Common interface for defining metrics based on :telemetry events.

Metrics are aggregations of Telemetry events with specific name, providing a view of the system's behaviour over time.

To give a more concrete example, imagine that somewhere in your code there is a function which sends an HTTP request, measures the time it took to get a response, and emits an event with the information:

:telemetry.execute([:http, :request, :stop], %{duration: duration})

You could define a counter metric, which counts how many HTTP requests were completed:

Telemetry.Metrics.counter("http.request.stop.duration")

or you could use a summary metric to see statistics about the request duration:

Telemetry.Metrics.summary("http.request.stop.duration")

This documentation is going to cover all the available metrics and how to use them, as well as options, and how to integrate those metrics with reporters.

Metrics

There are five metric types provided by Telemetry.Metrics:

• counter/2 which counts the total number of emitted events
• sum/2 which keeps track of the sum of selected measurement
• last_value/2 holding the value of the selected measurement from the most recent event
• summary/2 calculating statistics of the selected measurement, like maximum, mean, percentiles etc.
• distribution/2 which builds a histogram of selected measurement

The first argument to all metric functions is the metric name. Metric name can be provided as a string (e.g. "http.request.stop.duration") or a list of atoms ([:http, :request, :stop, :duration]). The metric name is automatically used to infer the telemetry event and measurement. For example, In the "http.request.stop.duration" example, the source event name is [:http, :request, :stop] and metric values are drawn from :duration measurement. Like this:

[:http , :request, :stop]      :duration
<----- event name ------> <-- measurement -->

You can also explicitly specify the event name and measurement if you prefer.

The second argument is a list of options. Below is the description of the options common to all metric types:

• :event_name - the source event name. Can be represented either as a string (e.g. "http.request") or a list of atoms ([:http, :request]). By default the event name is all but the last segment of the metric name.
• :measurement - the event measurement used as a source of a metric values. By default it is the last segment of the metric name. It can be:
  • an arbitrary term
  • a key in the event's measurements map
  • an unary function accepting the whole measurements map and returning the actual value to be used.
  • a binary function accepting the measurements and metadata maps and returning the actual value to be used.
• :tags - a subset of metadata keys by which aggregations will be broken down. Defaults to an empty list.
• :tag_values - a function that receives the metadata and returns a map with the tags as keys and their respective values. Defaults to returning the metadata itself.
• :keep - a predicate function that evaluates the metadata to conditionally record a given event. :keep and :drop cannot be combined. Defaults to nil.
• :drop - a predicate function that evaluates the metadata to conditionally skip recording a given event. :keep and :drop cannot be combined. Defaults to nil.
• :description - human-readable description of the metric. Might be used by reporters for documentation purposes. Defaults to nil.
• :unit - an atom describing the unit of selected measurement, typically in singular, such as :millisecond, :byte, :kilobyte, etc. It may also be a tuple indicating that a measurement should be converted from one unit to another before a metric is updated. Currently, only time and byte unit conversions are supported. We discuss those in detail in the "Converting Units" section.
• :reporter_options - a keyword list of reporter-specific options for the metric.

Breaking down metric values by tags

Sometimes it's not enough to have a global overview of all HTTP requests received or all DB queries made. It's often more helpful to break down this data, for example, we might want to have separate metric values for each unique database table and operation name (select, insert etc.) to see how these particular queries behave.

This is where tagging comes into play. All metric definitions accept a :tags option:

counter("db.query.count", tags: [:table, :operation])

The above definition means that we want to keep track of the number of queries, but we want a separate counter for each unique pair of table and operation. Tag values are fetched from event metadata - this means that in this example, [:db, :query] events need to include :table and :operation keys in their payload:

:telemetry.execute([:db, :query], %{duration: 198}, %{table: "users", operation: "insert"})
:telemetry.execute([:db, :query], %{duration: 112}, %{table: "users", operation: "select"})
:telemetry.execute([:db, :query], %{duration: 201}, %{table: "sessions", operation: "insert"})
:telemetry.execute([:db, :query], %{duration: 212}, %{table: "sessions", operation: "insert"})

The result of aggregating the events above looks like this:

The approach where we create a separate metric for some unique set of properties is called a multi-dimensional data model.

