The pg_pathman module provides optimized partitioning mechanism and functions to manage partitions.
Partitioning means splitting one large table into smaller pieces. Each row in such table is moved to a single partition according to the partitioning key. PostgreSQL supports partitioning via table inheritance: each partition must be created as a child table with CHECK CONSTRAINT. For example:
CREATE TABLE test (id SERIAL PRIMARY KEY, title TEXT);
CREATE TABLE test_1 (CHECK ( id >= 100 AND id < 200 )) INHERITS (test);
CREATE TABLE test_2 (CHECK ( id >= 200 AND id < 300 )) INHERITS (test);
Despite the flexibility, this approach forces the planner to perform an exhaustive search and to check constraints on each partition to determine whether it should be present in the plan or not. Large amount of partitions may result in significant planning overhead.
The pg_pathman module features partition managing functions and optimized planning mechanism which utilizes knowledge of the partitions' structure. It stores partitioning configuration in the pathman_config table; each row contains a single entry for a partitioned table (relation name, partitioning column and its type). During the initialization stage the pg_pathman module caches some information about child partitions in the shared memory, which is used later for plan construction. Before a SELECT query is executed, pg_pathman traverses the condition tree in search of expressions like:
VARIABLE OP CONST
where VARIABLE is a partitioning key, OP is a comparison operator (supported operators are =, <, <=, >, >=), CONST is a scalar value. For example:
WHERE id = 150
Based on the partitioning type and condition's operator, pg_pathman searches for the corresponding partitions and builds the plan. Currently pg_pathman supports two partitioning schemes:
- RANGE - maps rows to partitions using partitioning key ranges assigned to each partition. Optimization is achieved by using the binary search algorithm;
- HASH - maps rows to partitions using a generic hash function (only integer attributes are supported at the moment).
More interesting features are yet to come. Stay tuned!
- Replace INSERT triggers with a custom node (aka PartitionFilter)
- Implement concurrent partitioning (much more responsive)
- Implement HASH partitioning for non-integer attributes
- Optimize hash join (both tables are partitioned by join key)
- Implement LIST partitioning scheme
To install pg_pathman, execute this in the module's directory:
make install USE_PGXS=1
Modify the shared_preload_libraries parameter in postgresql.conf as following:
shared_preload_libraries = 'pg_pathman'
It is essential to restart the PostgreSQL instance. After that, execute the following query in psql:
CREATE EXTENSION pg_pathman;
Done! Now it's time to setup your partitioning schemes.
Important: Don't forget to set the
PG_CONFIGvariable in case you want to testpg_pathmanon a custom build of PostgreSQL. Read more here.
create_hash_partitions(relation TEXT,
attribute TEXT,
partitions_count INTEGER)Performs HASH partitioning for relation by integer key attribute. Creates partitions_count partitions and trigger on INSERT. All the data will be automatically copied from the parent to partitions.
create_range_partitions(relation TEXT,
attribute TEXT,
start_value ANYELEMENT,
interval ANYELEMENT,
premake INTEGER DEFAULT NULL)
create_range_partitions(relation TEXT,
attribute TEXT,
start_value ANYELEMENT,
interval INTERVAL,
premake INTEGER DEFAULT NULL)Performs RANGE partitioning for relation by partitioning key attribute. start_value argument specifies initial value, interval sets the range of values in a single partition, premake is the number of premade partitions (if not set then pathman tries to determine it based on attribute values). All the data will be automatically copied from the parent to partitions.
create_partitions_from_range(relation TEXT,
attribute TEXT,
start_value ANYELEMENT,
end_value ANYELEMENT,
interval ANYELEMENT)
create_partitions_from_range(relation TEXT,
attribute TEXT,
start_value ANYELEMENT,
end_value ANYELEMENT,
interval INTERVAL)Performs RANGE-partitioning from specified range for relation by partitioning key attribute. Data will be copied to partitions as well.
create_hash_update_trigger(parent TEXT)Creates the trigger on UPDATE for HASH partitions. The UPDATE trigger isn't created by default because of the overhead. It's useful in cases when the key attribute might change.
create_range_update_trigger(parent TEXT)Same as above, but for a RANGE-partitioned table.
