Bug #65085
Updated by Samuel Just about 2 months ago
The replica and the primary populate RepModify::last_complete and RepGather::pg_local_last_complete prior to doing PrimaryLogPG::log_operation. This causes the last_complete_ondisk value and consequently min_last_complete_ondisk to lag by an additional write cycle (mlcod on replicas will tend to lag last_update by two versions rather than one). The practical consequence of this is mainly the objects written in the last two log versions will be unavailable for replica read rather than the last one.
See https://tracker.ceph.com/issues/65013 for debugging process/discussion.