Today with node-node replication we perform independent commits on replicas in two instances.
- When a replica receives a new commit point from primary so that it can fsync and purge from its local xlog.
- Before engine close so that if promoted as primary it can continue and replay from translog from the latest checkpoint before the engine flip.
With remote storage, all segments are durably persisted in the store. So we can skip performing these commits on replica shards. Further, the commits will bump segment generation and segments_n file, which could potentially cause segment conflicts during failover cases.