We have identified an issue where concurrent commits would cause latency for readers on the followers. We would like to address this by optimizing the 2-phase commit protocol for the RWStore to identify the allocators that were modified on the leader. Currently, the followers will read all the allocators from the disk after a 2-phase commit. We would like to modify this so the followers can do the minimum work at a commit, reading only those allocators that have changed. Those allocators should all be in local cache on the followers. Thus, this should remove the IO costs on the followers to "catch up" with the updated allocators. Addressing this might take an additional week, but I think that it is worth the additional time to remove this overhead on the followers.
The javadoc on IHABufferStrategy#resetFromHARootBlock() flags some issues related to the overloaded use of that method. Update that javadoc and review callers of resetFromHARootBlock() when addressing this ticket.