Re: ALTER TABLE lock strength reduction patch is unsafe

Andres Freund <andres(at)2ndquadrant(dot)com> writes:
> On 2014-03-04 16:37:48 -0500, Tom Lane wrote:
>> However, it seems possible that we could have a mode in which a read-only
>> session did all its catalog fetches according to the transaction snapshot.
>> That would get us to a situation where the backend-internal lookups that
>> ruleutils relies on would give the same answers as queries done by
>> pg_dump. Robert's work on getting rid of SnapshotNow has probably moved
>> that much closer than it was before, but it's still not exactly a trivial
>> patch.

> The most interesting bit seems to be the cache invalidation handling. It
> would need to be something like PushCatalogSnapshot() which disables
> applying invals, including catchup interrupts and all. If the sinval
> queue hasn't overflown while that snapshot was up, everything is fine we
> just need to apply all pending invalidations, if it has, we need to do a
> InvalidateSystemCaches().

Yeah, at least within the transaction we would simply ignore invals.
To avoid causing sinval queue overrun (which would hurt everyone
system-wide), my inclination would be to just drop invals on the floor all
the time when in this mode, and forcibly do InvalidateSystemCaches at
transaction end. For pg_dump, at least, there is no value in working any
harder than that, since it's going to quit at transaction end anyhow.

regards, tom lane