On 2016/12/21 21:44, Etsuro Fujita wrote:
> On 2016/12/20 0:37, Tom Lane wrote:
>> Etsuro Fujita <fujita.etsuro@lab.ntt.co.jp> writes:
>>> On 2016/12/17 1:13, Tom Lane wrote:
>>>> So I think the rule could be
>>>> "When first asked to produce a path for a given foreign joinrel,
>>>> collect
>>>> the cheapest paths for its left and right inputs, and make a
>>>> nestloop path
>>>> (or hashjoin path, if full join) from those, using the join quals
>>>> needed
>>>> for the current input relation pair.
>>> Seems reasonable.
>>>> Use this as the fdw_outerpath for
>>>> all foreign paths made for the joinrel."
>>> I'm not sure that would work well for foreign joins with sort orders.
>>> Consider a merge join, whose left input is a 2-way foreign join with a
>>> sort order that implements a full join and whose right input is a sorted
>>> local table scan. If the EPQ subplan for the foreign join wouldn't
>>> produce the right sort order, the merge join might break during EPQ
>>> rechecks (note that in this case the EPQ subplan for the foreign join
>>> might produce more than a single row during an EPQ recheck).
>> How so? We only recheck one row at a time, therefore it can be
>> claimed to
>> have any sort order you care about.
> I'll have second thoughts about that.
I noticed I was wrong and you are right. Sorry for the noise.
Best regards,
Etsuro Fujita