Hi,
On 2021-01-04 19:11:43 +0100, Michael Banck wrote:
> Am Samstag, den 02.01.2021, 10:47 -0500 schrieb Stephen Frost:
> > * Michael Paquier (michael@paquier.xyz) wrote:
> > > On Fri, Jan 01, 2021 at 08:34:34PM +0100, Michael Banck wrote:
> > > > I think enough people use data checksums these days that it warrants to
> > > > be moved into the "normal part", like in the attached.
> > >
> > > +1. Let's see first what others think about this change.
> >
> > I agree with this, but I'd also like to propose, again, as has been
> > discussed a few times, making it the default too.
FWIW, I am quite doubtful we're there performance-wise. Besides the WAL
logging overhead, the copy we do via PageSetChecksumCopy() shows up
quite significantly in profiles here. Together with the checksums
computation that's *halfing* write throughput on fast drives in my aio
branch.
> This looks much better from the WAL size perspective, there's now almost
> no additional WAL. However, that is because pgbench doesn't do TOAST, so
> in a real-world example it might still be quite larger. Also, the vacuum
> runtime is still 15x longer.
That's obviously an issue.
> So maybe we should switch on wal_compression if we enable data checksums
> by default.
It unfortunately also hurts other workloads. If we moved towards a saner
compression algorithm that'd perhaps not be an issue anymore...
Greetings,
Andres Freund
