9

The following join has very different row estimates when doing a join on the partitions vs when joining on the whole table:

CREATE TABLE m_data.ga_session (
  session_id         BIGINT                   NOT NULL,
  visitor_id         BIGINT                   NOT NULL,
  transaction_id     TEXT,

  timestamp          TIMESTAMP WITH TIME ZONE NOT NULL,
  day_id             INTEGER                  NOT NULL,
  [...]

  device_category    TEXT                     NOT NULL,
  [...]
  operating_system   TEXT

);

For all partitions:

CREATE TABLE IF NOT EXISTS m_data.ga_session_20170127 ( CHECK (day_id = 20170127) ) INHERITS (m_data.ga_session);
-- the identifier are theoretically invalid, but they get truncated to 63 chars and nevertheless work
CREATE INDEX IF NOT EXISTS "ga_session__m_tmp.normalize_device_category(ga_session.device_category)" on m_data.ga_session_20170127 USING btree (m_tmp.normalize_device_category(device_category)) ;
CREATE INDEX IF NOT EXISTS "ga_session__m_tmp.normalize_operating_system(operating_system)" on m_data.ga_session_20170127 USING btree (m_tmp.normalize_operating_system(operating_system)) ;
ANALYZE m_data.ga_session_20170127;

EXPLAIN analyse
SELECT * 
  FROM m_data.ga_session_20170127 ga_session
    JOIN m_dim_next.device ON 
      device.device_category_name = m_tmp.normalize_device_category(ga_session.device_category)
      AND device.operating_system_name = m_tmp.normalize_operating_system(ga_session.operating_system);

The statistics for these indexes on the partitions are visible:

SELECT * FROM pg_stats WHERE tablename ilike 'ga_session_20170127%';

schemaname |tablename                                                       |attname                    |inherited |null_frac   |avg_width |n_distinct   
-----------|----------------------------------------------------------------|---------------------------|----------|------------|----------|-------------
m_data     |ga_session_20170127__m_tmp.normalize_device_category(device_cat |normalize_device_category  |false     |0           |10        |3            
m_data     |ga_session_20170127__m_tmp.normalize_operating_system(operating |normalize_operating_system |false     |0           |7         |14           

This (with statistics on the index on the partition) results in the following (fine) query plan estimates: 80146 estimated, 77503 actual

Hash Join  (cost=1.95..6103.53 rows=80146 width=262) (actual time=0.121..117.204 rows=77503 loops=1)
  Hash Cond: ((COALESCE(initcap(ga_session.device_category), 'Unknown'::text) = device.device_category_name) AND (COALESCE(replace(ga_session.operating_system, '(not set)'::text, 'Unknown'::text), 'Unknown'::text) = device.operating_system_name))
  ->  Seq Scan on ga_session_20170127 ga_session  (cost=0.00..2975.03 rows=77503 width=224) (actual time=0.010..9.203 rows=77503 loops=1)
  ->  Hash  (cost=1.38..1.38 rows=38 width=38) (actual time=0.064..0.064 rows=38 loops=1)
        Buckets: 1024  Batches: 1  Memory Usage: 11kB
        ->  Seq Scan on device  (cost=0.00..1.38 rows=38 width=38) (actual time=0.006..0.019 rows=38 loops=1)
Planning time: 1.460 ms
Execution time: 120.098 ms

What doesn't work is the join on the whole table, which estimates a completely wrong row count (832 estimated vs 876237 actual).

