Thread: Index only scan sometimes switches to sequential scan for small amount of rows

Hi,

Situation:

We have a table with 3,500,000+ rows, which contain items that need to
be printed or have been printed previously.

Most of these records have a status of 'PRINTED', we have a partial
index on this table WHERE status <> 'PRINTED'.
During normal operation there will be < 10 records matching 'NOT_YET_PRINTED'.
When using the index scan this is done in < 5ms, but when the
sequential scan is involved the query runs > 500ms.


We query this table often in the form:

SELECT *
  FROM print_list
  JOIN [...]
  JOIN [...]
 WHERE stats = 'NOT_YET_PRINTED'
 LIMIT 8;

This query is currently switching between a sequential scan on the
print_list table and an index scan on the previously mentioned index.

When doing an explain analyze on the queries we see that it sometimes
expects to return > 5000 records when in reality it is only < 5
records that are returned, for example:

   ->  Index Scan using print_list_status_idx on print_list
(cost=0.27..1138.53 rows=6073 width=56) (actual time=0.727..0.727
rows=0 loops=1)

Sometimes, this results in the planner choosing a sequential scan for
this query.

When analyzing pg_stats we have sometimes have the following: (Note:
'NOT_YET_PRINTED' has not been found during this analyze, these are
real values)

 attname                | status
 inherited              | f
 null_frac              | 0
 avg_width              | 4
 n_distinct             | 3
 most_common_vals       | {PRINTED}
 most_common_freqs      | {0.996567}
 histogram_bounds       | {PREPARED,ERROR}
 correlation            | 0.980644

A question about this specific entry, which some of you may be able to
shed some light on:

most_common_vals contains only 1 entry, why is this? I would expect to
see 3 entries, as it has n_distinct=3

When looking at
http://www.postgresql.org/docs/current/static/row-estimation-examples.html
we can see that an estimate > 5000 is what is to be expected for these
statistics:

# select ( (1 - 0.996567)/2 * 3500000 )::int;
 int4
------
 6008
(1 row)

But why does it not record the frequency of 'PREPARED' and 'ERROR' in
most_common_*?

Our current strategies in mitigating this problem is decreasing the
autovacuum_*_scale_factor for this specific table, therefore
triggering more analyses and vacuums.

This is helping somewhat, as if the problem occurs it often solved
automatically if autoanalyze analyzes this table, it is analyzed many
times an hour currently.

We can also increase the 'Stats target' for this table, which will
cause the statistics to contain information about 'NOT_YET_PRINTED'
more often, but even then, it may not find any of these records, as
they sometimes do not exist.

Could you help us to find a strategy to troubleshoot this issue further?

Some specific questions:
- We can see it is doing a sequential scan of the full table (3.5mio
records) even when it only expects 8000 records to be returned, we
would expect this not to happen so soon.
- Why is most_common_* not filled when there are only 3 distinct values?

Feike Steenbergen

On 25.3.2015 13:04, Feike Steenbergen wrote:
...
> When analyzing pg_stats we have sometimes have the following: (Note:
> 'NOT_YET_PRINTED' has not been found during this analyze, these are
> real values)
>
>  attname                | status
>  inherited              | f
>  null_frac              | 0
>  avg_width              | 4
>  n_distinct             | 3
>  most_common_vals       | {PRINTED}
>  most_common_freqs      | {0.996567}
>  histogram_bounds       | {PREPARED,ERROR}
>  correlation            | 0.980644
>
> A question about this specific entry, which some of you may be able to
> shed some light on:
>
> most_common_vals contains only 1 entry, why is this? I would expect to
> see 3 entries, as it has n_distinct=3

To be included in the MCV list, the value has to actually appear in the
random sample at least twice, IIRC. If the values are very rare (e.g. if
you only have such 10 rows out of 3.5M), that may not happen.

You may try increasing the statistics target for this column, which
should make the sample larger and stats more detailed (max is 10000,
which should use sample ~3M rows, i.e. almost the whole table).

> When looking at
> http://www.postgresql.org/docs/current/static/row-estimation-examples.html
> we can see that an estimate > 5000 is what is to be expected for these
> statistics:
>
> # select ( (1 - 0.996567)/2 * 3500000 )::int;
>  int4
> ------
>  6008
> (1 row)
>
> But why does it not record the frequency of 'PREPARED' and 'ERROR'
> in most_common_*?

Can you post results for this query?

SELECT stats, COUNT(*) FROM print_list group by 1

I'd like to know how frequent the other values are.

>
> Our current strategies in mitigating this problem is decreasing the
> autovacuum_*_scale_factor for this specific table, therefore
> triggering more analyses and vacuums.

I'm not sure this is a good solution. The problem is elsewhere, IMHO.

