Re: FailedAssertion on partprune

On 24 July 2018 at 16:55, Jaime Casanova <jaime(dot)casanova(at)2ndquadrant(dot)com> wrote:
> I was trying sqlsmith on REL_11_STABLE (commit
> 1b957e59b92dc44c14708762f882d7910463a9ac) with a database i have at
> hand, and got an assertion failure.
> It seems to happen during planning on prunning time but only when
> tables get bigger than certain size.
>
> I configured it with "--enable-debug --enable-depend
> --enable-cassert", attached is the backtrace and a script to reproduce
> the problem.

It looks like an AppendPath which has a non-NIL partitioned_rels List
won't always only contain leaf partitions as subpaths. The plan that
the planner is trying to make here is:

https://explain.depesz.com/s/26zj

In this case, the 2nd to 5th Append subplan's parents are actually
partitioned tables rather than leaf partitions.
make_partition_pruneinfo() assumes these will always be leaf
partitions. Looking more closely at accumulate_append_subpath() I can
see that there are cases where it won't pull sub-Append paths into the
main Append. That's what we're seeing in this plan.

The fix is either to have make_partition_pruneinfo() return NIL to
disable run-time pruning when any of the subpaths are not a leaf
partition (it might be a good safety precaution regardless), or we
could maybe think harder about what can be done in
accumulate_append_subpath() to flatten the Append/MergeAppend paths in
this case.

I've attached a patch for the former.

--
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services