Making a CL parallel capable has a cost, should we revise our decision for QuarkusClassLoader?

[gsmet]

We made QuarkusClassLoader parallel capable and this isn't cost free: the parallelLockMap is more often than not close to 1 MB in memory size.

While I don't have a problem with it if it actually makes things significantly faster, I think we should at least check if it's the case.

See here (sorry the selected line is not the right one but it's easy to find it):

[geoand]

Another option is to do what we did in RunnerClassLoader which effectively removes the cost of that map.

[quarkus-bot]

/cc @Sanne (core), @aloubyansky (core), @radcortez (core), @stuartwdouglas (core)

[Sanne]

Another option is to do what we did in RunnerClassLoader which effectively removes the cost of that map.

I like that solution, it's a clever locking scheme which removes the need for the map and also makes it more efficient.

I also think it could be improved more even (had some comments on the original PR), but even in its current form it's an improvement over the previous design.

Awesome @gsmet that you're actively watching for such aspects! Love it, we need more of that.

[gsmet]

Please don't do it right away as there are a few other things going on here: #42525

[geoand]

Yeah, let's leave that for once all the critical stuff is in

[geoand]

With #42525 in, we can no work on this