Postgres Pro
Downloads
F.3. aqo
Prev UpAppendix F. Additional Supplied ModulesHome Next

F.3. aqo

F.3.1. Installation and Setup
F.3.2. Usage
F.3.3. Reference
F.3.4. Examples
F.3.5. Author

The aqo module is a Postgres Pro Enterprise extension for cost-based query optimization. Using machine learning methods, more precisely, a modification of the k-NN algorithm, aqo improves cardinality estimation, which can optimize execution plans and, consequently, speed up query execution.

The aqo module can collect statistics on all the executed queries, excluding the queries that access system relations. The collected statistics is classified by query class. If the queries differ in their constants only, they belong to the same class. For each query class, aqo stores the cardinality quality, planning time, execution time, and execution statistics for machine learning. Based on this data, aqo builds a new query plan and uses it for the next query of the same class. aqo test runs have shown significant performance improvements for complex queries.

Important

Query optimization using the aqo module is not supported on standby.

aqo saves all the learning data (aqo_data), queries (aqo_query_texts), query settings (aqo_queries), and query execution statistics (aqo_query_stat) to files. When aqo starts, it loads this data to shared memory. You can access aqo data through functions and views.

Warning

Be aware that aqo may not work correctly right after extension upgrades that change its kernel and after Postgres Pro upgrades. Therefore, after each Postgres Pro upgrade, call aqo_reset() and run DROP EXTENSION aqo. However, after Postgres Pro minor release upgrades to versions to 13.11/14.8/15.3 or higher a call to aqo_reset() is not needed as aqo is reset automatically if required.

After a minor release upgrade, also run ALTER EXTENSION aqo UPDATE and keep in mind that aqo downgrade is impossible.

In the event of an automatic aqo reset or manually calling aqo_reset(), all the machine learning data gets lost, and new learning will be needed for future aqo use. Therefore, if some data, such as query texts, may help in aqo learning, back up this data in advance.

F.3.1. Installation and Setup

The aqo extension is included into Postgres Pro Enterprise. Once you have Postgres Pro Enterprise installed, complete the following steps to enable aqo:

  1. Add aqo to the shared_preload_libraries parameter in the postgresql.conf file: 

    shared_preload_libraries = 'aqo'

    The aqo library must be preloaded at the server startup, since adaptive query optimization needs to be enabled per cluster.

  2. Create the aqo extension using the following query: 

    CREATE EXTENSION aqo;

Once the extension is created, you can start optimizing queries.

The command 

DROP EXTENSION aqo;

will only remove aqo interface at the cluster level. aqo will still be running on all the backends while it is listed in shared_preload_libraries and at the server restart, will continue working in the operation mode specified in postgresql.conf. Besides, aqo will retain its internal state after consequent execution of DROP EXTENSION -> CREATE EXTENSION.

To remove all the data from the aqo storage, including the collected statistics, call aqo_reset(): 

SELECT aqo_reset();

To actually disable aqo at the cluster level, do the following: 

ALTER SYSTEM SET aqo.mode = 'disabled';
SELECT pg_reload_conf();
DROP EXTENSION aqo;

If you do not want aqo to be loaded at the server restart, remove the line 

shared_preload_libraries = 'aqo'

from the postgresql.conf file.

F.3.1.1. Configuration

With the default operation mode (controlled), aqo does not affect query performance. Depending on your database usage model, you should choose between the following modes:

  • intelligent — this mode auto-tunes your queries based on statistics collected per query class. See the description of the auto_tuning flag of the aqo_queries view for more details.

  • forced — this mode collects statistics for all queries altogether without any classification.

  • controlled — this mode uses the default planner for all new queries, but continues using the previously specified planning settings for already known query classes, if any.

  • learn — this mode collects statistics on all the executed queries and updates the data for query classes without auto-tuning queries.

  • frozen — this mode reads the collected statistics for already known query classes but does not collect any new data. You can use this mode to reduce the impact of aqo on query planning and execution.

  • disabled — this mode disables aqo for all queries, even for the known query classes. The collected statistics and aqo settings are saved and can be used in the future. You can use this mode to temporarily disable aqo without losing the collected statistics and configuration.

