Re: [HACKERS] Performance while loading data and indexing

Doug McNaught <doug(at)wireboard(dot)com> writes:
> "data=writeback" means that no data is journaled, just metadata (which
> is like XFS or Reiser). An fsync() call should still do what it
> normally does, commit the writes to disk before returning.
> "data=journal" journals all data and is the slowest and safest.
> "data=ordered" writes out data blocks before committing a journal
> transaction, which is faster than full data journaling (since data
> doesn't get written twice) and almost as safe. "data=writeback" is
> noted to keep obsolete data in the case of some crashes (since the
> data may not have been written yet) but a completed fsync() should
> ensure that the data is valid.

Thanks for the explanation.

> So I guess I'd probably use data=ordered for an all-on-one-fs
> installation, and data=writeback for a WAL-only drive.

Actually I think the ideal thing for Postgres would be data=writeback
for both data and WAL drives. We can handle loss of un-fsync'd data
for ourselves in both cases.

Of course, if you keep anything besides Postgres data files on a
partition, you'd possibly want the more secure settings.

regards, tom lane