split_range_partition(partition TEXT, value ANYELEMENT)Split RANGE partition in two by value.
merge_range_partitions(partition1 TEXT, partition2 TEXT)Merge two adjacent RANGE partitions. First, data from partition2 is copied to partition1, then partition2 is removed.
append_range_partition(p_relation TEXT)Append new RANGE partition.
prepend_range_partition(p_relation TEXT)Prepend new RANGE partition.
add_range_partition(relation TEXT,
start_value ANYELEMENT,
end_value ANYELEMENT)Create new RANGE partition for relation with specified range bounds.
drop_range_partition(partition TEXT)Drop RANGE partition and all its data.
attach_range_partition(relation TEXT,
partition TEXT,
start_value ANYELEMENT,
end_value ANYELEMENT)Attach partition to the existing RANGE-partitioned relation. The attached table must have exactly the same structure as the parent table, including the dropped columns.
detach_range_partition(partition TEXT)Detach partition from the existing RANGE-partitioned relation.
disable_partitioning(relation TEXT)Permanently disable pg_pathman partitioning mechanism for the specified parent table and remove the insert trigger if it exists. All partitions and data remain unchanged.
pg_pathman provides a couple of custom plan nodes which aim to reduce execution time, namely:
RuntimeAppend(overridesAppendplan node)RuntimeMergeAppend(overridesMergeAppendplan node)
RuntimeAppend and RuntimeMergeAppend have much in common: they come in handy in a case when WHERE condition takes form of:
VARIABLE OP PARAM
This kind of expressions can no longer be optimized at planning time since the parameter's value is not known until the execution stage takes place. The problem can be solved by embedding the WHERE condition analysis routine into the original Append's code, thus making it pick only required scans out of a whole bunch of planned partition scans. This effectively boils down to creation of a custom node capable of performing such a check.
There are at least several cases that demonstrate usefulness of these nodes:
/* create table we're going to partition */
CREATE TABLE partitioned_table(id INT NOT NULL, payload REAL);
/* insert some data */
INSERT INTO partitioned_table
SELECT generate_series(1, 1000), random();
/* perform partitioning */
SELECT create_hash_partitions('partitioned_table', 'id', 100);
/* create ordinary table */
CREATE TABLE some_table AS SELECT generate_series(1, 100) AS VAL;
id = (select ... limit 1)
EXPLAIN (COSTS OFF, ANALYZE) SELECT * FROM partitioned_table
WHERE id = (SELECT * FROM some_table LIMIT 1);
QUERY PLAN
----------------------------------------------------------------------------------------------------
Custom Scan (RuntimeAppend) (actual time=0.030..0.033 rows=1 loops=1)
InitPlan 1 (returns $0)
-> Limit (actual time=0.011..0.011 rows=1 loops=1)
-> Seq Scan on some_table (actual time=0.010..0.010 rows=1 loops=1)
-> Seq Scan on partitioned_table_70 partitioned_table (actual time=0.004..0.006 rows=1 loops=1)
Filter: (id = $0)
Rows Removed by Filter: 9
Planning time: 1.131 ms
Execution time: 0.075 ms
(9 rows)
/* disable RuntimeAppend node */
SET pg_pathman.enable_runtimeappend = f;
EXPLAIN (COSTS OFF, ANALYZE) SELECT * FROM partitioned_table
WHERE id = (SELECT * FROM some_table LIMIT 1);
QUERY PLAN
----------------------------------------------------------------------------------
Append (actual time=0.196..0.274 rows=1 loops=1)
InitPlan 1 (returns $0)
-> Limit (actual time=0.005..0.005 rows=1 loops=1)
-> Seq Scan on some_table (actual time=0.003..0.003 rows=1 loops=1)
-> Seq Scan on partitioned_table_0 (actual time=0.014..0.014 rows=0 loops=1)
Filter: (id = $0)
Rows Removed by Filter: 6
-> Seq Scan on partitioned_table_1 (actual time=0.003..0.003 rows=0 loops=1)
Filter: (id = $0)
Rows Removed by Filter: 5
... /* more plans follow */
Planning time: 1.140 ms
Execution time: 0.855 ms
(306 rows)
id = ANY (select ...)
