Re: pgsql: Charge cpu_tuple_cost * 0.5 for Append and MergeAppend nodes.

On Thu, Feb 22, 2018 at 8:31 AM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
> On Thu, Feb 22, 2018 at 12:35 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> Robert Haas <rhaas(at)postgresql(dot)org> writes:
>>> Charge cpu_tuple_cost * 0.5 for Append and MergeAppend nodes.
>>
>> This seems to have produced some plan instability in the buildfarm.
>
> I was worried about that. Looking at it now. I wish we had some way
> to figure out whether a given plan was likely to be unstable on the
> buildfarm other than committing it and seeing what happens. It's
> surprisingly difficult to write non-trivial tests that produce
> consistent EXPLAIN output, and annoyingly time-consuming to try to
> figure out why they don't.

I pushed a commit to try to fix this by making the three tables being
joined not all identical in terms of row count. That seems like a
clear improvement, but there's every possibility it won't solve the
problem entirely. I'll keep an eye on the buildfarm and see what
happens.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company