Re: BUG #15324: Non-deterministic behaviour from parallelised sub-query

On Mon, Aug 20, 2018 at 4:20 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>
> On Thu, Aug 16, 2018 at 9:25 AM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> > On Wed, Aug 15, 2018 at 4:40 PM, Stephen Frost <sfrost(at)snowman(dot)net> wrote:
> >> Greetings,
> >>
> >
> > Yeah, let me summarize the problems which require patches:
> > (a) Consider the presence of a LIMIT/OFFSET in a sub-select as making
> > it parallel-unsafe.
> >
>
> As mentioned up-thread, I have considered adding a check in
> max_parallel_hazard_walker, but it turns out that it will make the
> whole query parallel-unsafe even if one of the sub-selects has
> Limit/Offset. I think the better idea is to detect that during
> set_rel_consider_parallel. Attached patch
> prohibit_parallel_limit_subselect_v2 implements the fix for same.
>

I was trying this patch on back-branches and found that it doesn't
apply cleanly beyond PG11, so created separate patches for 10 and 9.6.
Further, I found that the test for this patch was not failing for
9.6 (without the patch) even though the code doesn't deal with this
problem. On further investigation, I found that it is because the
commit
655393a022bd653e2b48dbf20b69236981e35195 has not been backpatched to
9.6. I don't see any reason why we shouldn't backpatch this commit.
So, I have attached a patch (fix_parallel_hash_path_v1.patch) which we
can backpatch in 9.6.

Robert, your input will be highly appreciated here especially for the
back patch (to 9.6) I am proposing?

> > (b) Consider the presence of any window function calculation as
> > parallel-restricted operation.
> >
>
> For this, we need to mark all the window functions like row_number,
> rank, dense_rank, etc as parallel-restricted. Additionally, we also
> need to detect the presence of aggregate functions that act as window
> functions (when an OVER clause follows the call). Attached patch
> treat_window_func_calc_parallel_restricted_v1 implements the fix.
>

On again looking at this patch, I found that the test case in the
patch was not sufficient to reproduce the problem reported here, so I
have changed the test on the lines of what is reported in this thread.

--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com