Re: find_inheritance_children() and ALTER TABLE NO INHERIT

On 2015/12/03 15:30, Amit Langote wrote:
> On 2015/12/03 13:09, Tom Lane wrote:
>> Amit Langote <Langote_Amit_f8(at)lab(dot)ntt(dot)co(dot)jp> writes:
>>> Currently find_inheritance_children() is smart enough to skip a child
>>> table that it finds has been dropped concurrently after it gets a lock on
>>> the same. It does so by looking up the child relid in syscache. It seems
>>> it should also check if the table is still in the list of children of the
>>> parent.
>
>> I wonder whether we could improve matters by rechecking validity of the
>> pg_inherits tuple (which we saw already and could presumably retain the
>> TID of). There is at least one place where we do something like that now,
>> IIRC.
>
> Not sure whether sane but how about performing ordered scan on pg_inherits
> (systable_getnext_ordered())and using systable_recheck_tuple() in step
> with it? Does using ordered catalog scan ensure safety against deadlocks
> that the existing approach of ordered locking of child tables does?
> Perhaps I'm missing something.

Just leaving here a patch that does this. It still returns the list in
order determined by qsort(), IOW, not in the pg_inherits_parent_index
order to avoid broken tests. I could not figure how to do it the way you
suggested.

Thanks,
Amit