Provide interim sampling capability in exporter

[garthk]

There's plenty of progress on the OT specification. If you can wait, you don't need to use this capability. Reading material:

• Sampling Working Group Gitter
• PR: Proposal for sampling.priority: Proposal for sampling.priority open-telemetry/oteps#107
• PR: Add ParentNotRemoteOrElse and fix Probability sampler definition Rename ParentOrElse to ParentBased and generalize to support all cases open-telemetry/opentelemetry-specification#610
• PR: Add requirements for probability sampler Rename ProbabilitySampler to TraceIdRatioBasedSampler and add requirements open-telemetry/opentelemetry-specification#611

As the README change puts it:

Please treat this as an off-API workaround. We trust the OpenTelemetry specification will evolve to cover this use case, after which you'll be able to get this result without an implementation kludge.

That said, some Honeycomb customers need to be able to sample right now because of a change in their pricing model. It's literally a choice between sampling in a Honeycomb-sympathetic way, or turning it off. So, this branch provides a workaround:

• You set some attribute on your spans eg. hc_sample to indicate your preferred sample rate, as a positive integer: 4 means you're dropping 75% of spans

• You configure the exporter with samplerate_key set to something eg. "hc_sample"

• The exporter pops "hc_sample" out of the attribute map so it's no longer confusable for a measurement, and _moves it to the HC samplerate key

• The exporter also drops all but 1 in samplerate spans before export

It's crude, but it'll do for now.

[garthk]

open-telemetry/opentelemetry-specification#570 would provide an official span attribute eg.span.count with the number we need. I'll see if I can wire a sampler up as part of cleaning up and extending opentelemetry-beam/opentelemetry-examples#2.

[aj-foster]

Hi @garthk. Could you update us on the current state of affairs of sampling with Honeycomb, and if workarounds like this continue to be necessary?

[garthk]

👋 Honeycomb have since added native OpenTelemetry delivery, so you shouldn't need this exporter, for starters. Strikes me the correct approach now is to hook up processors to perform the sampling and tell Honeycomb your sample rate. According to their Python sampling , you can do it using environment variables with the Python SDK:

export OTEL_TRACES_SAMPLER="traceidratio"
export OTEL_TRACES_SAMPLER_ARG=0.5
export OTEL_RESOURCE_ATTRIBUTES="SampleRate=2"
# … and then run your software

… and those are in the OpenTelemetry Environment Variable Specification and also documented in the Erlang OTel docs.

For dynamic sampling, you'll want to set SampleRate as a span attribute rather than a resource. If you're adventurous, perhaps you can write your own Probability Sampling processor. I'm no longer working with Elixir, though, so I regret I can't give a hand.