To dynamically change the aqo mode in your current session, run the following command: 

ALTER SYSTEM SET aqo.mode = 'mode';
SELECT pg_reload_conf();

where mode is the name of the operation mode to use.

F.3.2. Usage

F.3.2.1. Choosing Operation Mode for Query Optimization

If you often run queries of the same class, for example, your application limits the number of possible query classes, you can use the intelligent mode to improve planning for these queries. In this mode, aqo analyzes each query execution and stores statistics. Statistics on queries of different classes is stored separately. If performance is not improved after 50 iterations, the aqo extension falls back to the default query planner.

Note

You can view the current query plan using the standard Postgres Pro EXPLAIN command with the ANALYZE option. For details, see the Section 14.1.

Since the intelligent mode tries to learn separately for different query classes, aqo may fail to provide performance improvements if queries in the workload are of multiple different classes or if the classes of the queries in the workload are constantly changing. For such workloads, reset the aqo extension to the controlled mode, or try using the forced mode.

In the forced mode, aqo does not classify the collected statistics by query classes and tries to optimize all queries together. Therefore, aqo_query_texts, aqo_queries and aqo_query_stat views do not get updated. Machine learning data collected in any other modes is inapplicable for the forced mode and vice versa. This mode can help you optimize workloads with multiple different query classes, and it consumes less memory than the intelligent mode. However, since the forced mode lacks intelligent tuning, performance may decrease for some queries. If you see performance issues in this mode, switch aqo to the controlled mode.

In the controlled mode, aqo does not collect statistics for new query classes, so they will not be optimized. For known query classes, aqo will continue collecting statistics and using optimized planning algorithms. So use the controlled mode only after aqo learned in the learn or intelligent mode. As there are no query classes in the forced mode, switching from it to the controlled mode actually means disabling aqo.

The learn mode collects statistics from all the executed queries and updates the data for query classes. This mode is similar to the intelligent mode, except that it does not provide intelligent tuning.

If you want to reduce the impact of aqo on query planning and execution, you can use it in the frozen mode. In this mode, aqo only reads the collected statistics for already known query classes but does not collect any new data.

Switching to the disabled mode is the only way to actually disable aqo without losing the statistics and aqo settings, which are saved and can be used in the future. Queries in this mode will be executed as though there were no aqo at all.

F.3.2.2. Fine-Tuning aqo

You must have superuser rights to access aqo views and configure advanced query settings.

When run in the intelligent or learn mode, aqo assigns a unique hash value to each query class to separate the collected statistics. If you switch to the forced mode, the statistics for all untracked query classes is stored in a common query class with hash 0. You can view all the processed query classes and their corresponding hash values in the aqo_query_texts view: 

SELECT * FROM aqo_query_texts;

To find out the class, that is, hash, of a query and aqo mode, enable aqo.show_hash and aqo.show_details environment variables and execute the query. The output will contain something like this: 

...
Planning Time: 23.538 ms
...
Execution Time: 249813.875 ms
...
Using aqo: true
...
AQO mode: LEARN
...
Query hash: -2439501042637610315

Each query class has an associated separate space, called feature space, in which the statistics for this query class is collected. Each feature space has associated feature subspaces, where the information about selectivity and cardinality for each query plan node is collected.

Each query class has its own optimization settings. These settings are shown in the aqo_queries view: 

SELECT * FROM aqo_queries;

The settings available are listed in the aqo_queries View table.

You can manually change these settings to adjust optimization for a particular query class. For example: 

 -- Add a new query class into the aqo_queries view:

SET aqo.mode='intelligent';
SELECT * FROM a, b WHERE a.id=b.id;
SET aqo.mode='controlled';

 -- Disable auto_tuning, enable both learn_aqo and use_aqo 
 -- for this query class:

SELECT count(*) FROM (SELECT queryid FROM aqo_queries) AS q1,
  LATERAL aqo_queries_update(q1.queryid, NULL, true, true, false) AS q2
  WHERE queryid = (SELECT queryid FROM aqo_query_texts 
  WHERE query_text LIKE 'SELECT * FROM a, b WHERE a.id=b.id;');

 -- Run EXPLAIN ANALYZE until the plan changes:

EXPLAIN ANALYZE SELECT * FROM a, b WHERE a.id=b.id;
EXPLAIN ANALYZE SELECT * FROM a, b WHERE a.id=b.id;

 -- Disable learning to stop statistics collection 
 -- and use the optimized plan:

 SELECT count(*) FROM (SELECT queryid FROM aqo_queries) AS q1,
  LATERAL aqo_queries_update(q1.queryid, NULL, false, true, false) AS q2 
  WHERE queryid = (SELECT queryid FROM aqo_query_texts 
  WHERE query_text LIKE 'SELECT * FROM a, b WHERE a.id=b.id;');

To stop intelligent tuning for a particular query class, disable the auto_tuning setting: 

SELECT count(*) FROM (SELECT queryid FROM aqo_queries) AS q1,
  LATERAL aqo_queries_update(q1.queryid, NULL, true, true, false) AS q2
  WHERE queryid = 'hash');

where hash is the hash value for this query class. As a result, aqo disables automatic change of the learn_aqo and use_aqo settings.

To disable further learning for a particular query class, use the following command: 

SELECT count(*) FROM (SELECT queryid FROM aqo_queries) AS q1,
  LATERAL aqo_queries_update(q1.queryid, NULL, false, true, false) AS q2
  WHERE queryid = 'hash');

where hash is the hash value for this query class.

To fully disable aqo for all queries and use the default Postgres Pro query planner, run: 

SELECT count(*) FROM (SELECT queryid FROM aqo_queries) AS q1,
  LATERAL aqo_queries_update(q1.queryid, NULL, false, false, false) AS q2
  WHERE queryid IN (SELECT queryid FROM aqo_query_texts);

F.3.3. Reference

F.3.3.1. Configuration Parameters

aqo.mode (text)

Defines the aqo operation mode. Possible values are listed in Section F.3.1.1.

Default: controlled.

aqo.show_hash (boolean)

Show a hash value that is computed from a query tree and uniquely identifies the class of queries or class of plan nodes. Starting with Postgres Pro 14, aqo uses the native query ID to identify a query class for consistency with other extensions, such as pg_stat_statements. So, the query ID can be taken from the Query Identifier field in EXPLAIN ANALYZE output of a query.

Default: off.

aqo.show_details (boolean)

Add some details to EXPLAIN output of a query, such as the prediction or feature-subspace hash, and show some additional aqo-specific on-screen information.

Default: off.

aqo.join_threshold (integer)

Ignore queries that contain smaller number of joins, which means that statistics for such queries will not be collected.

Default: 3.

aqo.statement_timeout (integer)

Defines the initial value of the smart statement timeout, in milliseconds, which is needed to limit the execution time when manually training aqo on special queries with a poor cardinality forecast. aqo can dynamically change the value of the smart statement timeout during this training. When the cardinality estimation error on nodes exceeds 0.1, the value of aqo.statement_timeout is automatically incremented exponentially, but remains not greater than statement_timeout.

Default: 0.

aqo.force_collect_stat (boolean)

Gather statistics on query executions even in the disabled mode. Although no predictions are made, some overhead will be added.

Default: off.

aqo.dsm_size_max (integer)

Defines the maximum size of dynamic shared memory, in MB, that aqo can allocate to store learning data. When this number is exceeded, an attempt to load the aqo_data view will fail with “out of memory” error.

Default: 100.

aqo.fs_max_items (integer)

Defines the maximum number of feature spaces that aqo can operate with. When this number is exceeded, learning on new query classes will no longer occur, and they will not appear in the views accordingly.

Default: 10000.

aqo.fss_max_items (integer)

Defines the maximum number of feature subspaces that aqo can operate with. When this number is exceeded, the selectivity and cardinality for new query plan nodes will no longer be collected, and new feature subspaces will not appear in the aqo_data view accordingly.

Default: 100000.

Enables searching neighbors with the same feature subspace among different query classes.

Default: off.

aqo.querytext_max_size (integer)

Defines the maximum size of the query in the aqo_query_texts view.

Default: 1000.

aqo.min_neighbors_for_predicting (integer)

Defines the minimum number of neighbors needed for the cardinality prediction. If there are fewer of them, aqo will not make any prediction.

