arielshaqed commented 1 month ago

This is an experimental feature. When reviewing please comment about its usability, but focus mainly on correctness.

The code in the controller re-uses an existing physical address, which is supposedly a no-no. However this case is fine: the physical address is used by the same object being updated. So the only way for UGC to remove it would be to lose these races:

Thread 1 gets object metadata
Thread 2 deletes the object
👉🏽 Thread 3 performs UGC and cleans up the physical object
Thread 1 writes a new entry for the object with the same physical address Thread 3 is UGC so it only cleans up objects that were abandoned several hours ago. Our UGC model allows us to lose the object in this (extreme) case - this is not a new lossage.

Fixes #8251.

arielshaqed commented 1 month ago

@ozkatz for API

(@N-o-Z for API and correctness and code and ..., of course)