On Mon, Aug 21, 2017 at 2:43 PM, Robert Haas <robertmhaas@gmail.com> wrote:
> Works for me. While I'm sure this won't eclipse previous achievements
> in this area, it still seems worthwhile. So committed with one minor
> change to shorten a long-ish line.
Buildfarm members with force_parallel_mode=regress are failing now. I
haven't had a chance to investigate exactly what's going on here, but
I think there are probably several issues:
1. It's definitely the case that the details about a sort operation
aren't propagated from a worker back to the leader. This has elicited
complaint previously.
2. pass_down_bound() probably gets the Gather node at the top of the
tree and stops, since it doesn't know how to pass down a bound through
a Gather. We could probably teach it to pass down the bound through
Gather or Gather Merge on the same theory as Append/MergeAppend.
3. However, even if it did that, it would only affect the leader, not
the workers, because each worker will of course have a separate copy
of each executor state node. We could fix that by having the Gather
or Gather Merge node also store the bound and propagate that to each
worker via DSM, and then have each worker call pass_down_bound itself.
(This would require a bit of refactoring of the API for
pass_down_bound, but it looks doable.)
In the short run, I'm not sure we have a better alternative than
removing this test - unless somebody has a better idea? - but it would
be good to work on all of the above problems.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
