Memory leak of spans with grpc server stats handler

[atoulme]

Description

See open-telemetry/opentelemetry-collector#11875 for original filing.

The gRPC server stats handler set by the OTLP receiver exposes a leak of spans over a long period of time.

From review of the code, it is possible the spans are not always closed by the stats_handler.go.

The span.End() call is made upon receiving the *stats.End event:

opentelemetry-go-contrib/instrumentation/google.golang.org/grpc/otelgrpc/stats_handler.go

Line 205 in eb2064c

span.End()

It seems in light of the leak, given the pprof dumps provided in the original report, that the event is not always received.

[dmathieu]

cc @dashpole as code owner.

[dashpole]

Updated the title and description. Looks like you are using the stats handler, rather than the interceptor.

[atoulme]

Is there a workaround we can apply to deactivate or not trace the stats handler?

[vianamjr]

I noticed something similar by just start using it with sentry. (We where using pure otlp with grpc collector - jaeger).

Since I started sentry and added:

import (
         ....
         sentryotel "github.com/getsentry/sentry-go/otel"
        ....
)

....

	tp := oteltrace.NewTracerProvider(
		oteltrace.WithSpanProcessor(sentryotel.NewSentrySpanProcessor()),
	)

	// set the global tracer.
	otel.SetTracerProvider(tp)
	otel.SetTextMapPropagator(sentryotel.NewSentryPropagator())

....

Also I noticed it going up even without any call (we have health check call, that I filtered out)

The leak seems to start. Not sure where to fix it, maybe sentry is missing something?

Also reported in: getsentry/sentry-go#959

The memory profile (objects):

[vianamjr]

I did some more investigation and it's on filtered calls. Is start a span, but never ends. (using sentry processor)

[vianamjr]

Did this changes:

https://github.com/vianamjr/opentelemetry-go-contrib/pull/1/files