Track excessive use of registrations for reflective usage

[Sanne]

Description

As the benchmark at https://github.com/quarkusio/spring-quarkus-perf-comparison starts producing interesting data, it also highlight that this not-humoungous application is registering 5600+ classes for reflective instantiaton, and 4200+ fields accessed reflectively, 13000+ methods "using reflection" [1].

I suppose these might have good reasons to be registered, however as a "finger in the air" first impression they seem to me wildly excessive: we should look into these, and see if we can bring those needs for reflective invocations under control.

We might need some tool to break them down / classify, and ideally we'd be able to discuss a strategy for keeping these under control in the long term.

1 - I'm not sure of what exactly such figures refer to, needs to double check.

Implementation ideas

Start looking at the primary offenders and investigate how practical it would be to decise alternative solutions.

Try to document the exploration process, to see if we can make future analysis easier and/or somewhat automated.

Perhaps output such stats as part of each CI build, and see if they can be plotted over time as part of other performance metrics? cc/ @stalep

[stalep]

Hm, that does for sure sound like a fairly high number of classes/methods/fields yes. I assume there must be a cascade of reflection going on?
I understand getting a metric on that would be interesting for sure, we'll look into it!

[diegolovison]

I will start working on it.

[edeandrea]

@Sanne where are you getting the numbers 5600+ classes for reflective instantiaton, and 4200+ fields accessed reflectively, 13000+ methods from?

From the latest run the numbers are

   22,158 types,  28,962 fields, and 104,915 methods found reachable
    7,304 types,     234 fields, and   6,519 methods registered for reflection

[Sanne]

@edeandrea from your slides :) Did I misunderstand?

[mkouba]

FYI this one - #51396 - could help to identify the build steps that produce a large number of Reflective*BuildItems.

[edeandrea]

@edeandrea from your slides :) Did I misunderstand?

That section is stating the difference between Quarkus & Spring (i.e. Quarkus has 5625 less classes using reflection than Spring does).

Those are the numbers reported by GraalVM

[diegolovison]

Definition of Done:

• Once Build metrics: make it possible to collect produced items #51396 is merged
• The metric is being collected from https://github.com/quarkusio/benchmarks/blob/main/results/spring-quarkus-perf-comparison/2025-12-03_01-53-38/metrics.json#L86-L93
• If necessary, update the root's metric source.
• Assert the metric is still presented on Horreum