Re: Support UPDATE table SET(*)=...

Andrew Gierth <andrew(at)tao11(dot)riddles(dot)org(dot)uk> writes:
> "Tom" == Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> writes:
> Tom> and not at all robust against schema changes (the same problem as
> Tom> with the patch's approach to UPDATE).

> Now this I think is wrong; I think it's just as robust against schema
> changes as using the composite value directly would be. Consider the
> case where foo and reference_foo match with the exception of dropped
> columns; the code as it is should just work, while a variant that used
> the composite values would have to explicitly deal with that.

AFAICS you've got that backwards.

As I illustrated upthread, after parse-time expansion we would have a
command that explicitly tells the system to insert or update only the
enumerated columns. That will *not* work as desired if columns are added
later, and (if it's in a rule) I would expect the system to have to fail
a DROP command trying to drop one of those named columns, the same way
you can't drop a column referenced in a view/rule today.

On the other hand, if it's a composite-type assignment, then dealing
with things like dropped columns becomes the system's responsibility,
and we already have code for that sort of thing.

> ... The alternative of
> set * = populate_record(foo, new_values)
> is less satisfactory since it introduces inconsistencies with the case
> where you _do_ want to specify explicit columns, unless you also allow
> set (a,b) = row_value
> which is required by the spec for T641 but which we don't currently
> have.

Right, but we should be trying to move in that direction. I see your
point though that (*) is more notationally consistent with that case.
Maybe we should be looking at trying to implement T641 in full and then
including (*) as a special case of that.

Anyway, my core point here is that we should avoid parse-time expansion
of the column set. The *only* reason for doing that is that it makes the
patch simpler, not that there is some functional advantage; in fact there
are considerable functional disadvantages as we've just been through.
So I do not want to commit a patch that institutionalizes those
disadvantages just for short-term implementation simplicity.

regards, tom lane