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 send 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 distribution metric to see how many queries were completed in particular time buckets:

Telemetry.Metrics.distribution("http.request.stop.duration", buckets: [100, 200, 300])

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 either an arbitrary term, a key in the event's measurements map, or a function accepting the whole measurements map 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.
• :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 unit conversions are supported. We discuss those in detail in the "Converting Units" section.

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:

count("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 conversions are supported, which means that both from_unit and to_unit need to be one of :second, :millisecond, :microsecond, :nanosecond, or :native. That's because most time measurements in the Erlang VM are done in the :native unit, which we need to convert to the desired precision.

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

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

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. 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_active_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")

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 procesing 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.Metriics 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.

You may search for available reporters on hex.pm. You can also read the Writing Reporters page for general information on how to write a reporter.