Add serializer-cache size limit to avoid Metaspace issues from caching Serializers

[mcolemanNOW]

I know this is a topic that has been discussed before (#489 & #1905), but it has been a few years so I wanted to open the discussion again.

I see that there are changes coming in 3.0 to limit the size of the serializer cache to 4000 (#1994). Are there any other plans to allow tuning this size, adding a 'super flush', or letting the consumer provide the cache implementation (ex. with weak references)?

Our app servers have a lot of DelegatingClassLoaders sticking around due to reflection methods being stored in the SerializerCache(like _accessorMethod in BeanPropertyWriter) and it would be great if we had more control over the caching strategy so metaspace doesn't grow unbounded.

[cowtowncoder]

There is nothing specific, except that I think there were some concrete talks about exposing cache sizes through builders -- so at least you could limit maximum number -- and I think also ability to specify custom cache implementations. Latter would then allow more advanced handling if cache could determine how to clear references to obsolete instances.

[cjimenez1]

Is this consideration being made for the codehaus jackson implementation? There is a serializercache there as well, which will lead to pretty much the same impact.

https://github.com/codehaus/jackson/blob/master/src/mapper/java/org/codehaus/jackson/map/ser/impl/SerializerCache.java#L34

[yawkat]

@cjimenez1 The codehaus implementation is Jackson 1.x and hasn't seen a release for 8 years. You should upgrade to Jackson 2.

[mcolemanNOW]

The cache limit hasn't been backported to Jackson 2 either. Could that be considered as well?

[cowtowncoder]

Yeah, Jackson 1.x has not been modified in literally years and I have no way to actually publish any of jars even if I did change the copy Github has. So issues it has will remain issues unless someone forks the build (or implements proper Maven build).

As to 2.x, sure, PRs welcome! I do not necessarily have time to make changes but can usually help others get PRs merged.

[cowtowncoder]

@pjfanning @ben-manes I think I'll try to do this (basically backport of #1994)