Sub-optimal memory leak capabilities with Pyroscope Java agent

[Nerja]

Hi,

I’m trying to use Grafana Pyroscope to find a simulated memory leak in a small example program. I have two methods called leak() and noLeak() where the leak() method saves data to a Set every time it is called and noLeak() only allocates some memory that will be garbage collected. When I use Async Profiler (2.9) locally with the –live option I see that the leak() method is “leaking” memory. When I try to use the Pyroscope agent instead (io.pyroscope:agent:0.11.3) I don’t see it nearly as clearly.

I used the following configuration when setting up the Pyroscope Java agent:

Config config = new Config.Builder()
                .setApplicationName("leaktest-1")
                .setFormat(Format.JFR)
                .setServerAddress("http://localhost:4040/")
                .setUploadInterval(Duration.ofSeconds(10))
                .setProfilingEvent(EventType.ALLOC)
                .setAllocLive(true)
                .build();

When I run the async profiler, locally, for 60s or more then I can easily see the memory leak. When I run the async profiler for 10s or less than I observe the same behavior as I see with Pyroscope.

I also tried to change the upload interval of the Pyroscope Java agent to 60s, but I am still not able to see the memory leak in a good way.

Slack thread: https://grafana.slack.com/archives/C049PLMV8TB/p1683294542388929