can follow along with this tutorial by cloning down the repo here.

Overview

In order to get this Telemetry pipeline up and running, we don’t have to write too much code.

We will:

  1. Define a supervisor module that imports Telemetry.Metrics
  2. Define a set of metrics for the Telemetry events we want to observe using the Telemetry.Metrics metrics definition functions
  3. Tell the supervisor to run the TelemetryMetricsStatsd GenServer with the list of metrics we defined in the previous step

Let’s do it!

Setting Up Telemetry.Metrics

First, we’ll add the Telemetry.Metrics library and the TelemetryMetricsStatsd reporter library to our application’s dependencies and run mix deps.get:

# mix.exs
defp deps do
  [
    {:telemetry_metrics, "~> 0.4"},
    {:telemetry_metrics_statsd, "~> 0.3.0"}
  ]
end

Now we’re ready to define a module that imports Telemetry.Metrics.

Step 1: Defining a Metrics Module

We’ll define a module that imports the Telemetry.Metrics library and acts as a Supervisor. Our Supervisor will start up the child GenServer provided by the TelemetryMetricsStatsd reporter. It will start that GenServer along with an argument of the list of Telemetry events to listen for, structured as metrics, via the :metrics option.

We’ll place our metrics module in lib/quantum/telemetry.ex

defmodule Quantum.Telemetry do
  use Supervisor
  import Telemetry.Metrics

  def start_link(arg) do
    Supervisor.start_link(__MODULE__, arg, name: __MODULE__)
  end

  def init(_arg) do
    children = [
      {TelemetryMetricsStatsd, metrics: metrics()}
    ]

    Supervisor.init(children, strategy: :one_for_one)
  end

  defp metrics do
    [
      # coming soon!
    ]
  end
end

We’ll come back to the metrics list in a bit. First, let’s teach our application to start this Supervisor when the app starts up but adding it to our application’s supervision tree in the Quantum.Application.start/2 function:

# lib/quantum/application.ex
def start(_type, _args) do
  children = [
    Quantum.Repo,
    QuantumWeb.Endpoint,
    Quantum.Telemetry
  ]

  opts = [strategy: :one_for_one, name: Quantum.Supervisor]
  Supervisor.start_link(children, opts)
end

Now we’re ready to specify which Telemetry events to handle as metrics.

Step 2: Specifying Events As Metrics

Our Telemetry.Metrics module, Quantum.Telemetry, is responsible for telling the TelemetryMetricsStatsd GenServer which Telemetry events to respond to and how to treat each event as a specific type of metric.

We want to handle the [:phoenix, :request] event described above. First, let’s consider what type of metrics we want to report for this event. Let’s say we want to increment a counter for each such event, thereby keeping track of the number of requests our app receives to the endpoint. Let’s also send a timing metric to report the duration of a given web request.

Now that we have a basic idea of what kind of metrics we want to construct and send to StatsD for our event, let’s take a look at how Telemetry.Metrics allows us to define these metrics.

Defining Our Metrics

The Telemetry.Metrics module provides a set of five metrics functions. These functions are responsible for formatting Telemetry event data as a given metric.

We’ll use the Telemetry.Metrics.counter/2 and the Telemetry.Metrics.summary/2 functions to define our metrics for the given event.

In our Quantum.Telemetry module, which imports Telemetry.Metrics, we’ll add the following to the metrics function:

# lib/quantum/telemetry.ex
defp metrics do
  [
    summary(
      "phoenix.request.duration",
      unit: {:native, :millisecond},
      tags: [:request_path]
    ),

    counter(
      "phoenix.request.count",
      tags: [:request_path]
    )
  ]
end

Each metric function takes in two arguments:

  • The event name
  • A list of options

And returns a struct that describes the given metric type. For example, the counter/2 function returns a %Telemetry.Metrics.Counter{} struct that looks like this:

%Telemetry.Metrics.Counter{
  description: Telemetry.Metrics.description(),
  event_name: :telemetry.event_name(),
  measurement: Telemetry.Metrics.measurement(),
  name: Telemetry.Metrics.normalized_metric_name(),
  reporter_options: Telemetry.Metrics.reporter_options(),
  tag_values: (:telemetry.event_metadata() -> :telemetry.event_metadata()),
  tags: Telemetry.Metrics.tags(),
  unit: Telemetry.Metrics.unit()
}

Now that we’ve defined our metrics list, we’re ready for the next step.

