RE: Re: Re: Generic Plans for Prepared Statement are 158155 times slowerthan Custom Plans

-----Original Message-----
From: Tom Lane [mailto:tgl@sss.pgh.pa.us]
Sent: Monday, April 29, 2019 8:06 PM
To: Bruce Momjian <bruce@momjian.us>
Cc: Naik, Sameer <Sameer_Naik@bmc.com>; pgsql-performance@lists.postgresql.org
Subject: [EXTERNAL] Re: Generic> Plans for Prepared Statement are 158155 times slower than Custom Plans

Bruce Momjian <bruce@momjian.us> writes:
> On Mon, Apr 29, 2019 at 10:36:20AM +0000, Naik, Sameer wrote:
>> Since Postgres 9.2, for prepared statements, the CBO automatically
>> switches from Custom Plan to Generic plan on the sixth iteration
>> (reference backend/ utils/cache/plancache.c).

> This is not totally true.

Yeah, that's a pretty inaccurate statement of the behavior.

The problem seems to be that the actual values being used for
c400129200 and c400127400 are quite common in the dataset, so that when considering

Filter: ... (c400129200 = '0'::citext) AND (c400127400 = 'DATASET1M'::citext)

the planner makes a roughly correct assessment that there are a lot of such rows, so it prefers to index on the basis of the giant OR clause instead, even though that's fairly expensive.  But, when considering the generic case

      ->  Index Scan using i776_0_400129200_t776 on t776  (cost=0.42..12.66 rows=1 width=52) (actual time=1190.399..5544.385 rows=48 loops=1)
            Index Cond: ((c400129200 = $1) AND (c400127400 = $2))

it's evidently guessing that just a few rows will match the index condition (no more than about 3 given the cost number), making this plan look much cheaper, so it goes with this plan.  I wonder what the actual distribution of those keys is.


In v10 and later, it's quite possible that creating extended stats on the combination of those two columns would produce a better estimate.  Won't help OP on 9.6, though.

This isn't the first time we've seen a plan-choice failure of this sort.
I've wondered if we should make the plancache simply disbelieve generic cost estimates that are actually cheaper than the custom plans, on the grounds that they must be estimation errors.  In principle a generic plan could never really be better than a custom plan; so if it looks that way on a cost basis, what that probably means is that the actual parameter values are outliers of some sort (e.g. extremely common), and the custom plan "knows" that it's going to be taking a hit from that, but the generic plan doesn't.  In this sort of situation, going with the generic plan could be really disastrous, which is exactly what the OP is seeing (and what we've seen reported before).

However, I'm not sure how to tune this idea so that it doesn't end up rejecting perfectly good generic plans.  It's likely that there will be some variation in the cost estimates between the generic and specific cases, even if the plan structure is exactly the same; and that variation could go in either direction.

            regards, tom lane