Re: Catalog Access (was: [GENERAL] Concurrency problem building indexes)

Martijn van Oosterhout <kleptog(at)svana(dot)org> writes:
> I think the basic problem is that DDL can't really work within a
> transaction. If I do an ALTER TABLE, some of these changes need to show
> up to concurrent transactions (maybe creating a unique index?).

The point is that DDL can't be MVCC. If for instance you add an index
to a table, once you commit every other transaction must *immediately*
start updating that index when they modify the table. They can't use
the excuse of "not my problem because the catalog change postdates the
snapshot I'm using". The drop-index case is even worse, since a
transaction that believes the index is still present is likely to try
to access/update a disk file that's not there anymore. Adding/dropping
columns, constraints, triggers, etc all have hazards of the same ilk.

> I think it's like Tom says in that email, it could be done, but the
> cost/benefit ratio isn't very good...

It's barely possible that we could make this happen, but it would be a
huge amount of work, and probably a huge amount of instability for a
very long time until we'd gotten all the corner cases sorted. I think
there are much more productive uses for our development effort.

regards, tom lane