Re: pg_partition_tree crashes for a non-defined relation

Hi,

On 2019/03/01 9:22, Michael Paquier wrote:
> On Thu, Feb 28, 2019 at 04:32:03PM -0300, Alvaro Herrera wrote:
>> Yeah, looks good, please push.
> 
> Done for this part.
> 
>> I would opt for returning the empty set for legacy inheritance too.
>>
>> More generally, I think we should return empty for anything that's
>> either not RELKIND_PARTITIONED_TABLE or has relispartition set.
> 
> I think that one option is to make the function return only the table
> itself if it is not a partitioned table, which would be more
> consistent with what pg_partition_root() does.
> 
> What I am writing next sounds perhaps a bit fancy, but in my opinion a
> normal table is itself a partition tree, made of one single member:
> itself.

That's what we discussed, but it seems that we ended up allowing regular
standalone tables (possibly in inheritance trees) only because it
*appeared* to work.  Alvaro already pointed out what appears to be a bug
in how we compute the value of level.  Instead of trying to fix it, I
agree we should just disallow tables that are not a partitioned
table/index or a partition (relispartition).  Maybe there won't be any use
cases, so we should change that while we still can.

So, maybe change the check in check_rel_can_be_partition() as follows:

    relkind = get_rel_relkind(relid);
    relispartition = get_rel_relispartition(relid);

    /* Only allow relation types that can appear in partition trees. */
    if (!relispartition &&
        relkind != RELKIND_PARTITIONED_TABLE &&
        relkind != RELKIND_PARTITIONED_INDEX)
        return false;

Thanks,
Amit