Re: partition routing layering in nodeModifyTable.c

Hi,

On Sat, Aug 3, 2019 at 3:01 AM Andres Freund <andres@anarazel.de> wrote:
> On 2019-07-31 17:04:38 +0900, Amit Langote wrote:
> > I looked into trying to do the things I mentioned above and it seems
> > to me that revising BeginDirectModify()'s API to receive the
> > ResultRelInfo directly as Andres suggested might be the best way
> > forward.  I've implemented that in the attached 0001.

> Fujita-san, do you have any comments on the FDW API change?  Or anybody
> else?
>
> I'm a bit woried about the move of BeginDirectModify() into
> nodeModifyTable.c - it just seems like an odd control flow to me. Not
> allowing any intermittent nodes between ForeignScan and ModifyTable also
> seems like an undesirable restriction for the future. I realize that we
> already do that for BeginForeignModify() (just btw, that already accepts
> resultRelInfo as a parameter, so being symmetrical for BeginDirectModify
> makes sense), but it still seems like the wrong direction to me.
>
> The need for that move, I assume, comes from needing knowing the correct
> ResultRelInfo, correct?  I wonder if we shouldn't instead determine the
> at plan time (in setrefs.c), somewhat similar to how we determine
> ModifyTable.resultRelIndex. Doesn't look like that'd be too hard?

I'd vote for that; I created a patch for that [1].

> Then we could just have BeginForeignModify, BeginDirectModify,
> BeginForeignScan all be called from ExecInitForeignScan().

I think so too.

Best regards,
Etsuro Fujita

[1] https://www.postgresql.org/message-id/CAPmGK15%3DoFHmWNND5vopfokSGfn6jMXVvnHa7K7P49F7k1hWPQ%40mail.gmail.com