A datadog adapter for the spandex library.
The package can be installed by adding spandex_datadog to your list of
dependencies in mix.exs:
def deps do
[
{:spandex_datadog, "~> 1.1"}
]
endTo start the datadog adapter, add a worker to your application's supervisor
# Example configuration
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
]
# in your supervision tree
worker(SpandexDatadog.ApiServer, [opts])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.
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:
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).
:system_time: The time (in native units) this event executed.
trace: The currentSpandex.Tracethat is being sent to theApiServer.
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).
:duration: The time (in native units) spent servicing the API call.
trace: TheSpandex.Tracethat was sent to theApiServer.
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).
:duration: The time (in native units) spent servicing the API call.
trace: The currentSpandex.Tracethat is being sent to theApiServer.:kind: The kind of exception raised.:error: Error data associated with the relevant kind of exception.:stacktrace: The stacktrace associated with the exception.
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).