Hi Nazir, Thanks for the comparison. The main benefit of this patch series seems to be in checkpointing rather than crash recovery, so I think further testing should focus on that.
> For now, ordinary reopen failures for both relation and SLRU fsyncs > return -errno. > > Done. Relation fsync requests now use io_max_concurrency as their > limit. These two changes LGTM. > However, I think increasing the number of worker processes > should solve this problem. Increasing the worker count may help with worker availability, but it doesn't address contention for storage bandwidth. One more point: when a worker calls pg_fsync() or pg_fdatasync(), it checks its own enableFsync. If the issuer has reloaded fsync=on but the worker still has the old off value, the worker can skip the sync and report success. Should the issuer's setting or the worker's determine whether the sync is performed? Best regards, Yuhang Qiu
