Hannu Krosing <hannu@tm.ee> writes:
> On Wed, 2002-07-31 at 10:22, Tom Lane wrote:
>> Hm. How about
>>
>> ERROR: Cannot insert into a view
>> You need an unconditional ON INSERT DO INSTEAD rule
> Seems more accurate, but actually you may also have two or more
> conditional rules that cover all possibilities if taken together.
> Maybe
> ERROR: Cannot insert into a view
> You need an ON INSERT DO INSTEAD rule that matches your INSERT
> Which covers both cases.
Actually not: the system insists that you provide an unconditional
DO INSTEAD rule. The other would require trying to prove (during
rule expansion) a theorem that the conditions of the available
conditional rules cover all possible cases.
Alternatively we could move the test for insertion-into-a-view out of
the rewriter and into a low level of the executor, producing an error
message only if some inserted tuple actually gets past the rule
conditions. I don't much care for that answer because (a) it turns a
once-per-query overhead check into once-per-tuple overhead, and
(b) if you fail to span the full space of possibilities in your rule
conditions, you might not find out about it until your application goes
belly-up in production. There's some version of Murphy's Law that says
rare conditions arise with very low probability during testing, and very
high probability as soon as you go live...
regards, tom lane