Memory usage increased between Quarkus 3.6 and 3.7 (classloading?) about quarkus

[gsmet]

If you can publish a 3.6 heap dump and a 3.7 one, it probably would help to all work on the same baseline.

[holly-cummins]

If you can publish a 3.6 heap dump and a 3.7 one, it probably would help to all work on the same baseline.

I've got the dumps, and will attach them, but I don't think I have the 'right' dump. I had two options for collection the dumps:

• Option 1. Manually trigger a dump. To make sure I was comparing apples to apples, I dumped when the app had completed its startup. At this point, all the classes associating with classloading are no longer live, and the steady state of 3.7 looks pretty similar to the steady state of 3.6. The startup cycle is so quick, I don't think I could reliably trigger a dump partway through startup in a way that was consistent enough to allow 3.6 and 3.7 to be compared. If I'm a bit slower for one, the heap looks different, but it's just that I was 0.2 seconds later with my typing and startup had progressed more ...
• Option 2. Dump on OOM. To do this, I have to constrain the heap in such a way that both 3.6 and 3.7 both dump. When I do this, the heap usage for 3.6 and 3.7 is the same, and the profile of what's loaded is also pretty similar. If 3.6 needs 128m, and 3.7 needs 140m, but I only give them 110m, that means my dump 'misses' the extra 30m that causes a problem in 3.7.

Dumps are here: https://drive.google.com/file/d/1gaGHo-qHu2GImdskJMiF-dNCocabGAfG/view?usp=sharing
I've done 3.6 and 3.7, for manually triggered post-startup dumps in an unconstrained heap, and dumps triggered on OOM in a constrained heap. The constrained heap dumps are bigger.

[holly-cummins]

Some of what I'm seeing in the dumps also looks similar to what Guillaume investigated in #35280, so I think this issue may be in a similar domain to that one.