Re: Datatypes and performance

Tom Lane wrote:
> "Mattias Kregert" <mattias(at)kregert(dot)se> writes:
> > [ misguided analysis ]
>
> > Journalling FS will fix the FS problems, so the files are ok.
> > PG journal will fix the PG problems so the tables will be ok.
>
> Only if the journal is all down to disk before the crash.
>
> The fundamental problem with fsync off is that it's likely to violate
> the WAL principle (write journal entry before data entry it describes).
> If you then crash, you have data entries that correspond to transactions
> that should not have been committed (because WAL replay won't guarantee
> recovering all of the transaction's effects). In other words, corrupt
> data.
>
> If we had a less costly way of guaranteeing write order than fsync, we'd
> use it, but there is no other portable method.

Uh oh... i thought the journal was always synced, and that the fsync option only affected table writes... :(

If I turn fsync on and then pull the power cord while a number of clients are doing lots of inserts/updates and stuff, will the fsync then guarantee that no data will be lost or corrupted?

/* m */