Using Interface-Based Projections Requires a Lot of Heap Memory

[i8r]

When using the feature of interface-based projection the amount of memory that is required is significant. This has been causing issues in production system in our services that have not been able to serve records with more 2 GB of heap even though the actual payload was estimated in the order of magnitude of 10 MB.

I have build a minimal working example to illustrate the issue. It's using:

• org.springframework.boot:spring-boot-starter-data-jpa:3.06
• openjdk Temurin-18.0.2.1+1

The example saves 100k entities in a H2 in-memory database (the same pattern can be observed when using e.g. postgres). Then it reads the ids of those entities 1) using an interface-based projection or 2) just plain Longs. While it is expected that the second representation is smaller than the first one the difference seems too big to be acceptable:

Steps

[mp911de]

Thanks for bringing this to our attention.

Projections are associated with an invocation handler and a backing object at least which requires much more memory than a plain long value.

We will have a look into our infrastructure and whether we can optimize our projection support.

[mp911de]

We have indeed quite an increased memory usage profile due to the recreation of DefaultMethodInvokingMethodInterceptor that holds internally a method cache.

I tweaked your example to 400.000 proxy objects resulting in about 1500MB of memory usage. After caching the DefaultMethodInvokingMethodInterceptor instance, we can get down to 320MB.

Spring Framework internally applies several recomputations therefore bulk creation can quickly get costly due to the allocation of various objects. Let me see whether there's additional potential for optimization (roughly another 20-30MB)

[i8r]

Thank you for looking into this!

I tweaked your example to 400.000 proxy objects resulting in about 1500MB of memory usage. After caching the DefaultMethodInvokingMethodInterceptor instance, we can get down to 320MB.

That's a good improvement! Still, using almost 1 KB to represent something as small as a Long is rather wasteful but much better than using 4 KB 🎉