Обсуждение: Cluster vs. non-cluster query planning

I'm running postgres 8.0.7, and I've got a table of orders with about
100,000 entries. I want to just look at the new orders, right now 104 of
them.

EXPLAIN ANALYZE
SELECT
    order_id
    FROM
        orders
    WHERE
        order_statuses_id = (SELECT id FROM order_statuses WHERE id_name
= 'new');

 Seq Scan on orders o  (cost=1.20..11395.51 rows=7029 width=8) (actual
time=286.038..287.662 rows=104 loops=1)
   Filter: (order_statuses_id = $0)
   InitPlan
     ->  Seq Scan on order_statuses  (cost=0.00..1.20 rows=1 width=4)
(actual time=0.030..0.039 rows=1 loops=1)
           Filter: ((id_name)::text = 'new'::text)
 Total runtime: 288.102 ms

The dreaded sequential scan.  I've got an index on order_statuses_id and
I've VACUUM ANALYZEd the table, but I'm currently clustered on the
primary key (order_id).

With enable_seqscan = off, I get:
-------------------------------------------------
 Index Scan using orders_status_btree_idx on orders o
(cost=4.64..12457.14 rows=7031 width=8) (actual time=0.164..0.664
rows=104 loops=1)
   Index Cond: (order_statuses_id = $0)
   InitPlan
     ->  Index Scan using order_statuses_id_name_key on order_statuses
(cost=0.00..4.64 rows=1 width=4) (actual time=0.128..0.134 rows=1 loops=1)
           Index Cond: ((id_name)::text = 'new'::text)
 Total runtime: 1.108 ms

If I hard-code the 'new' status ID, I get:
-------------------------------------------------
EXPLAIN ANALYZE
SELECT
    order_id
    FROM
        orders
    WHERE
        order_statuses_id = 1;

 Index Scan using orders_status_btree_idx on orders o
(cost=0.00..4539.65 rows=1319 width=8) (actual time=0.132..1.883
rows=104 loops=1)
   Index Cond: (order_statuses_id = 1)
 Total runtime: 2.380 ms

Here is the pg_stats entry for orders.order_statuses_id:
 schemaname | tablename |      attname      |  null_frac  | avg_width |
n_distinct | most_common_vals |          most_common_freqs
|    histogram_bounds     | correlation

------------+-----------+-------------------+-------------+-----------+------------+------------------+--------------------------------------+-------------------------+-------------
 public     | orders    | order_statuses_id | 0.000208333 |         4
|         14 | {8,24,10,25}     | {0.385417,0.242083,0.230625,0.07875} |
{1,7,7,9,9,9,9,9,23,26} |    0.740117

This is with SET STATISTICS = 16 on the column, since that's how many
different values the column can currently take.

Now, here's the thing - if I cluster on the index on order_statuses_id,
the original query produces:
 Index Scan using orders_status_btree_idx on orders o
(cost=1.20..978.94 rows=8203 width=8) (actual time=0.097..0.598 rows=104
loops=1)
   Index Cond: (order_statuses_id = $0)
   InitPlan
     ->  Seq Scan on order_statuses  (cost=0.00..1.20 rows=1 width=4)
(actual time=0.056..0.065 rows=1 loops=1)
           Filter: ((id_name)::text = 'new'::text)
 Total runtime: 1.042 ms

Estimated cost went way down. The pg_stats entry becomes:

 schemaname | tablename |      attname      | null_frac | avg_width |
n_distinct | most_common_vals |           most_common_freqs
|  histogram_bounds   | correlation

------------+-----------+-------------------+-----------+-----------+------------+------------------+----------------------------------------+---------------------+-------------
 public     | orders    | order_statuses_id |         0 |         4
|         12 | {8,24,10,25}     | {0.386458,0.244167,0.238333,0.0720833}
| {1,7,7,9,9,9,22,26} |           1

I'm hesitant to cluster on the order_statuses_id index, because there
are a lot of other queries using this table, many of which join on
order_id. I also feel like I ought to be able to get the planner to do
an index scan without hard-coding the order_statuses_id value.

Questions:
* What can I do to reduce the estimated row count on the query?
* Why does clustering drive down the estimated cost for the index scan
so much? Does a change in correlation from .72 to 1 make that much of a
difference?
* Can I convince my query planner to index scan without clustering on
the order_statuses_id index, or setting enable_seqscan = off?

Potential note of interest:  This is a very wide, monolithic table - no
less than 100 columns, with several check constraints, foreign key
constraints, and indexes, including three functional indexes.

Side question: Sometimes, when I VACUUM ANALYZE the table, the pg_stats
entry for order_statuses_id has almost all of the possible values in
most_common_vals, instead of just a handful.  Example:

 schemaname | tablename |      attname      | null_frac | avg_width |
n_distinct |         most_common_vals
|
most_common_freqs
| histogram_bounds | correlation

------------+-----------+-------------------+-----------+-----------+------------+-----------------------------------+------------------------------------------------------------------------------------------------------------------------------+------------------+-------------
 public     | orders    | order_statuses_id |         0 |         4
|         13 | {8,24,10,25,9,7,23,26,1,22,2,5,4} |

{0.393125,0.240208,0.226042,0.07875,0.0275,0.0145833,0.0110417,0.00291667,0.00229167,0.001875,0.000625,0.000625,0.000416667}

|                  |           1

This doesn't appear to influence whether the index scan is chosen, but I
am curious as to why this happens.

