5.10. Table Partitioning

5.10.1. Overview

Partitioning refers to splitting what is logically one large table into smaller physical pieces. Partitioning can provide several benefits:

• Query performance can be improved dramatically in certain situations, particularly when most of the heavily accessed rows of the table are in a single partition or a small number of partitions. The partitioning substitutes for leading columns of indexes, reducing index size and making it more likely that the heavily-used parts of the indexes fit in memory.

• When queries or updates access a large percentage of a single partition, performance can be improved by taking advantage of sequential scan of that partition instead of using an index and random access reads scattered across the whole table.

• Bulk loads and deletes can be accomplished by adding or removing partitions, if that requirement is planned into the partitioning design. Doing ALTER TABLE DETACH PARTITION or dropping an individual partition using DROP TABLE is far faster than a bulk operation. These commands also entirely avoid the VACUUM overhead caused by a bulk DELETE.

• Seldom-used data can be migrated to cheaper and slower storage media.

The benefits will normally be worthwhile only when a table would otherwise be very large. The exact point at which a table will benefit from partitioning depends on the application, although a rule of thumb is that the size of the table should exceed the physical memory of the database server.

PostgreSQL offers built-in support for the following forms of partitioning:

If your application needs to use other forms of partitioning not listed above, alternative methods such as inheritance and UNION ALL views can be used instead. Such methods offer flexibility but do not have some of the performance benefits of built-in declarative partitioning.

5.10.2. Declarative Partitioning

PostgreSQL offers a way to specify how to divide a table into pieces called partitions. The table that is divided is referred to as a partitioned table. The specification consists of the partitioning method and a list of columns or expressions to be used as the partition key.

All rows inserted into a partitioned table will be routed to one of the partitions based on the value of the partition key. Each partition has a subset of the data defined by its partition bounds. The currently supported partitioning methods are range, list, and hash.

Partitions may themselves be defined as partitioned tables, using what is called sub-partitioning. Partitions may have their own indexes, constraints and default values, distinct from those of other partitions. See CREATE TABLE for more details on creating partitioned tables and partitions.

It is not possible to turn a regular table into a partitioned table or vice versa. However, it is possible to add a regular or partitioned table containing data as a partition of a partitioned table, or remove a partition from a partitioned table turning it into a standalone table; see ALTER TABLE to learn more about the ATTACH PARTITION and DETACH PARTITION sub-commands.

Individual partitions are linked to the partitioned table with inheritance behind-the-scenes; however, it is not possible to use some of the generic features of inheritance (discussed below) with declaratively partitioned tables or their partitions. For example, a partition cannot have any parents other than the partitioned table it is a partition of, nor can a regular table inherit from a partitioned table making the latter its parent. That means partitioned tables and their partitions do not participate in inheritance with regular tables. Since a partition hierarchy consisting of the partitioned table and its partitions is still an inheritance hierarchy, all the normal rules of inheritance apply as described in Section 5.9 with some exceptions, most notably:

• Both CHECK and NOT NULL constraints of a partitioned table are always inherited by all its partitions. CHECK constraints that are marked NO INHERIT are not allowed to be created on partitioned tables.

• Using ONLY to add or drop a constraint on only the partitioned table is supported as long as there are no partitions. Once partitions exist, using ONLY will result in an error as adding or dropping constraints on only the partitioned table, when partitions exist, is not supported. Instead, constraints on the partitions themselves can be added and (if they are not present in the parent table) dropped.

• As a partitioned table does not have any data directly, attempts to use TRUNCATE ONLY on a partitioned table will always return an error.

• Partitions cannot have columns that are not present in the parent. It is not possible to specify columns when creating partitions with CREATE TABLE, nor is it possible to add columns to partitions after-the-fact using ALTER TABLE. Tables may be added as a partition with ALTER TABLE ... ATTACH PARTITION only if their columns exactly match the parent, including any oid column.

• You cannot drop the NOT NULL constraint on a partition's column if the constraint is present in the parent table.

Partitions can also be foreign tables, although they have some limitations that normal tables do not; see CREATE FOREIGN TABLE for more information.

Updating the partition key of a row might cause it to be moved into a different partition where this row satisfies the partition bounds.

