Re: pgsql: Avoid archiving XLOG_RUNNING_XACTS on idle server

On 2016-04-04 08:44:47 +0100, Simon Riggs wrote:
> On 4 April 2016 at 08:22, Andres Freund <andres(at)anarazel(dot)de> wrote:
> > No, because in the alternative proposal we determine that the system
> > indeed has been idle since the last time a WAL record was logged.

> Why would that change anything? If something aborts without writing an
> abort record, the system can easily be idle afterwards. So an idle system
> means nothing. Your objection applies to *all* cases, not just for this
> patch and if we revert, we block all cases.

The previous suggestion only considered a system idle if there's been a
previous snapshot logged (addressing the issue at hand) and no further
WAL record, but a select few excepted, have been generated since the
last logged WAL record.

Thus e.g. such an abort would only be ignored if it never persisted it
xid anywhere (as that'd be WAL logging), in which case there cannot be
an effect for HS.

>
> I haven't ignored any messages, regrettably I read them all. I've committed
> a much simpler patch, which is what committers do.

Unless it's completely obviously an improvement, most also send a
revised version of the patch for review/comments.

> That patch does exactly the same thing as the patch you prefer, just
> does it differently;

No, it doesn't; as explained above.

Greetings,

Andres Freund