Well the other option is to be safer about picking the neutered CacheKeysFactory. ATM hibernate-infinispan decided to pick that less-safe one in cases where it should not be used, which is the underlying cause of a few bug reports. FIxing that actually needs no changes at all to the SPI.

Now there is a question about whether we can properly "hint" that to the provider. Really the question boils down to whether the region is shared. We could provide that information as part of the SPI. But outside of that, I think the more proper solution to that piece is for hibernate-infinispan to make the safer assumption rather than the more performant one.