eqjoinsel_semi still sucks ...

On 2/5/2012 20:34, Tom Lane wrote:

On reflection I think that the idea of clamping ndistinct beforehand is
just wrong, and what we ought to do instead is apply a multiplier to the
selectivity estimate afterwards. In the case of a base rel we could
just multiply by the selectivity of its baserestrictinfo list. For join
rels it's a bit harder to guess how much a given input relation might
have been decimated, but if the join's estimated size is smaller than
the output size of the base rel the correlation var came from, we could
multiply by that ratio (on top of whatever correction came from the base
rel's restriction clauses).

I got stuck in some cases where (due to a tree of filters) the planner
underestimates the JOIN just because the ndistinct conveys a huge number
to the selectivity estimation formula. However, the estimation of both
input relations is made correctly and is limited.
I've tried to understand the logic through commits 0d3b231eebf,
97930cf578e and 7f3eba30c9d. But it is still not clear.
So, why the idea of clamping ndistinct is terrible in general? Could you
explain your reasons a bit more?

--
regards,
Andrey Lepikhov
Postgres Professional

On 23.06.2023 14:28, Andrey Lepikhov wrote:

On 2/5/2012 20:34, Tom Lane wrote:

On reflection I think that the idea of clamping ndistinct beforehand is
just wrong, and what we ought to do instead is apply a multiplier to the
selectivity estimate afterwards.  In the case of a base rel we could
just multiply by the selectivity of its baserestrictinfo list. For join
rels it's a bit harder to guess how much a given input relation might
have been decimated, but if the join's estimated size is smaller than
the output size of the base rel the correlation var came from, we could
multiply by that ratio (on top of whatever correction came from the base
rel's restriction clauses).

I got stuck in some cases where (due to a tree of filters) the planner
underestimates the JOIN just because the ndistinct conveys a huge
number to the selectivity estimation formula. However, the estimation
of both input relations is made correctly and is limited.
I've tried to understand the logic through commits 0d3b231eebf,
97930cf578e and 7f3eba30c9d. But it is still not clear.
So, why the idea of clamping ndistinct is terrible in general? Could
you explain your reasons a bit more?

Hi, I also have an interest in understanding the same issue, and I dug
into the above commits and the topic . I hope this letter will help to
sort out this issue.

I found some information on this topic in the discussion [1], more
specifically in the email [2].

First of all, according to the post [2] and 0d3b231ee commit, clampling
ndistinct is necessary because it allows us to eliminate non-MCV values,
or potentially even the least-common MCVs for the inner relation that
might not be analyze operation (part of the code is given below when not
all different values are taken into account):

*static void
compute_distinct_stats(VacAttrStatsP stats,
                       AnalyzeAttrFetchFunc fetchfunc,
                       int samplerows,
                       double totalrows)*
*{*
*...
*
*/*
     * Decide how many values are worth storing as most-common values. If
     * we are able to generate a complete MCV list (all the values in the
     * sample will fit, and we think these are all the ones in the table),
     * then do so.  Otherwise, store only those values that are
     * significantly more common than the values not in the list.
     *
     * Note: the first of these cases is meant to address columns with
     * small, fixed sets of possible values, such as boolean or enum
     * columns.  If we can *completely* represent the column population by
     * an MCV list that will fit into the stats target, then we should do
     * so and thus provide the planner with complete information.  But if
     * the MCV list is not complete, it's generally worth being more
     * selective, and not just filling it all the way up to the stats
     * target.
     */
    if (track_cnt < track_max && toowide_cnt == 0 &&
      stats->stadistinct > 0 &&
      track_cnt <= num_mcv)
    {
      /* Track list includes all values seen, and all will fit */
      num_mcv = track_cnt;
    }
    else
    {
      int       *mcv_counts;

      /* Incomplete list; decide how many values are worth keeping */
      if (num_mcv > track_cnt)
        num_mcv = track_cnt;

      if (num_mcv > 0)
      {
        mcv_counts = (int *) palloc(num_mcv * sizeof(int));
        for (i = 0; i < num_mcv; i++)
          mcv_counts[i] = track[i].count;

        num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
                       stats->stadistinct,
                       stats->stanullfrac,
                       samplerows, totalrows);
      }
    }
*
*...*
*}**
*

The reason why the idea of clamping is used may be related to the fact
that during the processing of eqjoinsel_inner, the ndistinct estimate
for the join key column decreases (in proportion to the selectivity of
the baserel restrictions), which leads to a proportional increase in the
number of selectivities for the join condition; we have to eliminate
values other than MCV or potentially the least-common MCV for the inner
relation, since the most common values are the ones that are most likely
to survive the decimation resulting from a lower restriction clause.