Default: 3.

aqo.predict_with_few_neighbors (boolean)

Enables aqo to make predictions with fewer neighbors than were found.

Default: on.

F.3.3.2. Views

F.3.3.2.1. aqo_query_texts

The aqo_query_texts view classifies all the query classes processed by aqo. For each query class, the view shows the text of the first analyzed query of this class.

Table F.2. aqo_query_texts View

Column NameDescription
queryidStores the query ID, that is, the feature-space hash, that uniquely identifies the query class.
query_textProvides the text of the first analyzed query of the given class.

F.3.3.2.2. aqo_queries

The aqo_queries view shows optimization settings for different query classes.

Table F.3. aqo_queries View

SettingDescription
queryidStores the query ID that uniquely identifies the query class.
learn_aqoEnables statistics collection for this query class.
use_aqoEnables the aqo cardinality prediction for the next execution of this query class. If the cost estimation model is inaccurate, this may slow down query execution.
fspace_hashProvides a unique identifier of the separate space in which the statistics for this query class is collected. By default, fspace_hash is equal to queryid. You can change this setting to a different queryid to optimize different query classes together. It may decrease the amount of memory for models and even improve query execution performance. However, changing this setting may cause unexpected aqo behavior, so make sure to use it only if you know what you are doing.
auto_tuning

Shows whether aqo can dynamically change use_aqo and learn_aqo settings for this query class. By default, auto-tuning is only enabled in the intelligent mode.

In more detail, when auto_tuning is on, if for several successive executions of a query for which use_aqo is off, the cardinality error remains sufficiently small and stable, aqo turns on use_aqo.

For queries with learn_aqo=true (it is so for new queries), several first executions are done both using aqo and without it. The faster the query is executed compared to the execution with the standard planner, the more likely aqo will be used for the next query execution. If after a certain number of executions the execution time with aqo appears to be worse than with the standard planner, aqo will never be used for this query class: auto_tuning, use_aqo and learn_aqo are set to off.

smart_timeoutShows the value of smart statement timeout for this query class.
count_increase_timeoutShows how many times the smart statement timeout increased for this query class.

F.3.3.2.3. aqo_data

The aqo_data view shows machine learning data for cardinality estimation refinement. To forget all the collected statistics for a particular query class, you can delete all rows from aqo_data with the corresponding fs.

Table F.4. aqo_data View

DataDescription
fsFeature-space hash.
fssFeature-subspace hash.
nfeaturesFeature-subspace size for the query plan node.
featuresLogarithm of the selectivity which the cardinality prediction is based on.
targetsCardinality logarithm for the query plan node.
reliabilityEquals:

  • 1 — indicates data obtained after normal execution of a query

  • 0.1 — indicates data obtained from a partially executed node (not needed as unreliable)

  • 0.9 — indicates data obtained from a finished node, but from a partially executed statement

oidsList of IDs of tables that were involved in the prediction for this node.

F.3.3.2.4. aqo_query_stat

The aqo_query_stat view shows statistics on query execution, by query class. The aqo extension uses this data when the auto_tuning option is enabled for a particular query class.

Table F.5. aqo_query_stat View

DataDescription
execution_time_with_aqoExecution time for queries run with aqo enabled.
execution_time_without_aqoExecution time for queries run with aqo disabled.
planning_time_with_aqoPlanning time for queries run with aqo enabled.
planning_time_without_aqoPlanning time for queries run with aqo disabled.
cardinality_error_with_aqoCardinality estimation error in the selected query plans with aqo enabled.
cardinality_error_without_aqoCardinality estimation error in the selected query plans with aqo disabled.
executions_with_aqoNumber of queries run with aqo enabled.
executions_without_aqoNumber of queries run with aqo disabled.

F.3.3.3. Functions

aqo adds several functions to Postgres Pro catalog.

