On Mon, Jun 13, 2016 at 7:51 PM, Tom Lane <t...@sss.pgh.pa.us> wrote:
> Robert Haas <robertmh...@gmail.com> writes:
> > On Mon, Jun 13, 2016 at 3:18 AM, Amit Kapila <amit.kapil...@gmail.com>
> >> In create_grouping_paths(), we are building partial_grouping_path and
> >> is used for gather path and other grouping paths (for partial paths).
> >> However, we don't use it for partial path list and sort path due to
> >> path target for Sort path is different from what we have expected. Is
> >> a problem in applying partial_grouping_path for partial pathlist?
> >> Attached patch just does that and I don't see error with patch.
> > It doesn't seem like a good idea to destructive modify
> > input_rel->partial_pathlist. Applying the projection to each path
> > before using it would probably be better.
> I think the real question here is why the code removed by 04ae11f62
> was wrong. It was unsafe to use apply_projection_to_path, certainly,
> but using create_projection_path directly would have avoided the
> stated problem. And it's very unclear that this new patch doesn't
> bring back that bug in a different place.
This new patch still doesn't seem to be right, but it won't bring back the
original problem because apply_projection_to_path will be only done if
grouped_rel is parallel_safe which means it doesn't have any
parallel-unsafe or parallel-restricted clause in quals or target list.
> I am not very happy that neither 04ae11f62 nor this patch include
> any regression test case proving that a problem existed and has
> been fixed.
It is slightly tricky to write a reproducible parallel-query test, but
point taken and I think we should try to have a test unless such a test is
really time consuming.