Leveraging Quarkus build-time metaprogramming capabilities to improve Jackson's serialization performance

[giscus[bot]]

Leveraging Quarkus build-time metaprogramming capabilities to improve Jackson's serialization performance

Quarkus: Supersonic Subatomic Java

https://quarkus.io/blog/quarkus-metaprogramming/

[lost22git]

It would be better to provide a reflect-free introspection for pojo (just like Micronaut did https://docs.micronaut.io/latest/guide/#introspection)

[mariofusco]

This is exactly what Jandex does https://smallrye.io/jandex/

[lost22git]

I don't think they implement the same functionality

[mensinda]

Would it be possible to also enable this feature for non REST endpoints via an annotation? We are heavily using Jackson for serialization / deserialization outside of the REST context. It would be great if we could also use this new generator and ObjectMapper for our non-REST use-case. Most of our beans are auto-generated, so adding an annotation for registering it with Quarkus wouldn't be a problem.

Also, this blog only covers the serialization part. Is reflection-free deserialization also implemented / planned?

[mariofusco]

Yes to both questions: we are evaluating how to extend this feature for any JSON serialization and I'm already at work on the deserialization part.

[Sineaggi]

How does the performance of standard jackson and this compare to jackson with the blackbird module?

[rquinio1A]

Very nice example on how to leverage build-time processors in a concrete case!

Do you know what Jackson is using reflection for ?
I always thought reflection was only used once for a given type, and then cached in the static ObjectMapper instance.
If not, it might be worth reporting the perf issue to Jackson project.

[mariofusco]

I'm not sure I'm understanding this comment, what do you mean with "cached in the static ObjectMapper"? How can you cache the use of reflection?

[rquinio1A]

Thinking about it, Jackson still needs to call type.newInstance() and type.setField(value).
What gets cached at runtime by Jackson is more or less what Jandex knows at build time: discovering types, fields, method return types, annotations, etc.

[mariofusco]

Correct. Once Jackson has discovered which java.lang.reflect.Method has to be called to read a given field of an object (and cached it) it still has to use reflection and call method.invoke(obj); every single time which is much slower than call directly obj.get() through the serializers generated at build time.

[mensinda]

Does / will this also work with records and classes with public fields only (no getters / setters)?

[mariofusco]

Yes.

[devgauravjava]

can we do the below with metaprogramming or creating an extension where i can dynamically create kafka consumer based on different application properties because the core logic will be same just topics will be different

[t1]

Nice and very promising approach. I just wonder: why did you choose Jackson and not JSON-B? The latter is a standard open for alternative implementations. And the performance of the reference implementation YASSON is even worse than Jackson... by far! And there is a TCK to measure your progress.