F.3.3.3.1. Storage Management Functions

Important

Functions aqo_queries_update, aqo_query_texts_update, aqo_query_stat_update, and aqo_data_update modify data files underlying aqo views. Therefore, call these functions only if you understand the logic of adaptive query optimization.

aqo_cleanup() → setof integer

Removes data related to query classes that are linked (may be partially) with removed relations. Returns the number of removed feature spaces (classes) and feature subspaces. Insensitive to removing other objects.

aqo_enable_class (queryid bigint) → void

Sets learn_aqo, use_aqo and auto_tuning (only in the intelligent mode) to true for a given query class.

aqo_disable_class (queryid bigint) → void

Sets learn_aqo, use_aqo and auto_tuning (only in the intelligent mode) to false for a given query class.

aqo_drop_class (queryid bigint) → integer

Removes all data related to a given query class from the aqo storage. Returns the number of records removed from the aqo storage.

aqo_reset() → bigint

Removes data from the aqo storage: machine learning data, query texts, statistics and query class preferences. Returns the number of records removed from the aqo storage.

aqo_queries_update (queryid bigint, fs bigint, learn_aqo boolean, use_aqo boolean, auto_tuning boolean) → boolean

Assigns new values to the following settings in the aqo_queries view for a given query class: fspace_hash, learn_aqo, use_aqo and auto_tuning. NULL value means “leave as is”.

aqo_query_texts_update (queryid bigint, query_text text) → boolean

Updates or inserts a record in a data file underlying the aqo_query_texts view for a given queryid.

aqo_query_stat_update (queryid bigint, execution_time_with_aqo double precision[], execution_time_without_aqo double precision[], planning_time_with_aqo double precision[], planning_time_without_aqo double precision[], cardinality_error_with_aqo double precision[], cardinality_error_without_aqo double precision[], executions_with_aqo bigint[], executions_without_aqo bigint[]) → boolean

Updates or inserts a record in a data file underlying the aqo_query_stat view for a given queryid.

aqo_data_update (fs bigint, fss integer, nfeatures integer, features double precision[][], targets double precision[], reliability double precision[], oids oid[]) → boolean

Updates or inserts a record in a data file underlying the aqo_data view for given fs and fss.

F.3.3.3.2. Memory Management Functions
aqo_memory_usage () → setof record

Displays sizes of aqo memory contexts and hash tables.

F.3.3.3.3. Analytics Functions
aqo_cardinality_error (controlled boolean) → setof record

Shows the cardinality error for each query class. If controlled is true, shows the error of the last execution with aqo enabled. If controlled is false, returns the average cardinality error for all logged executions with aqo disabled.

aqo_execution_time (controlled boolean) → setof record

Shows the execution time for each query class. If controlled is true, shows the execution time of the last execution with aqo enabled. If controlled is false, returns the average execution time for all logged executions with aqo disabled.

F.3.4. Examples

Example F.1. Learning on a Query

Consider optimization of a query using aqo.

When the query is executed for the first time, it is missing in tables underlying aqo views. So there is no data for predicting with aqo for each plan node, and “AQO not used” lines appear in the EXPLAIN output: 

postgres=# EXPLAIN (ANALYZE, SUMMARY OFF, TIMING OFF) select count(*) from score join course on score.cno=course.cno join student on score.sno=student.sno where degree<90 and test_preparation = 0;
                                                QUERY PLAN
----------------------------------------------------------------------------------------------------------
 Aggregate  (cost=308.28..308.29 rows=1 width=8) (actual rows=1 loops=1)
   AQO not used, fss=0
   ->  Hash Join  (cost=124.80..299.47 rows=3526 width=0) (actual rows=3649 loops=1)
         AQO not used, fss=2128507884
         Hash Cond: (score.sno = student.sno)
         ->  Hash Join  (cost=16.30..181.70 rows=3526 width=4) (actual rows=3649 loops=1)
               AQO not used, fss=-303037802
               Hash Cond: (score.cno = course.cno)
               ->  Seq Scan on score  (cost=0.00..156.00 rows=3526 width=8) (actual rows=3649 loops=1)
                     AQO not used, fss=-636613046
                     Filter: ((degree < 90) AND (test_preparation = 0))
                     Rows Removed by Filter: 1351
               ->  Hash  (cost=12.80..12.80 rows=280 width=4) (actual rows=10 loops=1)
                     Buckets: 1024  Batches: 1  Memory Usage: 9kB
                     ->  Seq Scan on course  (cost=0.00..12.80 rows=280 width=4) (actual rows=10 loops=1)
                           AQO not used, fss=-1076069505
         ->  Hash  (cost=71.00..71.00 rows=3000 width=4) (actual rows=3000 loops=1)
               Buckets: 4096  Batches: 1  Memory Usage: 138kB
               ->  Seq Scan on student  (cost=0.00..71.00 rows=3000 width=4) (actual rows=3000 loops=1)
                     AQO not used, fss=-1838231581
 Using aqo: true
 AQO mode: LEARN
 Query hash: -727505571757520766
 JOINS: 2
(24 rows)

