Re: Compression of full-page-writes

On 2015-01-02 12:06:33 -0500, Bruce Momjian wrote:
> On Fri, Jan 2, 2015 at 05:55:52PM +0100, Andres Freund wrote:
> > On 2015-01-02 11:52:42 -0500, Bruce Momjian wrote:
> > > Why are we not seeing the 33% compression and 15% performance
> > > improvement he saw? What am I missing here?
> >
> > To see performance improvements something needs to be the bottleneck. If
> > WAL writes/flushes aren't that in the tested scenario, you won't see a
> > performance benefit. Amdahl's law and all that.
> >
> > I don't understand your negativity about the topic.
>
> I remember the initial post from Masao in August 2013 showing a
> performance boost, so I assumed, while we had the concurrent WAL insert
> performance improvement in 9.4, this was going to be our 9.5 WAL
> improvement.

I don't think it makes sense to compare features/improvements that way.

> While the WAL insert performance improvement required no tuning and
> was never a negative

It's actually a negative in some cases.

> , I now see the compression patch as something that has negatives, so
> has to be set by the user, and only wins in certain cases. I am
> disappointed, and am trying to figure out how this became such a
> marginal win for 9.5. :-(

I find the notion that a multi digit space reduction is a "marginal win"
pretty ridiculous and way too narrow focused. Our WAL volume is a
*significant* problem in the field. And it mostly consists out of FPWs
spacewise.

> My negativity is not that I don't want it, but I want to understand why
> it isn't better than I remembered. You are basically telling me it was
> always a marginal win. :-( Boohoo!

No, I didn't. I told you that *IN ONE BENCHMARK* wal writes apparently
are not the bottleneck.

Greetings,

Andres Freund

--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services