Transforming event metadata for tagging

Finally, sometimes there is a need to modify event metadata before it's used for tagging. Each metric definition accepts a function in :tag_values option which transforms the metadata into desired shape. Note that this function is called for each event, so it's important to keep it fast if the rate of events is high.

Converting Units

It might happen that the unit of measurement we're tracking is not the desirable unit for the metric values, e.g. events are emitted by a 3rd-party library we do not control, or a reporter we're using requires specific unit of measurement.

For these scenarios, each metric definition accepts a :unit option in a form of a tuple:

summary("http.request.stop.duration", unit: {from_unit, to_unit})

This means that the measurement will be converted from from_unit to to_unit before being used for updating the metric. Currently, only time and byte conversions are supported.

Time Conversions

Most time measurements in the Erlang VM are done in the :native unit, which we need to convert to the desired precision. The supported time units are: :second, :millisecond, :microsecond, :nanosecond and :native.

For example, to convert HTTP request duration from :native time unit to milliseconds you'd write:

summary("http.request.stop.duration", unit: {:native, :millisecond})

Byte Conversions

Some metrics, like VM memory's usage are reported in bytes. You might want to convert this to megabytes, for example. The supported byte units are: :byte, :kilobyte and :megabyte.

In order to convert a metric value from bytes to megabytes, you can write the following:

last_value("vm.memory.total", unit: {:byte, :megabyte})

VM metrics

Telemetry.Metrics doesn't have a special treatment for the VM metrics - they need to be based on the events like all other metrics.

:telemetry_poller package (http://hexdocs.pm/telemetry_poller) exposes a bunch of VM-related metrics and also provides custom periodic measurements. You can add telemetry poller as a dependency:

{:telemetry_poller, "~> 0.4"}

By simply adding :telemetry_poller as a dependency, two events will become available:

• [:vm, :memory] - contains the total memory, as well as the memory used for binaries, processes, etc. See erlang:memory/0 for all keys;
• [:vm, :total_run_queue_lengths] - returns the run queue lengths for CPU and IO schedulers. It contains the total, cpu and io measurements;

You can consume those events with Telemetry.Metrics with the following sample metrics:

last_value("vm.memory.total", unit: :byte)
last_value("vm.total_run_queue_lengths.total")
last_value("vm.total_run_queue_lengths.cpu")
last_value("vm.total_run_queue_lengths.io")

If you want to change the frequency of those measurements, you can set the following configuration in your config file:

config :telemetry_poller, :default, period: 5_000 # the default

Or disable it completely with:

config :telemetry_poller, :default, false

The :telemetry_poller package also allows you to run your own poller, which is useful to retrieve process information or perform custom measurements periodically. For example, to keep track of the number of users, inside a supervision tree, you could do:

measurements = [
  {:process_info,
   event: [:my_app, :worker],
   name: MyApp.Worker,
   keys: [:message_queue_len, :memory]},

  {MyApp, :measure_users, []}
]

Supervisor.start_link([
  # Run the given measurements every 10 seconds
  {:telemetry_poller, measurements: measurements(), period: 10_000}
], strategy: :one_for_one)

Where MyApp.measure_users/0 could be written like this:

defmodule MyApp do
  def measure_users do
    :telemetry.execute([:my_app, :users], %{total: MyApp.users_count()}, %{})
  end
end

Now with measurements in place, you can define the metrics for the events above:

last_value("my_app.worker.memory", unit: :byte)
last_value("my_app.worker.message_queue_len")
last_value("my_app.users.total")

Optionally Recording Metric Events

There may be occasions where you want to record metrics differently or not at all based upon metadata. Rather than depending on changing event names with prefixes, you can instead provide a predicate function which returns true when you want the metric to be processed for this event and false when you do not.

Let's examine some a few examples where optional recording can be helpful.

Filtering on Metadata

Let's assume you are using an HTTP client library in your application which has the following event: [:http_client, :request, :stop]. You use this library in multiple places and you'd like to monitor the request duration.

You can use the event provided by the library but you have very different acceptable performance requirements for a critical request, so it would be better to provide a different metric name for monitoring. The client library includes a user configured :name option which you can set and is passed in the event metadata.

