Home

Awesome

Metrix

An Elixir library to log custom application metrics.

Metrix subscribes to the Twelve Factor App notion that logs are streams of time-ordered events and that events should be captured and recorded in the l2met logging convention.

What this means in pragmatic terms is that this library provides a few convenience methods to help you capture data from your application to your logging output in a well-structured key-value format:

measure#api.request.service=23.43ms path=/v1/user.json response_status=200
sample#api.response.size=1.34kb path=/v1/user.json
count#user.login user_id=32 other="data with spaces"

Note: Metrix does not do any calculations itself. It merely logs the data its given in a specific format. If you are looking for an in-app instrumentation library, you may want to look at exometer or folsom instead.

Benefits

Treating "logs as data" in this manner has several advantages, including:

Install

Add metrix to your applications in mix.exs:

def application do
  [mod: {YourApp, []},
   applications: [..., :metrix]]
end

And declare it as a dependency:

defp deps do
  [
    # ...
    {:metrix, "~> 1.0"}
  ]
end

Then update your dependencies:

$ mix deps.get

Usage

There are three types of metrics natively supported by Metrix: counts, samples and measurements.

Count

When you want to count the occurrences of an event in your app, use count:

import Metrix

count("app.event")
count("app.event", 3)

Which will output:

count#event.name=1
count#event.name=3

Event metadata can be attached by passing in a map as the first argument:

%{"path" => "/users/1"} |> count("app.event")

Which outputs:

count#event.name=1 path=/users/1

metadata can be a map or keyword list (though we'd encourage you to use a map to avoid unbounded atom creation).

[path: "/users/1"] |> count("app.event")

When passed to log processors that are capable of parsing structured logs, counts can be min, max, summed, stacked etc...

Sample

Samples are used to take periodic, point-in-time, measurements such as CPU load, hard disk space etc...

Samples are logged in the same fashion as count:

import Metrix

sample("file.size", "12.3kb")
%{"file" => "/images/hi.png"} |> sample("file.size", "12.3kb")

Which will output:

sample#file.size=12.3kb
sample#file.size=12.3kb file=/images/hi.png

Samples can be averaged, p50, p95, and p99d.

Measure

Measurements are measures of time, most often used to track execution time. As such, they wrap a block of code whose execution time is to be measured:

import Metrix

measure("api.request", fn -> HTTPotion.get "httpbin.org/get" end)

%{"path" => "/get"}
|> measure("api.request", fn -> HTTPotion.get "httpbin.org/get" end)

Measurements are taken in ms:

measure#api.request=142ms
measure#api.request=131ms path=/get

It's common to want to add metadata to a measurement that is used within the function call (the path above being a good example). Instead of providing a no-arg function to measure, you can provide a 1-arity function that accepts the metrics metadata (and can pattern match against it). For instance:

%{"path" => "/get", "client" => "elixir"}
|> measure("api.request", fn(%{"path" => path}) -> HTTPotion.get "httpbin.org#{path}" end)
measure#api.request=131ms path=/get client=elixir

There may be occasions when the latency measurement is pre-computed. In such a case, you can substitute a millisecond measurement argument in place of the function to measure:

latency_ms = 89.21
measure("api.request", latency_ms)

%{"path" => "/get"}
|> measure("api.request", latency_ms)

Which will output:

measure#api.request=89.21ms
measure#api.request=89.21ms path=/get

Measurements can be aggregated into median, p95 and p99 series:

Global context

Often times there is metadata you want applied to every measurement. For instance a source element indicating which server the output originated from or app which differentiates output from multiple components going to the same downstream processor. This type of universally applicable metadata can be set once using the global context:

Metrix.add_context(%{"source" => System.get_env("NODE_NAME")})
Metrix.count("event.name")

Outputs:

count#event.name=1 source=node.us-east.1a

The context can be cleared with Metrix.clear_context, though be aware it is global context and will be cleared for all output.

Global prefix

Similar to the global context, it's common to want to prefix all metrics within from same application with the same namespace. You can set and manage the global prefix as you do the global context:

Metrix.put_prefix(System.get_env("APP_PREFIX"))
Metrix.count "event.name"

Outputs:

count#app-prefix.event.name=1

The prefix can be cleared with Metrix.clear_prefix, though be aware it is global prefix and will be cleared for all output.

Configuration

More conventionally, Metrix allows its global context and prefix to be set via the app configuration:

config :metrix, context: %{"source" => "my-app"}
config :metrix, prefix: "my-prefix."

Metrix writes to Logger.info. To adjust the output target, set the logger configuration in config.exs. For instance, to write to stdout (the Elixir default) with no timestamp line info, do:

config :logger, :console,
  level: :info,
  format: "$message\n",
  colors: [enabled: false]

Heroku & Librato

Librato is my preferred choice for metrics visualization and long-term storage. It also plays very well with apps deployed to Heroku. Follow these instructions to get your Heroku app's Metrix log output streaming to Librato for processing.

Librato add-on

If your app is deployed to Heroku, just add the Librato add-on and all custom counts, samples and measurements will automatically be sent to Librato which will apply median, p95, p99 and a host of other real-time aggregations. In addition, Heroku's native logging will also be piped to Librato, giving you both platform and app metrics in one place.

External Librato account

If you already have a Librato account, you can still stream your data to from Heroku by setting up a custom log drain.

Development

To develop and test locally we assume the use of the asdf version manager and have a local .tool-versions to establish the correct versions of Erlang/Elixir for this project.

asdf install elixir 1.16.3-otp-26

Then run the local test suite:

mix deps.get
mix test

Todo

There are a few known missing pieces, including:

Contributions

This library was built as an Elixir alternative to the Ruby-based Scrolls library, which I've found to be indispensable. It was also built on top of Logfmt, which handles the mundane but critical task of actually formatting the output as correctly escaped key/value pairs.

Code contributors include:

Changelog

1.0.0

0.5.0

0.4.2

0.4.1

0.4.0

0.3.1

0.3.0