Re: ALTER TABLE lock strength reduction patch is unsafe Reply-To:

On 2014-03-04 11:40:10 -0500, Tom Lane wrote:
> Robert Haas <robertmhaas(at)gmail(dot)com> writes: > I think this is all too
> late for 9.4, though.
>
> I agree with the feeling that a meaningful fix for pg_dump isn't going
> to get done for 9.4. So that leaves us with the alternatives of (1)
> put off the lock-strength-reduction patch for another year; (2) push
> it anyway and accept a reduction in pg_dump reliability.
>
> I don't care for (2). I'd like to have lock strength reduction as
> much as anybody, but it can't come at the price of reduction of
> reliability.

I am sorry, but I think this is vastly overstating the scope of the
pg_dump problem. CREATE INDEX *already* doesn't require a AEL, and the
amount of problems that has caused in the past is surprisingly low. If
such a frequently used command doesn't cause problems, why are you
assuming other commands to be that problematic? And I think it's hard to
argue that the proposed changes are more likely to cause problems.

Let's try to go at this a bit more methodically. The commands that -
afaics - change their locklevel due to latest patch (v21) are:

01) ALTER TABLE .. ADD CONSTRAINT
02) ALTER TABLE .. ADD CONSTRAINT ... USING
03) ALTER TABLE .. ENABLE | DISABLE [ REPLICA | ALWAYS ] TRIGGER [ ALL ]
04) ALTER TABLE .. ALTER CONSTRAINT
05) ALTER TABLE .. REPLICA IDENTITY
06) ALTER TABLE .. ALTER COLUMN .. SET NOT NULL (*not* DROP NULL)
cmd_lockmode = ShareRowExclusiveLock;

07) ALTER TABLE ... ALTER COLUMN ... SET STATISTICS
08) ALTER TABLE ... CLUSTER ON ...
09) ALTER TABLE ... SET WITHOUT CLUSTER
10) ALTER TABLE ... SET (...)
11) ALTER TABLE ... RESET (...)
12) ALTER TABLE ... ALTER COLUMN ... SET (...)
13) ALTER TABLE ... ALTER COLUMN ... RESET (...)
14) ALTER TABLE ... VALIDATE CONSTRAINT constraint_name
cmd_lockmode = ShareUpdateExclusiveLock;

I have my reservations about ADD CONSTRAINT (including SET NOT NULL)
being unproblematic (mostly because I haven't thought through possible
consquences for the planner making different choices with added
constraints).

From the perspective of pg_dump consistency, except ADD CONSTRAINT, none
of those seem to have graver possible consequences than CREATE INDEX
(and DROP INDEX CONCURRENTLY) already being unsafe.

In fact all of those should actually end up being *safer*, even ending
up always being dumped consistently since they are all reconstructed
clientside by pg_dump.
You argue elsewhere that that's a fragile coincidence. But so what, even
if it changes, the consequences still are going to be *far* less
significant than missing various index, trigger, and whatnot commands.

I think the set of problems you mention are going to be really important
when we someday get around to make stuff like ALTER TABLE ... ADD/DROP
COLUMN require lower lock levels, but that's not what's proposed.

Greetings,

Andres Freund

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