EXPLAIN (COSTS OFF, ANALYZE) SELECT * FROM partitioned_table
WHERE id = any (SELECT * FROM some_table limit 4);
QUERY PLAN
-----------------------------------------------------------------------------------------------------------
Nested Loop (actual time=0.025..0.060 rows=4 loops=1)
-> Limit (actual time=0.009..0.011 rows=4 loops=1)
-> Seq Scan on some_table (actual time=0.008..0.010 rows=4 loops=1)
-> Custom Scan (RuntimeAppend) (actual time=0.002..0.004 rows=1 loops=4)
-> Seq Scan on partitioned_table_70 partitioned_table (actual time=0.001..0.001 rows=10 loops=1)
-> Seq Scan on partitioned_table_26 partitioned_table (actual time=0.002..0.003 rows=9 loops=1)
-> Seq Scan on partitioned_table_27 partitioned_table (actual time=0.001..0.002 rows=20 loops=1)
-> Seq Scan on partitioned_table_63 partitioned_table (actual time=0.001..0.002 rows=9 loops=1)
Planning time: 0.771 ms
Execution time: 0.101 ms
(10 rows)
/* disable RuntimeAppend node */
SET pg_pathman.enable_runtimeappend = f;
EXPLAIN (COSTS OFF, ANALYZE) SELECT * FROM partitioned_table
WHERE id = any (SELECT * FROM some_table limit 4);
QUERY PLAN
-----------------------------------------------------------------------------------------
Nested Loop Semi Join (actual time=0.531..1.526 rows=4 loops=1)
Join Filter: (partitioned_table.id = some_table.val)
Rows Removed by Join Filter: 3990
-> Append (actual time=0.190..0.470 rows=1000 loops=1)
-> Seq Scan on partitioned_table (actual time=0.187..0.187 rows=0 loops=1)
-> Seq Scan on partitioned_table_0 (actual time=0.002..0.004 rows=6 loops=1)
-> Seq Scan on partitioned_table_1 (actual time=0.001..0.001 rows=5 loops=1)
-> Seq Scan on partitioned_table_2 (actual time=0.002..0.004 rows=14 loops=1)
... /* 96 scans follow */
-> Materialize (actual time=0.000..0.000 rows=4 loops=1000)
-> Limit (actual time=0.005..0.006 rows=4 loops=1)
-> Seq Scan on some_table (actual time=0.003..0.004 rows=4 loops=1)
Planning time: 2.169 ms
Execution time: 2.059 ms
(110 rows)
NestLoopinvolving a partitioned table, which is omitted since it's occasionally shown above.
In case you're interested, you can read more about custom nodes at Alexander Korotkov's blog.
- You can easily add partition column containing the names of the underlying partitions using the system attribute called tableoid:
SELECT tableoid::regclass AS partition, * FROM partitioned_table;
- Though indices on a parent table aren't particularly useful (since it's empty), they act as prototypes for indices on partitions. For each index on the parent table,
pg_pathmanwill create a similar index on every partition.
Consider an example of HASH partitioning. First create a table with some integer column:
CREATE TABLE items (
id SERIAL PRIMARY KEY,
name TEXT,
code BIGINT);
INSERT INTO items (id, name, code)
SELECT g, md5(g::text), random() * 100000
FROM generate_series(1, 100000) as g;
Now run the create_hash_partitions() function with appropriate arguments:
SELECT create_hash_partitions('items', 'id', 100);
This will create new partitions and move the data from parent to partitions.
Here's an example of the query performing filtering by partitioning key:
SELECT * FROM items WHERE id = 1234;
id | name | code
------+----------------------------------+------
1234 | 81dc9bdb52d04dc20036dbd8313ed055 | 1855
(1 row)
EXPLAIN SELECT * FROM items WHERE id = 1234;
QUERY PLAN
------------------------------------------------------------------------------------
Append (cost=0.28..8.29 rows=0 width=0)
-> Index Scan using items_34_pkey on items_34 (cost=0.28..8.29 rows=0 width=0)
Index Cond: (id = 1234)
Notice that the Append node contains only one child scan which corresponds to the WHERE clause.
