On Thu, Feb 15, 2018 at 4:18 PM, Ashutosh Bapat
<ashutosh.bapat@enterprisedb.com> wrote:
> I happened to look at the patch for something else. But here are some
> comments. If any of those have been already discussed, please feel
> free to ignore. I have gone through the thread cursorily, so I might
> have missed something.
>
> In grouping_planner() we call query_planner() first which builds the
> join tree and creates paths, then calculate the target for scan/join
> rel which is applied on the topmost scan join rel. I am wondering
> whether we can reverse this order to calculate the target list of
> scan/join relation and pass it to standard_join_search() (or the hook)
> through query_planner().
>
I think the reason for doing in that order is that we can't compute
target's width till after query_planner(). See the below note in
code:
/*
* Convert the query's result tlist into PathTarget format.
*
* Note: it's desirable to not do this till after query_planner(),
* because the target width estimates can use per-Var width numbers
* that were obtained within query_planner().
*/
final_target = create_pathtarget(root, tlist);
Now, I think we can try to juggle the code in a way that the width can
be computed later, but the rest can be done earlier. However, I think
that will be somewhat major change and still won't address all kind of
cases (like for ordered paths) unless we can try to get all kind of
targets pushed down in the call stack. Also, I am not sure if that is
the only reason or there are some other assumptions about this calling
order as well.
>
> Here are some comments on the patch.
>
Thanks for looking into the patch. As of now, we are evaluating the
right approach for this patch, so let's wait for Robert's reply.
After we agree on the approach, I will address your comments.
--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com
