Re: Function scan FDW pushdown

Ashutosh Bapat писал 2021-06-15 16:15:
> Hi Alexander,

Hi.

The current version of the patch is based on asymetric partition-wise 
join.
Currently it is applied after 
v19-0001-Asymmetric-partitionwise-join.patch from
on 
https://www.postgresql.org/message-id/[email protected] 
.

>> So far I don't know how to visualize actual function expression used 
>> in
>> function RTE, as in postgresExplainForeignScan() es->rtable comes from
>> queryDesc->plannedstmt->rtable, and rte->functions is already 0.
> 
> The actual function expression will be part of the Remote SQL of
> ForeignScan node so no need to visualize it separately.

We still need to create tuple description for functions in 
get_tupdesc_for_join_scan_tuples(),
so I had to remove setting newrte->functions to NIL in 
add_rte_to_flat_rtable().
With rte->functions in place, there's no issues for explain.

> 
> The patch will have problems when there are multiple foreign tables
> all on different servers or use different FDWs. In such a case the
> function scan's RelOptInfo will get the fpinfo based on the first
> foreign table the function scan is paired with during join planning.
> But that may not be the best foreign table to join. We should be able
> to plan all the possible joins. Current infra to add one fpinfo per
> RelOptInfo won't help there. We need something better.

I suppose attached version of the patch is more mature.

> 
> The patch targets only postgres FDW, how do you see this working with
> other FDWs?

Not now. We introduce necessary APIs for other FDWs, but implementing 
TryShippableJoinPaths()
doesn't seem straightforward.

> 
> If we come up with the right approach we could use it for 1. pushing
> down queries with IN () clause 2. joining a small local table with a
> large foreign table by sending the local table rows down to the
> foreign server.


-- 
Best regards,
Alexander Pyhalov,
Postgres Professional