Re: [HACKERS] why not parallel seq scan for slow functions

On Tue, Jan 2, 2018 at 6:38 AM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> [ new patch ]

I think that grouping_planner() could benefit from a slightly more
extensive rearrangement. With your patch applied, the order of
operations is:

1. compute the scan/join target
2. apply the scan/join target to all paths in current_rel's pathlist
3. generate gather paths, possibly adding more stuff to current_rel's pathlist
4. rerun set_cheapest
5. apply the scan/join target, if parallel safe, to all paths in the
current rel's partial_pathlist, for the benefit of upper planning
steps
6. clear the partial pathlist if the target list is not parallel safe

I at first thought this was outright broken because step #3 imposes
the scan/join target without testing it for parallel-safety, but then
I realized that generate_gather_paths will apply that target list by
using apply_projection_to_path, which makes an is_parallel_safe test
of its own. But it doesn't seem good for step 3 to test the
parallel-safety of the target list separately for each path and then
have grouping_planner do it one more time for the benefit of upper
planning steps. Instead, I suggest that we try to get rid of the
logic in apply_projection_to_path that knows about Gather and Gather
Merge specifically. I think we can do that if grouping_planner does
this:

1. compute the scan/join target
2. apply the scan/join target, if parallel safe, to all paths in the
current rel's partial_pathlist
3. generate gather paths
4. clear the partial pathlist if the target list is not parallel safe
5. apply the scan/join target to all paths in current_rel's pathlist
6. rerun set_cheapest

That seems like a considerably more logical order of operations. It
avoids not only the expense of testing the scanjoin_target for
parallel-safety multiple times, but the ugliness of having
apply_projection_to_path know about Gather and Gather Merge as a
special case.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company