On 4 April 2017 at 22:32, Craig Ringer <craig@2ndquadrant.com> wrote:
> Hi all
>
> Here's the final set of three patches on top of what's already committed.
>
> The first is catalog_xmin logging, which is unchanged from the prior post.
>
> The 2nd is support for conflict with recovery, with changes that
> should address Andres's concerns there.
>
> The 3rd actually enables decoding on standby. Unlike the prior
> version, no attempt is made to check the walsender configuration for
> slot use, etc. The ugly code to try to mitigate races is also removed.
> Instead, if wal_level is logical the catalog_xmin sent by
> hot_standby_feedback is now the same as the xmin if there's no local
> slot holding it down. So we're always sending a catalog_xmin in
> feedback and we should always expect to have a valid local
> oldestCatalogXmin once hot_standby_feedback kicks in. This makes the
> race in slot creation no worse than the existing race between
> hot_standby_feedback establishment and the first queries run on a
> downstream, albeit with more annoying consequences. Apps can still
> ensure a slot created on standby is guaranteed safe and conflict-free
> by having a slot on the master first.
>
> I'm much happier with this. I'm still fixing some issues in the tests
> for 03 and tidying them up, but 03 should allow 01 and 02 to be
> reviewed in their proper context now.
Dammit. Attached.
--
Craig Ringer http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services