CREATE TABLE measurement (
    city_id         int not null,
    logdate         date not null,
    peaktemp        int,
    unitsales       int
);

DROP TABLE measurement_y2006m02;

ALTER TABLE measurement DETACH PARTITION measurement_y2006m02;

CREATE TABLE measurement_y2008m02 PARTITION OF measurement
    FOR VALUES FROM ('2008-02-01') TO ('2008-03-01')
    TABLESPACE fasttablespace;

CREATE TABLE measurement_y2008m02
  (LIKE measurement INCLUDING DEFAULTS INCLUDING CONSTRAINTS)
  TABLESPACE fasttablespace;

ALTER TABLE measurement_y2008m02 ADD CONSTRAINT y2008m02
   CHECK ( logdate >= DATE '2008-02-01' AND logdate < DATE '2008-03-01' );

\copy measurement_y2008m02 from 'measurement_y2008m02'
-- possibly some other data preparation work

ALTER TABLE measurement ATTACH PARTITION measurement_y2008m02
    FOR VALUES FROM ('2008-02-01') TO ('2008-03-01' );

CREATE INDEX measurement_usls_idx ON ONLY measurement (unitsales);

CREATE INDEX measurement_usls_200602_idx
    ON measurement_y2006m02 (unitsales);
ALTER INDEX measurement_usls_idx
    ATTACH PARTITION measurement_usls_200602_idx;
...

ALTER TABLE ONLY measurement ADD UNIQUE (city_id, logdate);

ALTER TABLE measurement_y2006m02 ADD UNIQUE (city_id, logdate);
ALTER INDEX measurement_city_id_logdate_key
    ATTACH PARTITION measurement_y2006m02_city_id_logdate_key;
...

5.10.3. Implementation Using Inheritance

While the built-in declarative partitioning is suitable for most common use cases, there are some circumstances where a more flexible approach may be useful. Partitioning can be implemented using table inheritance, which allows for several features not supported by declarative partitioning, such as:

• For declarative partitioning, partitions must have exactly the same set of columns as the partitioned table, whereas with table inheritance, child tables may have extra columns not present in the parent.

• Table inheritance allows for multiple inheritance.

• Declarative partitioning only supports range, list and hash partitioning, whereas table inheritance allows data to be divided in a manner of the user's choosing. (Note, however, that if constraint exclusion is unable to prune child tables effectively, query performance might be poor.)

• Some operations require a stronger lock when using declarative partitioning than when using table inheritance. For example, adding or removing a partition to or from a partitioned table requires taking an ACCESS EXCLUSIVE lock on the parent table, whereas a SHARE UPDATE EXCLUSIVE lock is enough in the case of regular inheritance.

DROP TABLE measurement_y2006m02;

ALTER TABLE measurement_y2006m02 NO INHERIT measurement;

CREATE TABLE measurement_y2008m02 (
    CHECK ( logdate >= DATE '2008-02-01' AND logdate < DATE '2008-03-01' )
) INHERITS (measurement);

CREATE TABLE measurement_y2008m02
  (LIKE measurement INCLUDING DEFAULTS INCLUDING CONSTRAINTS);
ALTER TABLE measurement_y2008m02 ADD CONSTRAINT y2008m02
   CHECK ( logdate >= DATE '2008-02-01' AND logdate < DATE '2008-03-01' );
\copy measurement_y2008m02 from 'measurement_y2008m02'
-- possibly some other data preparation work
ALTER TABLE measurement_y2008m02 INHERIT measurement;

5.10.4. Partition Pruning

Partition pruning is a query optimization technique that improves performance for declaratively partitioned tables. As an example:

SET enable_partition_pruning = on;                 -- the default
SELECT count(*) FROM measurement WHERE logdate >= DATE '2008-01-01';

Without partition pruning, the above query would scan each of the partitions of the measurement table. With partition pruning enabled, the planner will examine the definition of each partition and prove that the partition need not be scanned because it could not contain any rows meeting the query's WHERE clause. When the planner can prove this, it excludes (prunes) the partition from the query plan.

