Default otel reporter timeout of 900s is too long.

[tmc]

Since this timeout can occur when a process is closing/exiting it can mean a significant amount of time when the process is just retrying before it ultimately ends.

For example, it's easy to misconfigure the target service e.g.:

	status = StatusCode.UNAVAILABLE
	details = "DNS resolution failed for service: opencensus-collector:55680"

which can cause processes to easily hang at exit (and print nothing by default).

[owais]

I agree it is probably a bit too much but reducing it should not be the solution to services freezing at shutdown. At shutdown, exporters/processors should be given a grace-period to shutdown cleanly and if they don't then the pipeline should force them to shutdown and exit. We should have a different issue for that.

[lzchen]

@owais
What is the decision for this issue? Not increase the timeout but have a better shutdown mechanism for exporters/processors?

[owais]

I think these are two separate issues.

1. Exporter should not hang om shutdown and should have a grace period after which it force quits dropping in memory spans.
2. Exporter should have a shorter timeout on network ops in general as 900s is probably too long unless this is not a real problem and only manifests when network errors happen during shutdown. In that case we only need 1.

[lzchen]

I think we can address both. Java has default timeout to be 10 seconds, maybe we can do the same?

[lzchen]

Related #346

[owais]

SGTM

[srikanthccv]

@lzchen OTLP python also has the default timeout 10 seconds for Export.

opentelemetry-python/exporter/opentelemetry-exporter-otlp/src/opentelemetry/exporter/otlp/exporter.py

Lines 183 to 187 in 24edd3d

	self._timeout = (
	timeout
	or int(environ.get(OTEL_EXPORTER_OTLP_TIMEOUT, 0))
	or 10 # default: 10 seconds
	)

The 900 secs is to deal with the transient errors by using exponential back-off retry strategy. This might still be too long but just want to make sure we are not confusing one with other.

opentelemetry-python/exporter/opentelemetry-exporter-otlp/src/opentelemetry/exporter/otlp/exporter.py

Lines 247 to 298 in 24edd3d

	max_value = 900
	for delay in expo(max_value=max_value):
	if delay == max_value:
	return self._result.FAILURE
	try:
	self._client.Export(
	request=self._translate_data(data),
	metadata=self._headers,
	timeout=self._timeout,
	)
	return self._result.SUCCESS
	except RpcError as error:
	if error.code() in [
	StatusCode.CANCELLED,
	StatusCode.DEADLINE_EXCEEDED,
	StatusCode.PERMISSION_DENIED,
	StatusCode.UNAUTHENTICATED,
	StatusCode.RESOURCE_EXHAUSTED,
	StatusCode.ABORTED,
	StatusCode.OUT_OF_RANGE,
	StatusCode.UNAVAILABLE,
	StatusCode.DATA_LOSS,
	]:
	retry_info_bin = dict(error.trailing_metadata()).get(
	"google.rpc.retryinfo-bin"
	)
	if retry_info_bin is not None:
	retry_info = RetryInfo()
	retry_info.ParseFromString(retry_info_bin)
	delay = (
	retry_info.retry_delay.seconds
	+ retry_info.retry_delay.nanos / 1.0e9
	)
	logger.debug(
	"Waiting %ss before retrying export of span", delay
	)
	sleep(delay)
	continue
	if error.code() == StatusCode.OK:
	return self._result.SUCCESS
	return self._result.FAILURE

[lzchen]

@lonewolf3739
Thanks for the explanation. It's outlined that it is the same value as the Go implementation. If that's the case shouldn't we be consistent? Does 900s make sense for the "conventional" use case of actually retrying instead of this specific use case where the exporter hangs due to misconfiguration? If so, we should just solve (1) as outlined here.

[srikanthccv]

900 secs is too high for maximum backoff time. Generally it is either 32 or 64 seconds and sometimes can be little high based on use case. I couldn't find what is the recommended maximum value by OTLP but I assume it wouldn't be as high 900 seconds. Probably we can raise an issue on the spec repo.

[owais]

Thanks for digging into this @lonewolf3739. Agree we should clarify with spec or at least with other SIG maintainers.

And we should probably have separate tickets for possibly reducing exp-backoff max time and add grace period to exporters on shutdown.

[github-actions]

This issue was marked stale due to lack of activity. It will be closed in 30 days.