On Mon, Jun 13, 2016 at 3:42 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> I would not be surprised at a change to a parallel-query plan, but there's
>> no parallelism here, so what happened? This looks like a bug to me.
>> (Also, doing this query without COSTS OFF shows that the newly selected
>> plan actually has a greater estimated cost than the expected plan, which
>> makes it definitely a bug.)
>
> I looked into this and found that the costs are considered fuzzily the
> same, and then add_path prefers the slightly-worse path on the grounds
> that it is marked parallel_safe while the MinMaxAgg path is not. It seems
> to me that there is some fuzzy thinking going on there. On exactly what
> grounds is a path to be preferred merely because it is parallel safe, and
> not actually parallelized?
A parallel-safe plan can be joined to a partial path at a higher level
of the plan tree to form a new partial path. A non-parallel-safe path
cannot. Therefore, if two paths are equally good, the one which is
parallel-safe is more useful (just as a sorted path which is slightly
more expensive than an unsorted path is worth keeping around because
it is ordered). In practice, we don't yet have the ability for
parallel-safe paths from subqueries to affect planning at higher query
levels, but that's because the pathification stuff landed too late in
the cycle for me to fully react to it.
> Or perhaps the question to ask is whether a
> MinMaxAgg path can be marked parallel-safe.
That is a good question.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company