On 10/17/2012 07:28 AM, Dave Cramer wrote:
> So what is the consensus on this.
>
> I would lean towards removing serial types here
I certainly want to, as I think the status quo is not only inefficient
and incorrect but it fails to even function as designed (incorrectly
calling 'integer' columns generated by a sequence 'serial' no matter
what the sequence is) and cannot easily be fixed.
The questions that must still be answered are:
- How do we tell developers about the change?
- Is a compatibility flag required?
- What might it break? Do we need to test popular ORM systems or let
their developers know?
- What does the metadata spec actually require us to supply?
The 9.2 driver has been pushed and it's a backwards-incompatible change,
so it's probably a change that's only suitable for 9.3/master. I'm
inclined to just make the change in master, push a beta driver to the
website, and see it reports come in. I can test Hibernate and
EclipseLink out, but I'm pretty sure they won't care.
BTW, I'm quite surprised that PostgreSQL doesn't expose information
about column value generation in INFORMATION_SCHEMA.COLUMNS at the
moment; the "is_identity" column is always "NO" and "is_generated" is
always "NEVER". `pg_attribute` doesn't seem to convey it either.
Shouldn't we be able to ask PostgreSQL for this information?
I'll have a look and see how the driver currently reports whether fields
are generated, are identity fields, etc.
Looking at the number of issue reports that focus on metadata, I suspect
that's a real weak point of the current driver and really needs a
review. Yay. Something for all that abundant spare time.
--
Craig Ringer