Re: BUG #18588: Cannot force/let database use parallel execution in simple case.

On Fri, 23 Aug 2024 at 08:42, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Ah, I traced through it, and here's what's happening: at small enough
> estimated rowcounts, the parallel and non-parallel plans have fuzzily
> the same cost (parallel is a shade cheaper, but only a shade).
> Their other properties such as pathkeys are the same too. So we get
> to the tie-breaking logic in add_path, and what breaks the tie is
> the difference in parallel safety: the non-parallel plan is marked
> parallel_safe and the parallel one (which by this point is a Gather)
> is not.

I played around with the attached script and set some breakpoints in
cost_index(). I'm seeing the same thing as you with the parallel path
being only slightly cheaper, but when looking at cost_index(), it's
easy to see why.

It's only the cpu_run_cost that's divided by the parallel_divisor. In
this case, cpu_run_cost is just 7852.89 for the parallel path and the
parallel_divisor is 2.4. The run_cost is not divided and is much
higher at 501792, so dividing the CPU cost does not save much. Just a
few thousand in half a million, which is why the plans are fuzzily the
same cost.

If I make the cpu_tuple_cost 0.02 instead of 0.01, I get the parallel
plan. Possibly increasing effective_cache_size would be the best way
for Maxim to get the parallel plan. I wonder if that's just left at
the default 4GB... Not many people tune that.

David