On Fri, Feb 18, 2011 at 10:02 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> If by the first option you mean causing the failure report to be "syntax
> error" rather than anything more specific, then I agree that option
> sucks. I'd actually vote for putting the error test somewhere in
> statement execution code, well downstream of gram.y. The parser's job
> is to produce a parse tree, not to encapsulate knowledge about which
> combinations of options are supported.
OK. Proposed patch attached. It looks to me like an unlogged view is
inherently nonsensical, whereas an unlogged sequence is sensible but
we don't implement it (and may never implement it, absent some
evidence that the overhead of WAL logging sequence changes is worth
getting excited about), so I wrote the error messages to reflect that
distinction. I also added a couple of matching regression tests, and
documented that UNLOGGED works with SELECT INTO. I put the check for
views in DefineView adjacent to the other check that already cares
about relpersistence, and I put the one in DefineSequence to match, at
the top for lack of any compelling theory of where it ought to go. I
looked at stuffing it all the way down into DefineRelation but that
looks like it would require some other rejiggering of existing logic
and assertions, which seems pointless and potentially prone to
breaking things.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
