still, those figures are quite bad.

Apparently the patch was resolving only a part of the allocations. I've looked again and I see there is another hidden allocation, in the auto-wrapping of invocations to

org.hibernate.search.query.fieldcache.impl.IntFieldLoadingStrategy.collect(int)

(that's directly invoked by the FacetCollector).

Sadly there we enter in an area of more complex refactoring which I don't think is appropriate for the stable branches, especially as these field caches are completely different in the new Lucene we can't spend too much time on it. Adam if you want to try patching it yourself and share it as a pull request on github, we'll be happy to review and integrate it.

To clarify on the Infinispan relation: the way faceting works in Lucene and how indexes are stored in Infinispan are completely independent.