> This is helping somewhat, as if the problem occurs it often solved
> automatically if autoanalyze analyzes this table, it is analyzed
> many times an hour currently.
>
> We can also increase the 'Stats target' for this table, which will
> cause the statistics to contain information about 'NOT_YET_PRINTED'
> more often, but even then, it may not find any of these records, as
> they sometimes do not exist.

This is a better solution, IMHO.

>
> Could you help us to find a strategy to troubleshoot this issue
> further?

You might also make the index scans cheaper, so that the switch to
sequential scan happens later (when more rows are estimated). Try to
decreasing random_page_cost from 4 (default) to 1.5 or something like that.

It may hurt other queries, though, depending on the dataset size etc.

--
Tomas Vondra                http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

Hi, thanks for having a look and thinking with us

On 25 March 2015 at 13:45, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:
> Can you post results for this query?
>
> SELECT stats, COUNT(*) FROM print_list group by 1

   status   |  count
----------------+---------
 ERROR          |     159
 PREPARED       |   10162
 PRINTED        | 3551367
 TO_BE_PREPARED |       2
(4 rows)

>> We can also increase the 'Stats target' for this table, which will
>> cause the statistics to contain information about 'NOT_YET_PRINTED'
>> more often, but even then, it may not find any of these records, as
>> they sometimes do not exist.
>
> This is a better solution, IMHO.

We'll have a go at this, also if what you say about values having to
appear at least twice, the other values may make it into
most_common_*, which would make it clearer to us.

We're a bit hesitant to decrease random_page_cost (currently 3 in this
cluster) as a lot more is happening on this database.

I'm posting this as I am trying to understand what has happened.
TLDR: The problem seems to be fixed now.

By bumping the statistics_target we see that most_common_vals is
having its contents filled more often, causing way better estimates:

 attname                | status
 inherited              | f
 null_frac              | 0
 avg_width              | 4
 n_distinct             | 3
 most_common_vals       | {PRINTED,PREPARED,ERROR}
 most_common_freqs      | {0.996863,0.00307333,6.33333e-05}
 histogram_bounds       | (null)
 correlation            | 0.98207
 most_common_elems      | (null)
 most_common_elem_freqs | (null)
 elem_count_histogram   | (null)

Basically 100% of the records are accounted for in these statistics,
the planner now consistently estimates the number of rows to be very
small for other values.

Before bumping the target we didn't have information for 0.34% of the
rows, which in this case means roughly 11K rows.

What is the reasoning behind having at least 2 hits before including
it in the most_common_* columns?

Feike Steenbergen <feikesteenbergen@gmail.com> writes:
> On 25 March 2015 at 13:45, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:
>>> We can also increase the 'Stats target' for this table, which will
>>> cause the statistics to contain information about 'NOT_YET_PRINTED'
>>> more often, but even then, it may not find any of these records, as
>>> they sometimes do not exist.

>> This is a better solution, IMHO.

> We'll have a go at this, also if what you say about values having to
> appear at least twice, the other values may make it into
> most_common_*, which would make it clearer to us.

In principle increasing the stats target should fix this, whether or not
'NOT_YET_PRINTED' appears in the MCV list after any particular analyze;
because what will happen is that the frequency for 'PRINTED' will more
nearly approach 1, and so the estimated selectivity for other values
will drop even if they're not in the list.

            regards, tom lane

On 25 March 2015 at 19:07, Jeff Janes <jeff.janes@gmail.com> wrote:

> Also, I doubt that that is the problem in the first place.  If you collect a
> sample of 30,000 (which the default target size of 100 does), and the
> frequency of the second most common is really 0.00307333 at the time you
> sampled it, you would expect to find it 92 times in the sample. The chances
> against actually finding 1 instead of around 92 due to sampling error are
> astronomical.

It can be that the distribution of values is very volatile; we hope
the increased stats target (from the default=100 to 1000 for this
column) and frequent autovacuum and autoanalyze helps in keeping the
estimates correct.

It seems that it did find some other records (<> 'PRINTED), as is
demonstrated in the stats where there was only one value in the MCV
list: the frequency was 0.996567 and the fraction of nulls was 0,
therefore leaving 0.03+ for other values. But because none of them
were in the MCV and MCF list, they were all treated as equals. They
are certainly not equal.

I not know why some values were found (they are mentioned in the
histogram_bounds), but are not part of the MCV list, as you say, the
likeliness of only 1 item being found is very small.

Does anyone know the criteria for a value to be included in the MCV list?

> The problem seems to be rapidly changing stats, not too small of a target
> size (unless your original target size was way below the current default
> value, forgive me if you already reported that, I didn't see it anywhere).
> Maybe it would work better if you built the partial index where status =
> 'NOT_YET_PRINTED', instead of !='PRINTED'.

Thanks, we did create a partial index on 'NOT_YET_PRINTED' today to
help aiding these kind of queries.

