Re: Early locking option to parallel backup

On Mon, Nov 6, 2017 at 4:43 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> I wonder if we couldn't somehow repurpose the work that was done for
> parallel workers' locks. Lots of security-type issues to be handled
> if we're to open that up to clients, but maybe it's solvable. For
> instance, maybe only allowing it to clients sharing the same snapshot
> would help.

Interesting idea. There's a bunch of holes that would need to be
patched there; for instance, you can't have one session running DDL
while somebody else has AccessShareLock. Parallel query relies on the
parallel-mode restrictions to prevent that kind of thing from
happening, but it would be strange (and likely somewhat broken) to try
to enforce those here. It would be strange and probably bad if LOCK
TABLE a; LOCK TABLE b in one session and LOCK TABLE b; LOCK TABLE a in
another session failed to deadlock. In short, there's a big
difference between a single session using multiple processes and
multiple closely coordinated sessions.

Also, even if you did it, you still need a lot of PROCLOCKs. Workers
don't need to take all locks up front because they can be assured of
getting them later, but they've still got to lock the objects they
actually want to access. Group locking aims to prevent deadlocks
between cooperating processes; it is not a license to skip locking
altogether.

None of which is to say that the problems don't feel related somehow.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company