On Thu, Mar 23, 2017 at 12:50 PM, Robert Haas <robertmhaas@gmail.com> wrote:
> Commit 7aea8e4f2daa4b39ca9d1309a0c4aadb0f7ed81b allowed a parallel
> plan to be generated when for a RETURN QUERY or RETURN QUERY EXECUTE
> statement in a PL/pgsql block. As it turns out, the analysis that led
> to this decision was totally wrong-headed, because the plan will
> always be executed using SPI_cursor_fetch(portal, true, 50), which
> will cause ExecutePlan() to get invoked with a count of 50, which will
> cause it to run the parallel plan serially, without workers.
> Therefore, passing CURSOR_OPT_PARALLEL_OK is a bad idea here; all it
> can do is cause us to pick a parallel plan that's slow when executed
> serially instead of the best serial plan.
>
> The attached patch fixes it. I plan to commit this and back-patch it
> to 9.6, barring objections or better ideas.
I guess the downside of back-patching this is that it could cause a
plan change for somebody which ends up being worse. On the whole,
serial execution of queries intended to be run in parallel isn't
likely to work out well, but it's always possible somebody has a cases
where it happens to be winning, and this could break it. So maybe I
should do this only in master? Thoughts?
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company