Revise replication logic. Currently, any error that is not NotFound leads to replication, even if the HEAD response:
- says
AccessDenied;
- could not answer at all until timeout expiration;
- *any other reason that says nothing about object presence*.
That leads to network operations that could be useless. Also if a cluster is under critical load, continuous replication made that even worse.
Maybe that is an arch issue and we should allow SN's HEADs always.