Re: Declarative partitioning - another take

On Tue, Dec 27, 2016 at 3:59 AM, Amit Langote
<Langote_Amit_f8(at)lab(dot)ntt(dot)co(dot)jp> wrote:
> Patches 0001 to 0006 unchanged.

Committed 0001 earlier, as mentioned in a separate email. Committed
0002 and part of 0003. But I'm skeptical that the as-patched-by-0003
logic in generate_partition_qual() makes sense. You do this:

result = list_concat(generate_partition_qual(parent),
copyObject(rel->rd_partcheck));

/* Mark Vars with correct attnos */
result = map_partition_varattnos(result, rel, parent);

But that has the effect of applying map_partition_varattnos to
everything in rel->rd_partcheck in addition to applying it to
everything returned by generate_partition_qual() on the parent, which
doesn't seem right.

Also, don't we want to do map_partition_varattnos() just ONCE, rather
than on every call to this function? I think maybe your concern is
that the parent might be changed without a relcache flush on the
child, but I don't quite see how that could happen. If the parent's
tuple descriptor changes, surely the child's tuple descriptor would
have to be altered at the same time...

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company