Re: Prevent internal error at concurrent CREATE OR REPLACE FUNCTION

On Tue, 27 May 2025 08:33:42 +0200
Jim Jones <jim.jones@uni-muenster.de> wrote:

> Hi Yugo
> 
> On 26.05.25 18:39, Yugo Nagata wrote:
> > I can see the error when two concurrent transactions issue
> > "alter function f() immutable".
> 
> 
> I might have missed something in my last tests... I could now reproduce
> the behaviour you mentioned.
> 
> I've tested v2 and it works as described. CREATE OR REPLACE FUNCTION and
> ALTER TABLE no longer raise an error after the lock by the concurrent
> transaction was freed.
> 
> One quick question in v2-002:
> 
>      tup = SearchSysCacheCopy1(PROCOID, ObjectIdGetDatum(funcOid));
> -    if (!HeapTupleIsValid(tup)) /* should not happen */
> -        elog(ERROR, "cache lookup failed for function %u", funcOid);
> +    if (!HeapTupleIsValid(tup))
> +        ereport(ERROR,
> +                errcode(ERRCODE_UNDEFINED_OBJECT),
> +                errmsg("function \"%s\" does not exist",
> +                       NameListToString(stmt->func->objname)));
> 
> 
> Is it really ok to change this error message here? Did the addition of
> LockDatabaseObject change the semantics of the previous message? 

Yes. AcceptInvalidationMessages() is called in LockDatabaseObject() after wait,
and this enables the detection of object deletion during the wait.

Regards,
Yugo Nagata

-- 
Yugo Nagata <nagata@sraoss.co.jp>