On Mon, Dec 10, 2012 at 08:32:53AM -0500, Stephen Frost wrote:
> * Noah Misch (noah@leadboat.com) wrote:
> > On Fri, Dec 07, 2012 at 06:51:18PM -0500, Stephen Frost wrote:
> > > Now, what I've honestly been hoping for on this thread was for someone
> > > to speak up and point out why I'm wrong about this concern and explain
> > > how this patch addresses that issue. If that's been done, I've missed
> > > it..
>
> [...]
>
> So, apparently I'm not wrong about my concern, but no one seems to share
> my feelings on this change.
>
> I continue to hold that this could end up being a slippery slope for us
> to go down wrt 'correctness' vs. 'do whatever the user wants'.
I agree we should be reticent to compromise correctness for convenience.
Compromising mere bug-compatibility, trading one incorrect behavior for
another incorrect behavior, is not as bad. Furthermore, today's behavior in
question is not something I can see applications deliberately and successfully
relying upon.
> If we
> keep this to only COPY and where the table has to be truncated/created
> in the same transaction (which requires the user to have sufficient
> privileges to do non-MVCC-safe things on the table already), perhaps
> it's alright.
Extending it to cases not involving a just-created or just-truncated table
really would compromise correctness; errors could leave the table in an
otherwise-impossible state. Let's indeed not go there.
I see no particular need to restrict this to COPY; that's just the most
important case by far. As a side note, the calculus for whether to extend the
optimization to INSERT and UPDATE differs from that for WAL avoidance. WAL
avoidance can be a substantial loss when the total change is small. For
pre-hinting/freezing, the worst case is having needlessly checked a few local
variables to rule out the optimization.
> It'll definitely reduce the interest in finding a real
> solution though, which is unfortunate.
That effect seems likely, but I do not find it unfortunate. The change
variant I have advocated does not stand in contrast to some "real solution" to
PostgreSQL's treatment for readers of tables created or truncated by a
transaction not in the reader's snapshot. The two topics interact at arm's
length. Bundling them into one patch, artificially making them to stand or
fall as a pair, is not a win for PostgreSQL.
That does raise another disadvantage of making the change syntax-controlled:
if we someday implement the other improvement, COPY FREEZE will have minimal
reason not to be the default. FREEZE then becomes a relic noise word.
Thanks,
nm