If there is no information on a certain node in the aqo_data view, aqo will add the appropriate record there for future learning and predictions except for nodes with fss=0 in the EXPLAIN output. As each of features and targets in the aqo_data view is a logarithm to base e, to get the actual value, raise e to this power. For example: exp(0): 

         fs          |     fss     | nfeatures |                                      features                                      |       targets       | reliability |        oids
---------------------+-------------+-----------+------------------------------------------------------------------------------------+---------------------+-------------+---------------------
 -727505571757520766 |  2128507884 |         4 | {{-0.03438753143452488,-5.634789603169249,-0.3149847743198556,-8.006367567650246}} | {8.202208436436448} | {1}         | {16579,16555,16563}
 -727505571757520766 | -1076069505 |         0 |                                                                                    | {2.302585092994046} | {1}         | {16555}
 -727505571757520766 | -1838231581 |         0 |                                                                                    | {8.006367567650246} | {1}         | {16563}
 -727505571757520766 |  -303037802 |         3 | {{-0.03438753143452488,-5.634789603169249,-0.3149847743198556}}                    | {8.202208436436448} | {1}         | {16579,16555}
 -727505571757520766 |  -636613046 |         2 | {{-0.03438753143452488,-0.3149847743198556}}                                       | {8.202208436436448} | {1}         | {16579}
(6 rows)

When the query is executed for the second time, aqo recognizes the query and makes a prediction. Pay attention to the cardinality predicted by aqo and the value of aqo error (“error=0%”). 

postgres=# EXPLAIN (ANALYZE, SUMMARY OFF, TIMING OFF) select count(*) from score join course on score.cno=course.cno join student on score.sno=student.sno where degree<90 and test_preparation = 0;
                                               QUERY PLAN
---------------------------------------------------------------------------------------------------------
 Aggregate  (cost=305.86..305.87 rows=1 width=8) (actual rows=1 loops=1)
   AQO not used, fss=0
   ->  Hash Join  (cost=121.42..296.74 rows=3649 width=0) (actual rows=3649 loops=1)
         AQO: rows=3649, error=0%, fss=2128507884
         Hash Cond: (score.sno = student.sno)
         ->  Hash Join  (cost=12.93..178.65 rows=3649 width=4) (actual rows=3649 loops=1)
               AQO: rows=3649, error=0%, fss=-303037802
               Hash Cond: (score.cno = course.cno)
               ->  Seq Scan on score  (cost=0.00..156.00 rows=3649 width=8) (actual rows=3649 loops=1)
                     AQO: rows=3649, error=0%, fss=-636613046
                     Filter: ((degree < 90) AND (test_preparation = 0))
                     Rows Removed by Filter: 1351
               ->  Hash  (cost=12.80..12.80 rows=10 width=4) (actual rows=10 loops=1)
                     Buckets: 1024  Batches: 1  Memory Usage: 9kB
                     ->  Seq Scan on course  (cost=0.00..12.80 rows=10 width=4) (actual rows=10 loops=1)
                           AQO: rows=10, error=0%, fss=-1076069505
         ->  Hash  (cost=71.00..71.00 rows=3000 width=4) (actual rows=3000 loops=1)
               Buckets: 4096  Batches: 1  Memory Usage: 138kB
               ->  Seq Scan on student  (cost=0.00..71.00 rows=3000 width=4) (actual rows=3000 loops=1)
                     AQO: rows=3000, error=0%, fss=-1838231581
 Using aqo: true
 AQO mode: LEARN
 Query hash: -727505571757520766
 JOINS: 2
(24 rows)

In case of an error, values of features and targets must change, but as there was no error above, they did not change. 

