Re: Stopping logical replication protocol

On 3 Oct. 2016 10:15, "Michael Paquier" <michael(dot)paquier(at)gmail(dot)com> wrote:
>
> On Tue, Sep 27, 2016 at 11:05 AM, Craig Ringer <craig(at)2ndquadrant(dot)com>
wrote:
> > On 26 September 2016 at 21:52, Vladimir Gordiychuk <folyga(at)gmail(dot)com>
wrote:
> >>>You should rely on time I tests as little as possible. Some of the test
> >>> hosts are VERY slow. If possible something deterministic should be
used.
> >>
> >> That's why this changes difficult to verify. Maybe in that case we
should
> >> write benchmark but not integration test?
> >> In that case we can say, before this changes stopping logical
replication
> >> gets N ms but after apply changes it gets NN ms where NN ms less than
N ms.
> >> Is it available add this kind of benchmark to postgresql? I will be
grateful
> >> for links.
> >
> > It's for that reason that I added a message printed only in verbose
> > mode that pg_recvlogical emits when it's exiting after a
> > client-initiated copydone.
> >
> > You can use the TAP tests, print diag messages, etc. But we generally
> > want them to run fairly quickly, so big benchmark runs aren't
> > desirable. You'll notice that I left diag messages in to report the
> > timing for the results in your tests, I just changed the tests so they
> > didn't depend on very tight timing for success/failure anymore.
> >
> > We don't currently have any automated benchmarking infrastructure.
>
> Which seems like this patch is not complete yet.

Personally I think it is. I'm just explaining why I adjusted Vladimir's
tests to be less timing sensitive and rely on a qualitative property
instead.

That said, it's had recent change and it isn't a big intrusive change that
really need attention this cf.

> I am marking it as
> returned with feedback, but it may be a better idea to move it to next
> CF once a new version with updated tests shows up.

I'll move it now. I think the tests are fine, if Vladimir agrees, so IMO
it's ready to go. More eyes wouldn't hurt though.

If Vladimir wants benchmarking based tests that's a whole separate project
IMO. Something that will work robustly on the weird slow machines we have
isn't simple. Probably a new buildfarm option etc since we won't want to
clutter the really slow old niche boxes with it.

Vladimir, can I have your opinion on the latest patch or if you want to
make changes, an updated patch?