Memory leak on javaagent 1.5.1

[anuraaga]

Discussed in #3960

^{Originally posted by lmssy August 26, 2021}
Here is an OOM dump below:

javaagent version: 1.5.1
params: -Dotel.metrics.exporter=none -Dotel.exporter.otlp.traces.endpoint=http://myip:4317 -Dotel.traces.sampler=parentbased_traceidratio -Dotel.traces.sampler.arg=1 -Dio.opentelemetry.context.enableStrictContext=true

I have applications run on 1.3.1 and 1.4.1 for a long time without OOM, yeaterday I upgrade to 1.5.1 and my application crashed by OOM.

Please have a look, thanks.

[anuraaga]

If we go with open-telemetry/opentelemetry-java#3555, then one nuance when integrating into the agent will be checking the flag in our own SdkMeterConfigurer - not trivial since the configurer lamely doesn't yet pass ConfigProperties to the configurer.

[jsuereth]

GC of metric usage ONLY occurs when an exporter runs collection....

Agree we should NOT be performing any metric collection if there is no exporter, question is: How do you know there's no exporter given this is a pull-based API (generally)

[trask]

We discussed in the SIG today to only pass these span attributes to metrics until metrics views are in place, at which time we will go back to passing all span attributes.

This should also address #3927.

@jsuereth let us know if you want an option in the meantime to send all span attributes for testing purposes.

[jsuereth]

We discussed in the SiG.

TL;DR:

Long term we should be able to send all attributes from instrumentation and there will be default aggregation occurring that reduces to acceptable cardinality.
Short term just send us the reduced set of attributes until SDK specification has caught up.