QUERY PLAN                                                                                                                                                                                                                                             
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Hash Join  (cost=1.95..60056.78 rows=832 width=262) (actual time=0.037..1065.778 rows=876237 loops=1)                                                                                                                                                  
  Hash Cond: ((COALESCE(initcap(ga_session.device_category), 'Unknown'::text) = device.device_category_name) AND (COALESCE(replace(ga_session.operating_system, '(not set)'::text, 'Unknown'::text), 'Unknown'::text) = device.operating_system_name)) 
  ->  Append  (cost=0.00..33759.37 rows=876238 width=225) (actual time=0.005..132.070 rows=876237 loops=1)                                                                                                                                             
        ->  Seq Scan on ga_session  (cost=0.00..0.00 rows=1 width=319) (actual time=0.000..0.000 rows=0 loops=1)                                                                                                                                       
        ->  Seq Scan on ga_session_20170125 ga_session_1  (cost=0.00..3648.38 rows=94438 width=226) (actual time=0.005..10.606 rows=94438 loops=1)                                                                                                     
        ->  Seq Scan on ga_session_20170126 ga_session_2  (cost=0.00..3185.81 rows=82581 width=225) (actual time=0.014..8.982 rows=82581 loops=1)                                                                                                      
        ->  Seq Scan on ga_session_20170127 ga_session_3  (cost=0.00..2975.03 rows=77503 width=224) (actual time=0.002..8.797 rows=77503 loops=1)                                                                                                      
        ->  Seq Scan on ga_session_20170128 ga_session_4  (cost=0.00..2936.83 rows=76083 width=225) (actual time=0.003..7.873 rows=76083 loops=1)                                                                                                      
        ->  Seq Scan on ga_session_20170129 ga_session_5  (cost=0.00..3716.18 rows=96618 width=224) (actual time=0.002..9.318 rows=96618 loops=1)                                                                                                      
        ->  Seq Scan on ga_session_20170130 ga_session_6  (cost=0.00..3833.19 rows=99619 width=224) (actual time=0.002..9.453 rows=99619 loops=1)                                                                                                      
        ->  Seq Scan on ga_session_20170131 ga_session_7  (cost=0.00..3488.79 rows=90579 width=225) (actual time=0.002..8.298 rows=90579 loops=1)                                                                                                      
        ->  Seq Scan on ga_session_20170201 ga_session_8  (cost=0.00..3615.58 rows=93958 width=224) (actual time=0.002..9.199 rows=93958 loops=1)                                                                                                      
        ->  Seq Scan on ga_session_20170202 ga_session_9  (cost=0.00..3286.56 rows=85256 width=224) (actual time=0.006..8.021 rows=85256 loops=1)                                                                                                      
        ->  Seq Scan on ga_session_20170203 ga_session_10  (cost=0.00..3073.02 rows=79602 width=225) (actual time=0.002..7.727 rows=79602 loops=1)                                                                                                     
  ->  Hash  (cost=1.38..1.38 rows=38 width=38) (actual time=0.016..0.016 rows=38 loops=1)                                                                                                                                                              
        Buckets: 1024  Batches: 1  Memory Usage: 11kB                                                                                                                                                                                                  
        ->  Seq Scan on device  (cost=0.00..1.38 rows=38 width=38) (actual time=0.002..0.004 rows=38 loops=1)                                                                                                                                          
Planning time: 1.017 ms                                                                                                                                                                                                                                
Execution time: 1090.213 ms   

This then in turn results in wrong join choices (nested loops) when using that join result in more joins (not shown here).

I actually had the wrong row estimates on the partitions as well before I ran ANALYSE again on the partitions, so it seems that the query planner does not take the index based statistics into account when using the whole table.

Is there any way to make the query planner gather statistics on the level of the parent table or take the individual statistics of the partitions into account when building the query plan?

  • 2
    The question would work better for me if you provided (the relevant parts of) your actual table and index definitions. How did you implement partitions? With inheritance? Then there is no such thing as "indexes created on the whole table" Each index can only span one physical table, i.e. a single partition. – Erwin Brandstetter Feb 22 '17 at 12:02
  • Sorry, I added the information. The indexes itself are on each table (I got confused as we only call a function to create the index on all inheriting tables). – Jan Katins Feb 22 '17 at 14:00
1

Make sure not only partitions are indexed, but also the master table is indexed in same way and ANALYZEd.

This could make the planner include index-based estimates on a single partition, but ignore them on master table level.

If expression index or statistics for master table is missing, the planner is not able to infer join cardinality from this condition - even if it has perfect statistics for partitions.

It's just a guess because you did not provide full schema. Please let me know if this helps.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.