         fs          |     fss     | nfeatures |                                      features                                      |       targets       | reliability |        oids
---------------------+-------------+-----------+------------------------------------------------------------------------------------+---------------------+-------------+---------------------
 -727505571757520766 |  2128507884 |         4 | {{-0.03438753143452488,-5.634789603169249,-0.3149847743198556,-8.006367567650246}} | {8.202208436436448} | {1}         | {16579,16555,16563}
 -727505571757520766 | -1076069505 |         0 |                                                                                    | {2.302585092994046} | {1}         | {16555}
 -727505571757520766 | -1838231581 |         0 |                                                                                    | {8.006367567650246} | {1}         | {16563}
 -727505571757520766 |  -303037802 |         3 | {{-0.03438753143452488,-5.634789603169249,-0.3149847743198556}}                    | {8.202208436436448} | {1}         | {16579,16555}
 -727505571757520766 |  -636613046 |         2 | {{-0.03438753143452488,-0.3149847743198556}}                                       | {8.202208436436448} | {1}         | {16579}
(6 rows)

Let's change a constant in the query, and you will notice that the prediction is made with an error: 

    postgres=# EXPLAIN (ANALYZE, SUMMARY OFF, TIMING OFF) select count(*) from score join course on score.cno=course.cno join student on score.sno=student.sno where degree<80 and test_preparation = 0;
                                                   QUERY PLAN
---------------------------------------------------------------------------------------------------------
 Aggregate  (cost=305.86..305.87 rows=1 width=8) (actual rows=1 loops=1)
   AQO not used, fss=0
   ->  Hash Join  (cost=121.42..296.74 rows=3649 width=0) (actual rows=3551 loops=1)
         AQO: rows=3649, error=3%, fss=2128507884
         Hash Cond: (score.sno = student.sno)
         ->  Hash Join  (cost=12.93..178.65 rows=3649 width=4) (actual rows=3551 loops=1)
               AQO: rows=3649, error=3%, fss=-303037802
               Hash Cond: (score.cno = course.cno)
               ->  Seq Scan on score  (cost=0.00..156.00 rows=3649 width=8) (actual rows=3551 loops=1)
                     AQO: rows=3649, error=3%, fss=-636613046
                     Filter: ((degree < 80) AND (test_preparation = 0))
                     Rows Removed by Filter: 1449
               ->  Hash  (cost=12.80..12.80 rows=10 width=4) (actual rows=10 loops=1)
                     Buckets: 1024  Batches: 1  Memory Usage: 9kB
                     ->  Seq Scan on course  (cost=0.00..12.80 rows=10 width=4) (actual rows=10 loops=1)
                           AQO: rows=10, error=0%, fss=-1076069505
         ->  Hash  (cost=71.00..71.00 rows=3000 width=4) (actual rows=3000 loops=1)
               Buckets: 4096  Batches: 1  Memory Usage: 138kB
               ->  Seq Scan on student  (cost=0.00..71.00 rows=3000 width=4) (actual rows=3000 loops=1)
                     AQO: rows=3000, error=0%, fss=-1838231581
 Using aqo: true
 AQO mode: LEARN
 Query hash: -727505571757520766
 JOINS: 2
(24 rows)

However, instead of recalculating features and targets, aqo added new values of selectivity and cardinality for this query to aqo_data: 

         fs          |     fss     | nfeatures |                                                                               features                                                                               |                targets                | reliability |        oids
---------------------+-------------+-----------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------+---------------------------------------+-------------+---------------------
 -727505571757520766 |  1141621836 |         0 |                                                                                                                                                                      | {0}                                   | {1}         | {16579,16555,16563}
 -727505571757520766 |  2128507884 |         4 | {{-0.030949078292235133,-5.634789603169249,-0.3149847743198556,-8.006367567650246},{-0.34221288089027607,-5.634789603169249,-0.3149847743198556,-8.006367567650246}} | {8.202208436436448,8.174984532943087} | {1,1}       | {16579,16555,16563}
 -727505571757520766 | -1076069505 |         0 |                                                                                                                                                                      | {2.302585092994046}                   | {1}         | {16555}
 -727505571757520766 | -1838231581 |         0 |                                                                                                                                                                      | {8.006367567650246}                   | {1}         | {16563}
 -727505571757520766 |  -303037802 |         3 | {{-0.030949078292235133,-5.634789603169249,-0.3149847743198556},{-0.34221288089027607,-5.634789603169249,-0.3149847743198556}}                                       | {8.202208436436448,8.174984532943087} | {1,1}       | {16579,16555}
 -727505571757520766 |  -636613046 |         2 | {{-0.030949078292235133,-0.3149847743198556},{-0.34221288089027607,-0.3149847743198556}}

