I wrote:
> 4. The fact that the test succeeds on many machines implies that the
> leader process is usually doing *all* of the work. This is in itself not
> very good. Even on the machines where it fails, the fact that the tuple
> counts are usually a pretty large fraction of the expected values
> indicates that the leader usually did most of the work. We need to take
> a closer look at why that is.
I've spent some time poking into this, and it seems the bottom line is
that the time needed to launch a parallel worker and get it started on
running the query plan is comparable to the time needed to scan all 10000
rows of tenk1. Maybe this isn't surprising, I dunno; but it doesn't give
me a warm feeling about how much exercise the parallel scan machinery is
really getting in select_parallel.sql. Not sure what to do about that ---
I don't really want to make the regression test case large enough (and
slow enough) to provide a more realistic scenario.
In the meantime, I've been able to make the failure reproducible by the
expedient of sticking "pg_usleep(1000)" into heap_parallelscan_nextpage(),
thus slowing down the leader's scan enough so that the workers can get
started.
> However, the bottom line here is that parallel scan is completely broken
> for rescans, and it's not (solely) the fault of nodeGatherMerge; rather,
> the issue is that nobody bothered to wire up parallelism to the rescan
> parameterization mechanism. I imagine that related bugs can be
> demonstrated in 9.6 with little effort.
I've so far been unable to break it for cases involving only Gather.
The issue is triggered, basically, by having a Sort or HashAgg node
below Gather[Merge], since either of those can decide that they don't
need to rescan their child. While it's not terribly hard to get the
planner to make such plans, you always end up with another Sort or
HashAgg above the Gather, and that masks the problem because the upper
node makes the same decision that it needn't rescan its child, protecting
the Gather from being run more than once. The particular plan shape that
a2b70c89c used,
-> Finalize GroupAggregate -> Gather Merge -> Partial GroupAggregate ->
Sort -> Parallel Seq Scan on tenk1
does exhibit the problem, but it requires GatherMerge so that no extra
sort is needed above the parallel subplan.
This may mean that we don't need to risk back-patching the fix into 9.6.
I'm not totally convinced of that yet, but I can't show that it's needed
given 9.6's limited support for parallelism.
regards, tom lane