Step 3: Start The TelemetryMetricsStatsd GenServer with the Metrics List

The list of metrics structs gets passed to the TelemetryMetricsStatsd GenServer when it gets started up:

defmodule Quantum.Telemetry do
  use Supervisor
  import Telemetry.Metrics

  def start_link(arg) do
    Supervisor.start_link(__MODULE__, arg, name: __MODULE__)
  end

  def init(_arg) do
    children = [
      {TelemetryMetricsStatsd, metrics: metrics()} # here!
    ]

    Supervisor.init(children, strategy: :one_for_one)
  end

  defp metrics do
    [
      summary(
        "phoenix.request.duration",
        unit: {:native, :millisecond},
        tags: [:request_path]
      ),

      counter(
        "phoenix.request.count",
        tags: [:request_path]
      )
    ]
  end
end

This kicks off the following process:

  • When the TelemetryMetricsStatsd starts, it stores events in ETS along with their handler and a config map including this list of metric structs
  • Later, when TelemetryMetricsStatsd is responding to executed events, it looks up the event in ETS and uses the metrics structs stored in that config map to format the appropriate metrics for sending to StatsD

Seeing It In Action

Note that in our call to counter/2 and summary/2, we’re using the :tag option to specify which tags will be applied to the metric when it gets sent to StatsD. The TelemetryMetricsStatsD reporter will, when it receives our [:phoenix, :request] event, grab any values for the tag keys that are present in the event metadata and use them to construct the metric.

So, when we execute our Telemetry with the conn passed in as the metadata argument:

# lib/quantum_web/controllers/user_controller.ex
def new(conn, _params)
  :telemetry.execute([:phoenix, :request], %{duration: System.monotonic_time() - start}, conn)
end

The TelemetryMetricsStatsD will format a counter and summary metric tagged with the value of the :request_path key found in the conn.

So, if we run our app and send some web requests, we’ll see the following metrics reported to StatsD:

{
  counters: {
    'phoenix.request.count.-register-new': 2,
  },
  timers: {
    'phoenix.request.count.-register-new': [ 0, 0 ],
  },
  timer_data: {
    'phoenix.request.duration.-register-new': {
      count_90: 2,
      mean_90: 0,
      upper_90: 0,
      sum_90: 0,
      sum_squares_90: 0,
      std: 0,
      upper: 0,
      lower: 0,
      count: 2,
      count_ps: 0.2,
      sum: 0,
      sum_squares: 0,
      mean: 0,
      median: 0
    }
  }
}

Under The Hood

The Quantum.Telemetry module is, believe it or not, the only code we have to write in order to send these metrics to StatsD for the "phoenix.router_dispatch.stop" event. The Telemetry libraries take care of everything else for us under the hood.

Let’s take a closer look at how it all works.

  1. The Telemetry.Metrics supervisor that we defined in Quantum.Telemetry defines a list of metrics that we want to emit to StatsD for a given Telemetry event
  2. When the supervisor starts, it starts the TelemetryMetricsStatsd GenServer and gives it this list
  3. When the TelemetryMetricsStatsd GenServer starts, it calls :telemetry.attach/4 for each listed event, storing it in an ETS table along with the handler callback and a config map that includes the metrics definitions. The handler callback it gives to :telemetry.attach/4 is its own TelemetryMetricsStatsd.EventHandler.handle_event/4 function.
  4. Later, when a Telemetry event is executed via a call to :telemetry.execute/3, Telemetry looks up the handler callback and config (including metrics definitions) for the given event in ETS
  5. The :telemetry.execute/3 function then calls this handler callback, TelemetryMetricsStatsd.EventHandler.handle_event/4, with the event name, event measurement map, event metadata and metrics config
  6. The TelemetryMetricsStatsd.EventHandler.handle_event/4 function formats the appropriate metric using all of this information and sends it to StatsD over UDP

Phew!

Let’s take a deeper dive into this process by taking a look at some source code.

TelemetryMetricsStatsd Attaches Events to Handlers and Config Data

When our supervisor starts the TelemetryMetricsStatsd GenServer, the GenServer’s init/1 function calls on TelemetryMetricsStatsd.EventHandler.attach/7 with a set of arguments that includes the metrics list we provided. This in turn executes a call to :telemetry.attach/4:

# telemetry_metrics_statsd/lib/telemetry_metrics_statsd/event_handler.ex

def attach(metrics, reporter, mtu, prefix, formatter, global_tags) do
  metrics_by_event = Enum.group_by(metrics, & &1.event_name)

  for {event_name, metrics} <- metrics_by_event do
    handler_id = handler_id(event_name, reporter)

    :ok =
      :telemetry.attach(handler_id, event_name, &__MODULE__.handle_event/4, %{
        reporter: reporter,
        metrics: metrics,
        mtu: mtu,
        prefix: prefix,
        formatter: formatter,
        global_tags: global_tags
      })
  end
end

The call to :telemetry.attach/4 will create an ETS entry that stores the event name along with the handler callback function,&TelemetryMetricsStatsd.EventHandler.handle_event/4, and a config map that contains the metrics definitions for the event.

