Re: ALTER TABLE lock strength reduction patch is unsafe

On 2014-03-03 19:15:27 -0500, Tom Lane wrote:
> Noah Misch <noah(at)leadboat(dot)com> writes:
> > Just to be clear, that list is not a commentary on the particular patch at
> > hand. Those are merely the kinds of regressions to look for in a patch
> > affecting this area of the code.
>
> A complaint on pgsql-bugs just now reminded me of a specific area that
> needs to be looked at hard: how bad are the implications for pg_dump?
>
> Up to now, pg_dump could be reasonably confident that once it had
> AccessShareLock on every table it intended to dump, there would be no
> schema changes happening on those tables until it got done.

The guarantee wasn't actually that strong. It already was quite possible
that indexes got created/dropped during that time, which probably is the
by far most frequent DDL run in production.

> This greatly
> ameliorates the snapshot-skew problems that arise from its habit of doing
> some things for itself and other things via backend-internal functions
> (which historically used SnapshotNow and now use a fresh MVCC snapshot,
> either way potentially quite newer than the transaction snapshot pg_dump's
> own queries will use).

Yea, I wonder if we shouldn't start to make them use a different
snapshot. It's the pg_get_*def() functions, or is there something else?

Afair (I really haven't rechecked) all the actions that have a changed
locklevels affect things that pg_dump recreates clientside, using a
repeatable read snapshot, so there shouldn't be much change there?

Greetings,

Andres Freund

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