Re: Odd system-column handling in postgres_fdw join pushdown patch

On 2016/04/14 4:57, Robert Haas wrote:
> On Wed, Apr 13, 2016 at 2:36 PM, Robert Haas <robertmhaas@gmail.com> wrote:
>> On Wed, Apr 13, 2016 at 2:11 PM, Robert Haas <robertmhaas@gmail.com> wrote:
>>> So, clearly that's not good.  It should at least be consistent.  But
>>> more than that, the fact that postgres_fdw sets the xmax to 0xffffffff
>>> is also pretty wacky.  We might use such a value as a sentinel for
>>> some data type, but for transaction IDs that's just some random normal
>>> transaction ID, and it's NOT coming from t1.  I haven't tracked down
>>> where it *is* coming from yet, but can't imagine it's any place very
>>> principled.
>>
>> And, yeah, it's not very principled.
>>
>> rhaas=# select ft1.xmin, ft1.xmax, ft1.cmin from ft1;
>>   xmin |    xmax    | cmin
>> ------+------------+-------
>>     96 | 4294967295 | 16392
>>     96 | 4294967295 | 16392
>>     96 | 4294967295 | 16392
>>     96 | 4294967295 | 16392
>> (4 rows)
>>
>> What's happening here is that heap_getattr() is being applied to a
>> HeapTupleHeaderData which contains DatumTupleFields.  So 96 is
>> datum_len_, 4294967295 is the -1 recorded in datum_typmod, and 16392
>> is the compose type OID recorded in datum_typeid, which happens in
>> this case to be the OID of ft1.  Isn't that special?
>>
>> It's hard for me to view this as anything other than a bug in
>> postgres_fdw - which of course means that this open item boils down to
>> the complaint that the way system columns are handled by join pushdown
>> isn't bug-compatible with the existing behavior....

> OK, here's a patch.  What I did is:

Thank you for taking care of this.

> 1. For a regular FDW scan, zero the xmin, xmax, and cid of the tuple
> before returning it from postgres_fdw, so that we don't expose the
> datum-tuple fields.   I can't see any reason this isn't safe, but I
> might be missing something.

I'm not sure that is really safe.

> 2. When a join is pushed down, deparse system columns using something
> like "CASE WHEN r1.* IS NOT NULL THEN 0 END", except for the table OID
> column, which gets deparsed with the table OID in place of 0.  This
> delivers the correct behavior in the presence of outer joins.

I think that that would cause useless data transfer for such culumns. 
Why not set values locally for such columns?

Best regards,
Etsuro Fujita