TelemetryMetricsStatsd.EventHandler Handles Executed Events

Later, the [:phoenix, :request] event is executed in our UserController:

# lib/quantum_web/controllers/user_controller.ex
def new(conn, _params)
  :telemetry.execute([:phoenix, :request], %{duration: System.monotonic_time() - start}, conn)
end

The :telemetry.execute/3 function looks up the event in ETS. It fetches the handler callback function, along with the config that was stored for that event, including the list of metric definitions.

Telemetry will then call the callback function, TelemetryMetricsStatsd.EventHandler.handle_event/4, with the provided measurement map and metadata, along with stored config it looked up for the event in ETS.

TelemetryMetricsStatsd.EventHandler.handle_event/4 will format the metric according to the metrics definitions stored in ETS for the event and send the resulting metric to StatsD.

Here we can see that the TelemetryMetricsStatsd.EventHandler.handle_event/4 iterates over the metric definitions for the event and constructs the appropriate metric from the event data using the given measurement and metadata maps along with the metric struct from the list of metrics stored in the config. It then publishes the metric to StatsD over UDP via the call to publish_metrics/2

# telemetry_metrics_statsd/lib/telemetry_metrics_statsd/event_handler.ex

def handle_event(_event, measurements, metadata, %{
      reporter: reporter,
      metrics: metrics,
      mtu: mtu,
      prefix: prefix,
      formatter: formatter_mod,
      global_tags: global_tags
    }) do
  packets =
    # iterate over the stored metric definitions
    for metric <- metrics do
      # get the measurement for the metric type from the measurements map
      case fetch_measurement(metric, measurements) do
        {:ok, value} ->
          # collect metric tags specified in the metric struct
          tag_values =
            global_tags
            |> Map.new()
            |> Map.merge(metric.tag_values.(metadata))
          tags = Enum.map(metric.tags, &{&1, Map.fetch!(tag_values, &1)})
          # format the metric given the metric type, value and tags
          Formatter.format(formatter_mod, metric, prefix, value, tags)

        :error ->
          :nopublish
      end
    end
    |> Enum.filter(fn l -> l != :nopublish end)

  case packets do
    [] ->
      :ok

    packets ->
      # publish metrics to StatsD
      publish_metrics(reporter, Packet.build_packets(packets, mtu, "\n"))
  end
end

Conclusion

The Telemetry.Metrics and TelemetryMetricsStatsd libraries make it even easier for us to handle Telemetry events and report metrics based on those events. All we have to do is define a Supervisor that uses Telemetry.Metrics and tell that Supervisor to start the TelemetryMetricsStatsd GenServer with a list of metric definitions.

That’s it! The TelemetryMetricsStatsd library will take care of calling :telemetry.attach/3 to store events in ETS along with a handler callback function and the metrics list for that event. Later, when a Telemetry event is executed, Telemetry will lookup the event and its associated handler function and metrics list and invoke the handler function with this data. The handler function, TelemetryMetricsStatsd.EventHandler.handle_event/4, will iterate over the list of metric structs that was stored for the event in ETS and construct the appropriate StatsD metric given the metric type and tags, the event measurement map and metadata. All for free!

Next Up

In this post, we saw how Telemetry.Metrics and TelemetryMetricsStatsd abstracted away the need to define custom handlers and callback functions, attach those handlers to events and implement our own metric reporting logic. But our Telemetry pipeline still needs a little work.

We’re still on the hook for emitting all of our own Telemetry events.

In order to really be able to observe the state of our production Phoenix app, we need to be reporting on much more than just one endpoint’s request duration and count. We want to be able to handle information-rich events describing web requests across the app, database queries, the behavior and state of the Erlang VM, the behavior and state of any workers in our app, and more.

Instrumenting all of that by hand, by executing custom Telemetry events wherever we need, them will be tedious and time-consuming. On top of that, it will be a challenge to standardize event naming conventions, measurements and metadata across the app.

In next week’s post, we’ll examine Phoenix and Ecto’s out-of-the-box Telemetry events and use Telemetry.Metrics to observe a wide-range of such events, thus eliminating the need for us to execute our own custom events for most of our observability use-cases.

Caught a mistake or want to contribute to the article? Edit this page on GitHub!

Sophie DeBenedetto

Sophie is an engineer and teacher at The Flatiron School. She loves teaching and learning and finding the Elixir School community felt like the perfect fit!