By using the EXPLAIN command and the enable_partition_pruning configuration parameter, it's possible to show the difference between a plan for which partitions have been pruned and one for which they have not. A typical unoptimized plan for this type of table setup is:

SET enable_partition_pruning = off;
EXPLAIN SELECT count(*) FROM measurement WHERE logdate >= DATE '2008-01-01';
                                    QUERY PLAN
-----------------------------------------------------------------------------------
 Aggregate  (cost=188.76..188.77 rows=1 width=8)
   ->  Append  (cost=0.00..181.05 rows=3085 width=0)
         ->  Seq Scan on measurement_y2006m02  (cost=0.00..33.12 rows=617 width=0)
               Filter: (logdate >= '2008-01-01'::date)
         ->  Seq Scan on measurement_y2006m03  (cost=0.00..33.12 rows=617 width=0)
               Filter: (logdate >= '2008-01-01'::date)
...
         ->  Seq Scan on measurement_y2007m11  (cost=0.00..33.12 rows=617 width=0)
               Filter: (logdate >= '2008-01-01'::date)
         ->  Seq Scan on measurement_y2007m12  (cost=0.00..33.12 rows=617 width=0)
               Filter: (logdate >= '2008-01-01'::date)
         ->  Seq Scan on measurement_y2008m01  (cost=0.00..33.12 rows=617 width=0)
               Filter: (logdate >= '2008-01-01'::date)

Some or all of the partitions might use index scans instead of full-table sequential scans, but the point here is that there is no need to scan the older partitions at all to answer this query. When we enable partition pruning, we get a significantly cheaper plan that will deliver the same answer:

SET enable_partition_pruning = on;
EXPLAIN SELECT count(*) FROM measurement WHERE logdate >= DATE '2008-01-01';
                                    QUERY PLAN
-----------------------------------------------------------------------------------
 Aggregate  (cost=37.75..37.76 rows=1 width=8)
   ->  Append  (cost=0.00..36.21 rows=617 width=0)
         ->  Seq Scan on measurement_y2008m01  (cost=0.00..33.12 rows=617 width=0)
               Filter: (logdate >= '2008-01-01'::date)

Note that partition pruning is driven only by the constraints defined implicitly by the partition keys, not by the presence of indexes. Therefore it isn't necessary to define indexes on the key columns. Whether an index needs to be created for a given partition depends on whether you expect that queries that scan the partition will generally scan a large part of the partition or just a small part. An index will be helpful in the latter case but not the former.

Partition pruning can be performed not only during the planning of a given query, but also during its execution. This is useful as it can allow more partitions to be pruned when clauses contain expressions whose values are not known at query planning time; for example, parameters defined in a PREPARE statement, using a value obtained from a subquery or using a parameterized value on the inner side of a nested loop join. Partition pruning during execution can be performed at any of the following times:

• During initialization of the query plan. Partition pruning can be performed here for parameter values which are known during the initialization phase of execution. Partitions which are pruned during this stage will not show up in the query's EXPLAIN or EXPLAIN ANALYZE. It is possible to determine the number of partitions which were removed during this phase by observing the “Subplans Removed” property in the EXPLAIN output.

• During actual execution of the query plan. Partition pruning may also be performed here to remove partitions using values which are only known during actual query execution. This includes values from subqueries and values from execution-time parameters such as those from parameterized nested loop joins. Since the value of these parameters may change many times during the execution of the query, partition pruning is performed whenever one of the execution parameters being used by partition pruning changes. Determining if partitions were pruned during this phase requires careful inspection of the loops property in the EXPLAIN ANALYZE output. Subplans corresponding to different partitions may have different values for it depending on how many times each of them was pruned during execution. Some may be shown as (never executed) if they were pruned every time.

Partition pruning can be disabled using the enable_partition_pruning setting.

Note

Execution-time partition pruning currently only occurs for the Append node type, not for MergeAppend or ModifyTable nodes. That is likely to be changed in a future release of PostgreSQL.

5.10.5. Partitioning and Constraint Exclusion

Constraint exclusion is a query optimization technique similar to partition pruning. While it is primarily used for partitioning implemented using the legacy inheritance method, it can be used for other purposes, including with declarative partitioning.

Constraint exclusion works in a very similar way to partition pruning, except that it uses each table's CHECK constraints — which gives it its name — whereas partition pruning uses the table's partition bounds, which exist only in the case of declarative partitioning. Another difference is that constraint exclusion is only applied at plan time; there is no attempt to remove partitions at execution time.