Let's create a default distribution metric and one for the high performance call.

distribution(
  "http.client.request.duration",
  event: [:http_client, :request, :stop],
  drop: &(match?(%{name: :fast_client}, &1))
)

distribution(
  "http.fast.client.request.duration",
  event: [:http_client, :request, :stop],
  keep: &(match?(%{name: :fast_client}, &1))
)

With this configuration, you can now monitor these requests separately. In the first example, any events where the client's name is :fast_client will be dropped for that metric. Conversely, any matching events in the second example metric will be kept.

Note: only :keep OR :drop may be set, never both.

Other Uses

Another potential use case for :keep | :drop could be per metric sampling rates. As long as your function returns a boolean, it can determine if the event should be processed.

Reporter Support

Event optional recording must be supported by the reporter you are using. Check your reporter's documentation before relying on this functionality.

The keep function should be evaluated by reporters prior to tag_values using the raw :telemetry.medatadata() values from the event.

Reporters

So far, we have talked about metrics and how to describe them, but we haven't discussed how those metrics are consumed and published to a system that provides data visualization, aggregation, and more. The job of subscribing to events and processing the actual metrics is a responsibility of reporters.

Generally speaking, a reporter is a process that you would start in your supervision tree with a list of metrics as input. For example, Telemetry.Metrics ships with a Telemetry.Metrics.ConsoleReporter module, which prints data to the terminal as an example. You would start it as follows:

metrics = [
  last_value("my_app.worker.memory", unit: :byte),
  last_value("my_app.worker.message_queue_len"),
  last_value("my_app.users.total")
]

Supervisor.start_link([
  {Telemetry.Metrics.ConsoleReporter, metrics: metrics}
], strategy: :one_for_one)

Reporters take metric definitions as an input, subscribe to relevant events and aggregate data when the events are emitted. Reporters may push metrics to StatsD, some time-series database, or exposing a HTTP endpoint for Prometheus to scrape. In a nutshell, Telemetry.Metrics defines only how metrics of particular type should behave and reporters provide the actual implementation for these aggregations.

Official reporters, maintained by the Observability Working Group of the Erlang Ecosystem Foundation, can be found on the BEAM Telemetry organization on GitHub. You may also find community reporters on hex.pm. You can also read the Writing Reporters page for general information on how to write a reporter.

Wiring it all up

Over the previous sections we discussed how to setup metrics and pass them to reporters and how to configure a poller for measurements. We can wire it all up into a single module as shown below. The example below would be used in the context of a Phoenix application, where we have web metrics, database metrics (through Ecto) as well as from the database, Phoenix metrics as well as VM metrics.

The first step is to add both :telemetry_metrics and :telemetry_poller as dependencies:

[
  {:telemetry_poller, "~> 0.4"},
  {:telemetry_metrics, "~> 0.4"}
]

Then you could define a module that wires everything up:

defmodule MyAppWeb.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 = [
      {:telemetry_poller,
       measurements: periodic_measurements(),
       period: 10_000},
      # Or TelemetryMetricsPrometheus or TelemetryMetricsFooBar
      {TelemetryMetricsStatsd, metrics: metrics()}
    ]

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

  defp metrics do
    [
      # VM Metrics
      last_value("vm.memory.total", unit: :byte),
      last_value("vm.total_run_queue_lengths.total"),
      last_value("vm.total_run_queue_lengths.cpu"),
      last_value("vm.total_run_queue_lengths.io"),

      last_value("my_app.worker.memory", unit: :byte),
      last_value("my_app.worker.message_queue_len"),

      # Database Time Metrics
      summary("my_app.repo.query.total_time", unit: {:native, :millisecond}),
      summary("my_app.repo.query.decode_time", unit: {:native, :millisecond}),
      summary("my_app.repo.query.query_time", unit: {:native, :millisecond}),
      summary("my_app.repo.query.idle_time", unit: {:native, :millisecond}),
      summary("my_app.repo.query.queue_time", unit: {:native, :millisecond}),

      # Phoenix Time Metrics
      summary("phoenix.endpoint.stop.duration",
              unit: {:native, :millisecond}),
      summary(
        "phoenix.router_dispatch.stop.duration",
        unit: {:native, :millisecond},
        tags: [:plug]
      )
    ]
  end

  defp periodic_measurements do
    [
      {:process_info,
       event: [:my_app, :worker],
       name: Rumbl.Worker,
       keys: [:message_queue_len, :memory]}
    ]
  end
end