On 06/01/2016 05:15 PM, Tom Lane wrote:
> Andreas Karlsson <andreas@proxel.se> writes:
>> On 06/01/2016 04:44 PM, Tom Lane wrote:
>>> I don't understand why you think you need the CREATE OR REPLACE FUNCTION
>>> commands? We only need to change proargtypes, and the updates did that.
>>> The catcache can take care of itself.
>
>> Maybe I did something wrong (I try to avoid doing manual catalog
>> updates), but when I tested it, I needed to run the CREATE OR REPLACE
>> FUNCTION command to later be able to set the parallel safety. See the
>> example below.
>
> In this particular example, the problem seems to be that you forgot to
> update pronargs; it works for me when I add "pronargs = 2" to the UPDATE.
> Your "working" example is actually creating a new function, not replacing
> the old pg_proc entry.
>
> BTW, it strikes me that the pronamespace tests in these queries are
> redundant: the OID is unique, so what matters is the search path
> in the regprocedure lookups.
Thanks, I have fixed this.
>> The reason I use to_regprocedure is so that these scripts work for
>> people who have installed the extensions in beta1 and therefore only
>> have the new signatures. If they run these scripts they would get an
>> error if I used the cast. Is it ok if these scripts break for beta1 users?
>
> Ugh. This is more of a mess than I thought. I don't like update scripts
> that might silently fail to do what they're supposed to, but leaving
> beta1 users in the lurch is not very nice either.
>
> I wonder ... could we get away with using regproc rather than
> regprocedure? These function names are probably unique anyway ...
Yeah, I would have strongly preferred to be able to use the cast. All
these functions have unique names within the core, but there is the
small risk of someone adding a user function with the same name. I do
not like either option.
The attached patch still uses to_regprocedure, but I can change to using
::regproc if that is what you prefer.
Andreas