> Questions:
> * What can I do to reduce the estimated row count on the query?
> * Why does clustering drive down the estimated cost for the index scan
> so much? Does a change in correlation from .72 to 1 make that much of
> a difference?
> * Can I convince my query planner to index scan without clustering on
> the order_statuses_id index, or setting enable_seqscan = off?


After some more digging on the mailing list, I found some comments on
effective_cache_size.  Bringing it up from the default of 1000 does pust
the estimated cost for the index scan below that of the sequential scan,
but not by much.

With SET effective_cache_size = 1000:
 Seq Scan on orders o  (cost=1.20..11395.53 rows=7029 width=8) (actual
time=280.148..281.512 rows=105 loops=1)
   Filter: (order_statuses_id = $0)
   InitPlan
     ->  Seq Scan on order_statuses  (cost=0.00..1.20 rows=1 width=4)
(actual time=0.012..0.020 rows=1 loops=1)
           Filter: ((id_name)::text = 'new'::text)
 Total runtime: 281.700 ms

With SET effective_cache_size = 10000:
 Index Scan using orders_status_btree_idx on orders o
(cost=1.20..9710.91 rows=7029 width=8) (actual time=0.050..0.372
rows=105 loops=1)
   Index Cond: (order_statuses_id = $0)
   InitPlan
     ->  Seq Scan on order_statuses  (cost=0.00..1.20 rows=1 width=4)
(actual time=0.016..0.024 rows=1 loops=1)
           Filter: ((id_name)::text = 'new'::text)

The ratios between estimated costs are still nowhere near the ratio of
actual costs.

Nolan Cafferky <Nolan.Cafferky@rbsinteractive.com> writes:
> After some more digging on the mailing list, I found some comments on
> effective_cache_size.  Bringing it up from the default of 1000 does pust
> the estimated cost for the index scan below that of the sequential scan,
> but not by much.

The first-order knob for tuning indexscan vs seqscan costing is
random_page_cost.  What have you got that set to?

            regards, tom lane

Tom Lane wrote:

>The first-order knob for tuning indexscan vs seqscan costing is
>random_page_cost.  What have you got that set to?
>
>
This is currently at the default of 4. All of my other planner cost
constants are at default values as well. Dropping it to 1 drops the
estimated cost by a comparable ratio:

 Index Scan using orders_status_btree_idx on orders o
(cost=1.20..3393.20 rows=7026 width=8) (actual time=0.050..0.314
rows=105 loops=1)
   Index Cond: (order_statuses_id = $0)
   InitPlan
     ->  Seq Scan on order_statuses  (cost=0.00..1.20 rows=1 width=4)
(actual time=0.017..0.025 rows=1 loops=1)
           Filter: ((id_name)::text = 'new'::text)
 Total runtime: 0.498 ms

But, I'm guessing that random_page_cost = 1 is not a realistic value.

Nolan Cafferky <Nolan.Cafferky@rbsinteractive.com> writes:
> But, I'm guessing that random_page_cost = 1 is not a realistic value.

Well, that depends.  If all your data can be expected to fit in memory
then it is a realistic value.  (If not, you should be real careful not
to make performance decisions on the basis of test cases that *do* fit
in RAM...)

In any case, if I recall your numbers correctly you shouldn't need to
drop it nearly that far to get the thing to make the right choice.
A lot of people run with random_page_cost set to 2 or so.

            regards, tom lane

Nolan Cafferky <Nolan.Cafferky@rbsinteractive.com> writes:

But, I'm guessing that random_page_cost = 1 is not a realistic value.

Well, that depends.  If all your data can be expected to fit in memory
then it is a realistic value.  (If not, you should be real careful not
to make performance decisions on the basis of test cases that *do* fit
in RAM...)

In any case, if I recall your numbers correctly you shouldn't need to
drop it nearly that far to get the thing to make the right choice.
A lot of people run with random_page_cost set to 2 or so.

On Mon, May 01, 2006 at 07:35:02PM -0400, Tom Lane wrote:
> Nolan Cafferky <Nolan.Cafferky@rbsinteractive.com> writes:
> > But, I'm guessing that random_page_cost = 1 is not a realistic value.
>
> Well, that depends.  If all your data can be expected to fit in memory
> then it is a realistic value.  (If not, you should be real careful not
> to make performance decisions on the basis of test cases that *do* fit
> in RAM...)
>
> In any case, if I recall your numbers correctly you shouldn't need to
> drop it nearly that far to get the thing to make the right choice.
> A lot of people run with random_page_cost set to 2 or so.

Also, the index scan cost estimator comments indicate that it does a
linear interpolation between the entimated cost for a perfectly
correlated table and a table with 0 correlation, but in fact the
interpolation is exponential, or it's linear based on the *square* of
the correlation, which just doesn't make much sense.

I did some investigating on this some time ago, but never got very  far
with it. http://stats.distributed.net/~decibel/summary.txt has some
info, and http://stats.distributed.net/~decibel/ has the raw data.
Graphing that data, if you only include correlations between 0.36 and
0.5, it appears that there is a linear correlation between correlation
and index scan time.

Of course this is very coarse data and it'd be great if someone did more
research in this area, preferably using pg_bench or other tools to
generate the data so that others can test this stuff as well. But even
with as rough as this data is, it seems to provide a decent indication
that it would be better to actually interpolate linearly based on
correlation, rather than correlation^2. This is a production machine so
I'd rather not go mucking about with testing such a change here.
--
Jim C. Nasby, Sr. Engineering Consultant      jnasby@pervasive.com
Pervasive Software      http://pervasive.com    work: 512-231-6117
vcard: http://jim.nasby.net/pervasive.vcf       cell: 512-569-9461