This is discussed in [2], I copied a fragment where it is clearly seen
below:

*If you think about what is happening in eqjoinsel_inner with the patch,
we are reducing the ndistinct estimate for the join key column
proportionally to the selectivity of whatever baserel restrictions
apply.  This then results in proportionally increasing the selectivity
number for the join condition --- in other words, we're more or less
cancelling out the effects of one or the other relation's base
restrictions.  So that's pretty broken in general. **The reason it is
important for semi/antijoin inner relations is that this is actually the
only way that restrictions applied to the inner rel get to impact the
join size estimate at all, since set_joinrel_size_estimates is not going
to factor the inner rel size into what it multiplies the join
selectivity against.    In short, I was mistakenly extrapolating from
the observation that it helped to hack the ndistinct estimate for a
semijoin's inner rel, to the conclusion that we should do that for all
join input rels. *

*...*
*Yeah, those are estimating that all the outer rows have join partners,
because there are more distinct values in the sub-select than there are
in the outer relation.
*

In addition, the answer to the question why clamping is necessary is
also in the comment to commit 97930cf5, and this not only reduced the
number of rows coming out of a table and  the number of
possibly-distinct values of a join variable, but join restriction
clauses that might have been applied at a lower level of join.

*That patch accounted for baserel restriction clauses that reduced the
number of rows coming out of a table (and hence the number of
possibly-distinct values of a join variable), but not for join
restriction clauses that might have been applied at a lower level of
join.  To account for the latter, look up the sizes of the min_lefthand
and min_righthand inputs of the current join, and clamp with those in
the same way as for the base relations.*

1.
/messages/by-id/5156.1314829311@sss.pgh.pa.us

2. /messages/by-id/raw/5156.1314829311@sss.pgh.pa.us

----

Regards,

Alena Rybakina

Alena Rybakina <lena.ribackina@yandex.ru> writes:

On 23.06.2023 14:28, Andrey Lepikhov wrote:

On 2/5/2012 20:34, Tom Lane wrote:

On reflection I think that the idea of clamping ndistinct beforehand is
just wrong, and what we ought to do instead is apply a multiplier to the
selectivity estimate afterwards.

I got stuck in some cases where (due to a tree of filters) the planner
underestimates the JOIN just because the ndistinct conveys a huge
number to the selectivity estimation formula. However, the estimation
of both input relations is made correctly and is limited.
I've tried to understand the logic through commits 0d3b231eebf,
97930cf578e and 7f3eba30c9d. But it is still not clear.
So, why the idea of clamping ndistinct is terrible in general? Could
you explain your reasons a bit more?

Hi, I also have an interest in understanding the same issue, and I dug
into the above commits and the topic . I hope this letter will help to
sort out this issue.

Note that nothing's actually been done about the idea I expressed at the
start of this thread. The behavior is better now: if you try the example
in v12 or later you get

regression=# explain analyze select * from tenk1 a where not exists(select 1 from tenk1 b where a.thousand = b.unique2 and b.two = 0);
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------
Hash Anti Join (cost=532.50..1059.38 rows=5000 width=244) (actual time=1.993..3.959 rows=5090 loops=1)
Hash Cond: (a.thousand = b.unique2)
-> Seq Scan on tenk1 a (cost=0.00..445.00 rows=10000 width=244) (actual time=0.003..0.502 rows=10000 loops=1)
-> Hash (cost=470.00..470.00 rows=5000 width=4) (actual time=1.960..1.960 rows=5000 loops=1)
Buckets: 8192 Batches: 1 Memory Usage: 240kB
-> Seq Scan on tenk1 b (cost=0.00..470.00 rows=5000 width=4) (actual time=0.003..1.449 rows=5000 loops=1)
Filter: (two = 0)
Rows Removed by Filter: 5000
Planning Time: 0.760 ms
Execution Time: 4.087 ms
(10 rows)

I believe this improvement just stems from commit a314c3407 ("Clamp
semijoin selectivity to be not more than inner-join selectivity"),
and so this example looks good only because the actual semijoin
output size happens to be the same as the inner-join output size.
With a slight adjustment, that's no longer true and the estimate
still sucks:

regression=# explain analyze select * from tenk1 a, tenk1 b where a.thousand = b.fivethous and b.two = 0;
QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------
Hash Join (cost=532.50..1127.50 rows=10000 width=488) (actual time=2.000..5.742 rows=10000 loops=1)
...

regression=# explain analyze select * from tenk1 a where exists(select 1 from tenk1 b where a.thousand = b.fivethous and b.two = 0);
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------
Hash Semi Join (cost=532.50..1127.50 rows=10000 width=244) (actual time=1.529..3.474 rows=5000 loops=1)
...

regression=# explain analyze select * from tenk1 a where not exists(select 1 from tenk1 b where a.thousand = b.fivethous and b.two = 0);
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------
Hash Anti Join (cost=532.50..1027.50 rows=1 width=244) (actual time=1.564..3.476 rows=5000 loops=1)
...

So we still have an issue to fix. I've not had time to pursue
the idea I expressed at the start of the thread. Also, it's a
bit scary to change this logic with only a few examples to look
at. If you could reduce the problem cases you're looking at
to something sharable, that could help move things along.

regards, tom lane