Home

Awesome

SpandexDatadog

Package Version Hex Docs Downloads License Last Updated

A datadog adapter for the :spandex library.

Installation

The package can be installed by adding :spandex_datadog to your list of dependencies in mix.exs:

def deps do
  [
    {:spandex_datadog, "~> 1.4"}
  ]
end

To start the datadog adapter, add a worker to your application's supervisor

# Example configuration

# Note: You should put ApiServer before any other children in the list that
# might try to send traces before the ApiServer has started up, for example
# Ecto.Repo and Phoenix.Endpoint

spandex_opts =
  [
    host: System.get_env("DATADOG_HOST") || "localhost",
    port: System.get_env("DATADOG_PORT") || 8126,
    batch_size: System.get_env("SPANDEX_BATCH_SIZE") || 10,
    sync_threshold: System.get_env("SPANDEX_SYNC_THRESHOLD") || 100,
    http: HTTPoison
  ]

children = [
  # ...
  {SpandexDatadog.ApiServer, spandex_opts},
  MyApp.Repo,
  MyAppWeb.Endpoint,
  # ...
]

opts = [strategy: :one_for_one, name: MyApp.Supervisor]
Supvervisor.start_link(children, opts)

Distributed Tracing

Distributed tracing is supported via headers x-datadog-trace-id, x-datadog-parent-id, and x-datadog-sampling-priority. If they are set, the Spandex.Plug.StartTrace plug will act accordingly, continuing that trace and span instead of starting a new one. Both x-datadog-trace-id and x-datadog-parent-id must be set for distributed tracing to work. You can learn more about the behavior of x-datadog-sampling-priority in the Datadog priority sampling documentation.

Telemetry

This library includes telemetry events that can be used to inspect the performance and operations involved in sending trace data to Datadog. Please refer to the telemetry documentation to see how to attach to these events using the standard conventions. The following events are currently exposed:

[:spandex_datadog, :send_trace, :start]

This event is executed when the SpandexDatadog.ApiServer.send_trace/2 API function is called, typically as a result of the finish_trace/1 function being called on your Tracer implementation module.

NOTE: This event is executed in the same process that is calling this API function, but at this point, the Spandex.Trace data has already been passed into the send_trace function, and thus the active trace can no longer be modified (for example, it is not possible to use this event to add a span to represent this API itself having been called).

Measurements

Metadata

[:spandex_datadog, :send_trace, :stop]

This event is executed when the SpandexDatadog.ApiServer.send_trace/2 API function completes normally.

NOTE: This event is executed in the same process that is calling this API function, but at this point, the Spandex.Trace data has already been passed into the send_trace function, and thus the active trace can no longer be modified (for example, it is not possible to use this event to add a span to represent this API itself having been called).

Measurements

Metadata

[:spandex_datadog, :send_trace, :exception]

This event is executed when the SpandexDatadog.ApiServer.send_trace/2 API function ends prematurely due to an error or exit.

NOTE: This event is executed in the same process that is calling this API function, but at this point, the Spandex.Trace data has already been passed into the send_trace function, and thus the active trace can no longer be modified (for example, it is not possible to use this event to add a span to represent this API itself having been called).

Measurements

Metadata

API Sender Performance

Originally, the library had an API server and spans were sent via GenServer.cast, but we've seen the need to introduce backpressure, and limit the overall amount of requests made. As such, the Datadog API sender accepts batch_size and sync_threshold options.

Batch size refers to traces, not spans, so if you send a large amount of spans per trace, then you probably want to keep that number low. If you send only a few spans, then you could set it significantly higher.

Sync threshold is how many simultaneous HTTP pushes will be going to Datadog before it blocks/throttles your application by making the tracing call synchronous instead of async.

Ideally, the sync threshold would be set to a point that you wouldn't reasonably reach often, but that is low enough to not cause systemic performance issues if you don't apply backpressure.

A simple way to think about it is that if you are seeing 1000 request per second and batch_size is 10, then you'll be making 100 requests per second to Datadog (probably a bad config). With sync_threshold set to 10, only 10 of those requests can be processed concurrently before trace calls become synchronous.

This concept of backpressure is very important, and strategies for switching to synchronous operation are often surprisingly far more performant than purely asynchronous strategies (and much more predictable).

Copyright and License

Copyright (c) 2021 Zachary Daniel & Greg Mefford

Released under the MIT License, which can be found in the repository in LICENSE.