Important: pay attention to the fact that
pg_pathmanexcludes the parent table from the query plan.
To access parent table use ONLY modifier:
EXPLAIN SELECT * FROM ONLY items;
QUERY PLAN
------------------------------------------------------
Seq Scan on items (cost=0.00..0.00 rows=1 width=45)
Consider an example of RANGE partitioning. Let's create a table containing some dummy logs:
CREATE TABLE journal (
id SERIAL,
dt TIMESTAMP NOT NULL,
level INTEGER,
msg TEXT
);
-- similar index will also be created for each partition
CREATE INDEX ON journal(dt);
-- generate some data
INSERT INTO journal (dt, level, msg)
SELECT g, random() * 6, md5(g::text)
FROM generate_series('2015-01-01'::date, '2015-12-31'::date, '1 minute') as g;
Run the create_range_partitions() function to create partitions so that each partition would contain the data for one day:
SELECT create_range_partitions('journal', 'dt', '2015-01-01'::date, '1 day'::interval);
It will create 365 partitions and move the data from parent to partitions.
New partitions are appended automaticaly by insert trigger, but it can be done manually with the following functions:
-- append new partition with specified range
SELECT add_range_partition('journal', '2016-01-01'::date, '2016-01-07'::date);
-- append new partition with default range
SELECT append_range_partition('journal');
The first one creates a partition with specified range. The second one creates a partition with default interval and appends it to the partition list. It is also possible to attach an existing table as partition. For example, we may want to attach an archive table (or even foreign table from another server) for some outdated data:
CREATE FOREIGN TABLE journal_archive (
id INTEGER NOT NULL,
dt TIMESTAMP NOT NULL,
level INTEGER,
msg TEXT
) SERVER archive_server;
SELECT attach_range_partition('journal', 'journal_archive', '2014-01-01'::date, '2015-01-01'::date);
Important: the definition of the attached table must match the one of the existing partitioned table, including the dropped columns.
To merge to adjacent partitions, use the merge_range_partitions() function:
SELECT merge_range_partitions('journal_archive', 'journal_1');
To split partition by value, use the split_range_partition() function:
SELECT split_range_partition('journal_366', '2016-01-03'::date);
To detach partition, use the detach_range_partition() function:
SELECT detach_range_partition('journal_archive');
Here's an example of the query performing filtering by partitioning key:
SELECT * FROM journal WHERE dt >= '2015-06-01' AND dt < '2015-06-03';
id | dt | level | msg
--------+---------------------+-------+----------------------------------
217441 | 2015-06-01 00:00:00 | 2 | 15053892d993ce19f580a128f87e3dbf
217442 | 2015-06-01 00:01:00 | 1 | 3a7c46f18a952d62ce5418ac2056010c
217443 | 2015-06-01 00:02:00 | 0 | 92c8de8f82faf0b139a3d99f2792311d
...
(2880 rows)
EXPLAIN SELECT * FROM journal WHERE dt >= '2015-06-01' AND dt < '2015-06-03';
QUERY PLAN
------------------------------------------------------------------
Append (cost=0.00..58.80 rows=0 width=0)
-> Seq Scan on journal_152 (cost=0.00..29.40 rows=0 width=0)
-> Seq Scan on journal_153 (cost=0.00..29.40 rows=0 width=0)
(3 rows)
There are several user-accessible GUC variables designed to toggle the whole module or specific custom nodes on and off:
pg_pathman.enable--- disable (or enable)pg_pathmancompletelypg_pathman.enable_runtimeappend--- toggleRuntimeAppendcustom node on\offpg_pathman.enable_runtimemergeappend--- toggleRuntimeMergeAppendcustom node on\off
To permanently disable pg_pathman for some previously partitioned table, use the disable_partitioning() function:
SELECT disable_partitioning('range_rel');
All sections and data will remain unchanged and will be handled by the standard PostgreSQL inheritance mechanism.
##Feedback Do not hesitate to post your issues, questions and new ideas at the issues page.
Ildar Musin [email protected] Postgres Professional Ltd., Russia
Alexander Korotkov [email protected] Postgres Professional Ltd., Russia
Dmitry Ivanov [email protected] Postgres Professional Ltd., Russia