On 25 March 2015 at 22:45, Jeff Janes <jeff.janes@gmail.com> wrote:

> How can the avg_width be 4 when the vast majority of entries are 7
> characters long?

The datatype is an enum, as I understand it, an enum type always
occupies 4 bytes

Sorry, didn't respond to all your questions:

> What version of PostgreSQL are running?  'select version();'

PostgreSQL 9.3.4 on x86_64-pc-linux-gnu, compiled by gcc-4.6.real
(Ubuntu/Linaro 4.6.3-1ubuntu5) 4.6.3, 64-bit

> What do you get when to do "analyze verbose print_list"?

# analyze verbose print_list ;
INFO:  analyzing "print_list"
INFO:  "print_list": scanned 53712 of 53712 pages, containing 3626950
live rows and 170090 dead rows; 300000 rows in sample, 3626950
estimated total rows
ANALYZE
Time: 6656.037 ms

On 26.3.2015 08:48, Jeff Janes wrote:
>
> OK, this is starting to look like a long-standing bug to me.
>
> If it only sees 3 distinct values, and all three are present at least
> twice, it throws all of them into the MCV list. But if one of those 3
> were present just once, then it tests them to see if they qualify.
> The test for inclusion is that it has to be present more than once,
> and that it must be "over-represented" by 25%.
>
> Lets say it sampled 30000 rows and found 29,900 of one value, 99 of
> another, and 1 of a third.
>
> But that turns into the second one needing to be present 12,500 times.
> The average value is present 10,000 times (30,000 samples with 3
> distinct values) and 25 more than that is 12,500.  So it excluded.
>
> It seems to me that a more reasonable criteria is that it must be
> over-represented 25% compared to the average of all the remaining values
> not yet accepted into the MCV list.  I.e. all the greater ones should be
> subtracted out before computing the over-representation threshold.

That might work IMO, but maybe we should increase the coefficient a bit
(say, from 1.25 to 2), not to produce needlessly long MCV lists.


> It is also grossly inconsistent with the other behavior.  If they are
> "29900; 98; 2" then all three go into the MCV.

Isn't the mincount still 12500? How could all three get into the MCV?


--
Tomas Vondra                http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

On 26.3.2015 17:35, Jeff Janes wrote:
> On Thu, Mar 26, 2015 at 5:44 AM, Tomas Vondra
> <tomas.vondra@2ndquadrant.com <mailto:tomas.vondra@2ndquadrant.com>> wrote:
>
>> That might work IMO, but maybe we should increase the coefficient a
>> bit (say, from 1.25 to 2), not to produce needlessly long MCV
>> lists.
>
> That wouldn't work here, because at the point of decision the value
> present 99 times contributes half the average, so the average is 50,
> and of course it can't possibly be twice of that.

Oh, right. How could I miss that? ;-)

> I have a patch, but is there a way to determine how it affects a
> wide variety of situations? I guess run `make installcheck`, then
> analyze, then dump pg_stats, with the patch and without the patch,
> and then compare the dumpsj?

I doubt there's such way. I'd argue that if you can show this always
generates longer MCV lists, we can assume the stats are probably more
accurate, and thus the plans should be better.

Of course, there's always the possibility that the plan was good by
luck, and improving the estimates will result in a worse plan. But I
don't think we can really fix that :-(

>>> It is also grossly inconsistent with the other behavior. If they
>>> are "29900; 98; 2" then all three go into the MCV.
>>
>> Isn't the mincount still 12500? How could all three get into the
>> MCV?
>
> If all observed values are observed at least twice, it takes a
> different path through the code. It just keeps them all in the MCV
> list. That is what is causing the instability for the OP. If the 3rd
> most common is seen twice, then all three are kept. If it is seen
> once, then only the most common is kept. See if statements at 2494
> and 2585
>
> else if (toowide_cnt == 0 && nmultiple == ndistinct)
>
>         if (track_cnt == ndistinct ....

Aha, I think I see it now. I've been concentrating on this code:

   avgcount = (double) samplerows / ndistinct;
   /* set minimum threshold count to store a value */
   mincount = avgcount * 1.25;
   if (mincount < 2)
      mincount = 2;

but this is actually too late, because first we do this:

  else if (toowide_cnt == 0 && nmultiple == ndistinct)
  {
     stats->stadistinct = ndistinct;
  }

and that only happens if each item is observed at least 2x in the sample
(and the actual Haas and Stokes estimator it not used).

And then we do this:

  if (track_cnt == ndistinct && toowide_cnt == 0 &&
      stats->stadistinct > 0 && track_cnt <= num_mcv)
  {
    num_mcv = track_cnt;
  }

so that we track everything.

If at least one value is seen only 1x, it works differently, and we use
the code with (1.25*avgcount) threshold.

I wonder where the 1.25x threshold comes from - whether it's something
we came up with, or if it comes from some paper. I guess the former.


--
Tomas Vondra                http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services