Re: [HACKERS] Performance while loading data and indexing

Bruce Momjian <pgman(at)candle(dot)pha(dot)pa(dot)us> writes:

> Doug McNaught wrote:
> > Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> writes:
> >
> > > We'd be happiest with a filesystem that journals its own metadata and
> > > not the user data in the file(s). I dunno if there are any.
> >
> > ext3 with data=writeback? (See my previous message to Bruce).
>
> OK, so that makes ext3 crash safe without lots of overhead?

Metadata is journaled so you shouldn't lose data blocks or directory
entries. Some data blocks (that haven't been fsync()'ed) may have old
or wrong data in them, but I think that's the same as ufs, right? And
WAL replay should take care of that.

It'd be very interesting to do some tests of the various journaling
modes. I have an old K6 that I might be able to turn into a
hit-the-reset-switch-at-ramdom-times machine. What kind of tests
should be run?

-Doug