On 08/31/2016 09:20 PM, Vik Fearing wrote:
> On 08/24/2016 06:16 PM, Robert Haas wrote:
>> On Tue, Aug 23, 2016 at 6:11 PM, Tomas Vondra
>> <tomas.vondra@2ndquadrant.com> wrote:
>>> Could someone please explain how the unlogged tables are supposed to fix the
>>> catalog bloat problem, as stated in the initial patch submission? We'd still
>>> need to insert/delete the catalog rows when creating/dropping the temporary
>>> tables, causing the bloat. Or is there something I'm missing?
>>
>> No, not really. Jim just asked if the idea of partitioning the
>> columns was completely dead in the water, and I said, no, you could
>> theoretically salvage it. Whether that does you much good is another
>> question.
>>
>> IMV, the point here is that you MUST have globally visible dependency
>> entries for this to work sanely. If they're not in a catalog, they
>> have to be someplace else, and backend-private memory isn't good
>> enough, because that's not globally visible. Until we've got a
>> strategy for that problem, this whole effort is going nowhere - even
>> though in other respects it may be a terrific idea.
>
> Why not just have a regular-looking table, with a "global temporary"
> relpersistence (I don't care which letter it gets) and when a backend
> tries to access it, it uses its own private relfilenode instead of
> whatever is in pg_class, creating one if necessary. That way the
> structure of the table is fixed, with all the dependencies and whatnot,
> but the content is private to each backend. What's wrong with this idea?
>
It's an improvement (and it's pretty much exactly what I proposed
upthread). But it does not solve the problems with pg_statistic for
example (each backend needs it's own statistics. So we'd either bloat
the pg_statistic (if we manage to solve the problem that the table has
the same oid in all backends), or we would need in-memory tuples (just
like discussed in the thread so far).
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services