On Mon, Nov 14, 2016 at 4:42 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> No, I'm not at all proposing to assume that. But I may be willing to
>> assume that "we don't hold a CatalogSnapshot between Bind and Execute
>> unless we're also holding a transaction snapshot". I need to do a bit
>> more research to see if that's actually true, though.
>
> Turns out it's not true.
>
> I still don't much like having the main loop in PostgresMain know about
> this hack, so I experimented with putting the InvalidateCatalogSnapshot()
> calls into places in postgres.c that were already dealing with transaction
> commit/abort or snapshot management. I ended up needing five such calls
> (as in the attached patch), which is just about equally ugly. So at this
> point I'm inclined to hold my nose and stick a call into step (1) of the
> main loop instead.
Seems like a good idea.
> Also, wherever we end up putting those calls, is it worth providing a
> variant invalidation function that only kills the catalog snapshot when
> it's the only one outstanding? (If it isn't, the transaction snapshot
> should be older, so there's no chance of advancing our xmin by killing
> it.) In principle this would save some catalog snapshot rebuilds for
> inside-a-transaction-block cases, but I'm not sure it's worth sweating
> that when we're doing client message exchange anyway.
I think that would be a fairly worthwhile thing to do.
> Lastly, I find myself disliking the separate CatalogSnapshotStale flag
> variable. The other special snapshots in snapmgr.c are managed by setting
> the pointer to NULL when it's not valid, so I wonder why CatalogSnapshot
> wasn't done that way. Since this patch is touching almost every use of
> that flag already, it wouldn't take much to switch it over.
I think I had some reason why I did it that way, but I don't think it
was anything important, so I don't object to you revising it.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company