On Tue, Dec 8, 2020 at 10:46 AM James Coleman <jtc331@gmail.com> wrote:
>
> While I haven't actually tracked down to guarantee this is handled
> elsewhere, a thought experiment -- I think -- shows it must be so.
> Here's why: suppose we don't have a limit here, but the query return
> order is different in different backends. Then we would have the same
> problem you bring up. In that case this code is already setting
> consider_parallel=true on the rel. So I don't think we're changing any
> behavior here.
>
AFAICS, the patch seems very reasonable and specifically targets
lateral subqueries with limit/offset. It seems like the uncorrelated
case is the only real concern.
I generally agree that the current patch is probably not changing any
behavior in the uncorrelated case (and like yourself, haven't yet
found a case for which it breaks), but I'm not sure Brian's concerns
can be ruled out entirely.
How about a minor update to the patch to make it slightly more
restrictive, to exclude the case when there are no lateral
cross-references, so we'll be allowing parallelism only when we know
the lateral subquery will be evaluated anew for each source row?
I was thinking of the following patch modification:
BEFORE:
- if (limit_needed(subquery))
+ if (!rte->lateral && limit_needed(subquery))
AFTER:
- if (limit_needed(subquery))
+ if ((!rte->lateral || bms_is_empty(rel->lateral_relids)) &&
+ limit_needed(subquery))
Thoughts?
Regards,
Greg Nancarrow
Fujitsu Australia