The fact that constraint exclusion uses CHECK constraints, which makes it slow compared to partition pruning, can sometimes be used as an advantage: because constraints can be defined even on declaratively-partitioned tables, in addition to their internal partition bounds, constraint exclusion may be able to elide additional partitions from the query plan.

The default (and recommended) setting of constraint_exclusion is neither on nor off, but an intermediate setting called partition, which causes the technique to be applied only to queries that are likely to be working on inheritance partitioned tables. The on setting causes the planner to examine CHECK constraints in all queries, even simple ones that are unlikely to benefit.

The following caveats apply to constraint exclusion:

• Constraint exclusion is only applied during query planning, unlike partition pruning, which can also be applied during query execution.

• Constraint exclusion only works when the query's WHERE clause contains constants (or externally supplied parameters). For example, a comparison against a non-immutable function such as CURRENT_TIMESTAMP cannot be optimized, since the planner cannot know which child table the function's value might fall into at run time.

• Keep the partitioning constraints simple, else the planner may not be able to prove that child tables might not need to be visited. Use simple equality conditions for list partitioning, or simple range tests for range partitioning, as illustrated in the preceding examples. A good rule of thumb is that partitioning constraints should contain only comparisons of the partitioning column(s) to constants using B-tree-indexable operators, because only B-tree-indexable column(s) are allowed in the partition key.

• All constraints on all children of the parent table are examined during constraint exclusion, so large numbers of children are likely to increase query planning time considerably. So the legacy inheritance based partitioning will work well with up to perhaps a hundred child tables; don't try to use many thousands of children.

5.10.6. Declarative Partitioning Best Practices

The choice of how to partition a table should be made carefully as the performance of query planning and execution can be negatively affected by poor design.

One of the most critical design decisions will be the column or columns by which you partition your data. Often the best choice will be to partition by the column or set of columns which most commonly appear in WHERE clauses of queries being executed on the partitioned table. WHERE clause items that match and are compatible with the partition key can be used to prune unneeded partitions. However, you may be forced into making other decisions by requirements for the PRIMARY KEY or a UNIQUE constraint. Removal of unwanted data is also a factor to consider when planning your partitioning strategy. An entire partition can be detached fairly quickly, so it may be beneficial to design the partition strategy in such a way that all data to be removed at once is located in a single partition.

Choosing the target number of partitions that the table should be divided into is also a critical decision to make. Not having enough partitions may mean that indexes remain too large and that data locality remains poor which could result in low cache hit ratios. However, dividing the table into too many partitions can also cause issues. Too many partitions can mean longer query planning times and higher memory consumption during both query planning and execution. When choosing how to partition your table, it's also important to consider what changes may occur in the future. For example, if you choose to have one partition per customer and you currently have a small number of large customers, consider the implications if in several years you instead find yourself with a large number of small customers. In this case, it may be better to choose to partition by HASH and choose a reasonable number of partitions rather than trying to partition by LIST and hoping that the number of customers does not increase beyond what it is practical to partition the data by.

Sub-partitioning can be useful to further divide partitions that are expected to become larger than other partitions, although excessive sub-partitioning can easily lead to large numbers of partitions and can cause the same problems mentioned in the preceding paragraph.

It is also important to consider the overhead of partitioning during query planning and execution. The query planner is generally able to handle partition hierarchies with up to a few hundred partitions fairly well, provided that typical queries allow the query planner to prune all but a small number of partitions. Planning times become longer and memory consumption becomes higher as more partitions are added. This is particularly true for the UPDATE and DELETE commands. Another reason to be concerned about having a large number of partitions is that the server's memory consumption may grow significantly over a period of time, especially if many sessions touch large numbers of partitions. That's because each partition requires its metadata to be loaded into the local memory of each session that touches it.

With data warehouse type workloads, it can make sense to use a larger number of partitions than with an OLTP type workload. Generally, in data warehouses, query planning time is less of a concern as the majority of processing time is spent during query execution. With either of these two types of workload, it is important to make the right decisions early, as re-partitioning large quantities of data can be painfully slow. Simulations of the intended workload are often beneficial for optimizing the partitioning strategy. Never assume that more partitions are better than fewer partitions and vice-versa.