Now the prediction has no error: 

postgres=# EXPLAIN (ANALYZE, SUMMARY OFF, TIMING OFF) select count(*) from score join course on score.cno=course.cno join student on score.sno=student.sno where degree<80 and test_preparation = 0;
                                               QUERY PLAN
---------------------------------------------------------------------------------------------------------
 Aggregate  (cost=305.10..305.11 rows=1 width=8) (actual rows=1 loops=1)
   AQO not used, fss=0
   ->  Hash Join  (cost=121.42..296.22 rows=3551 width=0) (actual rows=3551 loops=1)
         AQO: rows=3551, error=0%, fss=2128507884
         Hash Cond: (score.sno = student.sno)
         ->  Hash Join  (cost=12.93..178.39 rows=3551 width=4) (actual rows=3551 loops=1)
               AQO: rows=3551, error=0%, fss=-303037802
               Hash Cond: (score.cno = course.cno)
               ->  Seq Scan on score  (cost=0.00..156.00 rows=3551 width=8) (actual rows=3551 loops=1)
                     AQO: rows=3551, error=0%, fss=-636613046
                     Filter: ((degree < 80) AND (test_preparation = 0))
                     Rows Removed by Filter: 1449
               ->  Hash  (cost=12.80..12.80 rows=10 width=4) (actual rows=10 loops=1)
                     Buckets: 1024  Batches: 1  Memory Usage: 9kB
                     ->  Seq Scan on course  (cost=0.00..12.80 rows=10 width=4) (actual rows=10 loops=1)
                           AQO: rows=10, error=0%, fss=-1076069505
         ->  Hash  (cost=71.00..71.00 rows=3000 width=4) (actual rows=3000 loops=1)
               Buckets: 4096  Batches: 1  Memory Usage: 138kB
               ->  Seq Scan on student  (cost=0.00..71.00 rows=3000 width=4) (actual rows=3000 loops=1)
                     AQO: rows=3000, error=0%, fss=-1838231581
 Using aqo: true
 AQO mode: LEARN
 Query hash: -727505571757520766
 JOINS: 2
(24 rows)


Example F.2. Using the aqo_query_stat View

The aqo_query_stats view shows statistics on the query planning time, query execution time and cardinality error. Based on this data you can make a decision whether to use aqo predictions for different query classes.

Let's query the aqo_query_stats view: 

select queryid, cardinality_error_with_aqo, cardinality_error_without_aqo,execution_time_with_aqo, execution_time_without_aqo, planning_time_with_aqo, planning_time_without_aqo from aqo_query_stat \gx
-[ RECORD 1 ]-----------------+------------------------------------------------------------------------------------------------------------
queryid                       | 8041624334006338922
cardinality_error_with_aqo    | {0.14932737556062836,0,0.507421202801325,0.00040469447777891077}
cardinality_error_without_aqo | {0.1493979460962751,0.018403615483185476}
execution_time_with_aqo       | {0.004760108,0.008743075,0.006608304,0.012392751}
execution_time_without_aqo    | {0.005775926,0.012730316}
planning_time_with_aqo        | {0.006927997,0.004247339,0.005005022,0.004169717}
planning_time_without_aqo     | {0.001783542,0.001706121}

The retrieved data is for the query from Example F.1, which was executed once without aqo for each of the parameters degree<80 and degree<90 and twice with aqo for each of these parameters. It is clear that with aqo, the cardinality error decreases to 0.0004, while the minimum cardinality error without aqo is 0.15. Besides, the execution time with aqo is lower than without it. So the conclusion is that aqo learns well on this query, and the prediction can be used for this query class.


F.3.5. Author

Oleg Ivanov

Prev Up Next
F.2. amcheck Home F.4. auth_delay
epub pdf