Robert Haas <robertmhaas@gmail.com> writes:
> On Thu, Apr 14, 2022 at 1:36 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Not really following? ISTM the point of ffaa44cb5 was that if we don't
>> include the CatalogSnapshot in our advertised xmin, then the length
>> of time we can safely use it is *zero*.
> No, I don't think so. I'm proposing that you shouldn't be taking a
> catalog snapshot unless you already hold some other snapshot, and that
> the catalog snapshot should always be newer than whatever that other
> snapshot is. If we made that be true, then your catalog snapshot
> couldn't ever be the thing holding back xmin -- and then I think it
> doesn't need to be registered.
If you don't register it, then you need to also make sure that it's
destroyed whenever that older snapshot is. Which I think would require
either a lot of useless/inefficient CatalogSnapshot destructions, or
infrastructure that's more or less isomorphic to the RegisteredSnapshot
heap.
> That is exactly my argument, but I'm not actually sure whether it is
> in fact better. I was responding to Andres's statement that
> CatalogSnapshot was hiding a lot of bugs because it makes it look like
> we have a snapshot when we don't really.
Well, we DO have a snapshot, and it is 100% perfectly safe to use, if it's
registered. Andres' complaint is that that snapshot might get invalidated
when you weren't expecting it, but I'm not really convinced that we have
all that many bugs of that ilk. Wouldn't CLOBBER_CACHE_ALWAYS testing
find them?
regards, tom lane
