Keep instantiated OffsetMapCodecManager so that metrics will not be recreated every commit #859

[Stefan Pries (priesus)]

We can see the OffsetMapCodecManager being instantiated every second or so, which creates 3 new Meters (Timer, Gauge, Counter) every time.

Because they are referenced from the registeredMeters list in PCMetrics, the duplicate Meters and Tags won't be cleaned up by garbage collection. This creates a significant memory leak over time (issue #859).

This is my first open source contribution, please let me know if this change doesn't make sense or what contribution guidelines I have missed!

Checklist

• Documentation (if applicable)
• Changelog

[Stefan Pries (priesus)]

To provide additional context:
In our production environment we see that per 24 hours roughly 1 million Meter.Id objects created that are not being garbage-collected. This seems independent of consumer throughput.

The problem is that OffsetMapCodecManager gets instantiated a lot and each time the same 3 metrics are created once more.

[Stefan Pries (priesus)]

Nacho Muñoz Gómez (@nachomdo) or Roman Kolesnev (@rkolesnev) Are you maybe part of the reviewer group and can have a look at this PR?

[Stefan Pries (priesus)]

Stephen Hegarty (@stephenheg) John Byrne (@johnbyrnejb) Would any of you be available to review this fix? I was hoping to get the memory leak fixed before black friday :-)

[Stefan Pries (priesus)]

Martyn Ye (@sangreal) Are you part of the reviewer group and can help out here with a review?

[Martyn Ye (sangreal)]

sorry, I am not a official Confluent reviewer

[Roman Kolesnev (rkolesnev)]

LGMT.

Stefan Pries (@priesus) - Changelog / readme needs to be updated please as well - you can see how its updated in other PRs - its the CHANGELOG.adoc file and then once its updated - regenerate readme to take in the changes - mvn asciidoc-template:build .

John Byrne (@johnbyrnejb) - could you please do formal approval / merge to master once docs are updated?

[Roman Kolesnev (rkolesnev)]

Nacho Muñoz Gómez (@nachomdo) or Roman Kolesnev (@rkolesnev) Are you maybe part of the reviewer group and can have a look at this PR?

Hey Stefan Pries (@priesus) - i dont work for Confluent anymore and dont use Parallel Consumer in my current day to day - but i try to keep an eye on it when i have a bit of time...

[Stefan Pries (priesus)]

Roman Kolesnev (@rkolesnev) Thanks so much for the feedback! I've updated the changelog & readme.
John Byrne (@johnbyrnejb) Let me know if this looks good to you as well!

[Stefan Pries (priesus)]

John Byrne (@johnbyrnejb) I believe my changes broke the master branch build. I'm unable to see the build logs and I'm also unable to run the tests myself because our company policy doesn't let us use the maven repositories required by this project.
Could you help me figure out what broke?
Sorry for the troubles :-(

[Stefan Pries (priesus)]

Just to clarify: To verify my changes I had to copy the /src/main/java into one of our projects and I was able to fix the memory leak with the changes from this PR.

So I assume the build issue is related to a test. Let me see if I can troubleshoot this somehow.

[Stefan Pries (priesus)]

John Byrne (@johnbyrnejb) So far I am still unable to get the tests to run locally, because I am unable to fetch the io.stubbs.truth dependencies. Would you be able to help check what's broken in the current build?

[Antony Stubbs (astubbs)]

John Byrne (John Byrne (@johnbyrnejb)) So far I am still unable to get the tests to run locally, because I am unable to fetch the io.stubbs.truth dependencies. Would you be able to help check what's broken in the current build?

it wouldn't have bene broken, but the truth def is on githubs maven distribution system. I'm going to publish it to maven central soon to avoid this problem.