Using arenas to replace allocation trickery

[jzelinskie]

Behind a flag, this proposal for memory arenas has landed in Go.
It's quite possible that leveraging this could eliminate a lot of the low-level nastiness in the z package.

I haven't done a deep dive, but it seems like it could be an improvement.

[manishrjain]

Are you talking about the use of jemalloc in z package? We have been using memory arenas since 2017 in Badger 1, so the memory arena isn't a new thing. The benefit of jemalloc is to be able to allocate many small objects and release them independently of each other. A memory arena wouldn't help with that.

In fact, even with arenas, we've seen that jemalloc does a better job of releasing memory than Go does. Let me know if you meant something else.

[manishrjain]

Closing this issue. Feel free to re-open if needed.