Postgres partitioning as easy as pie. Works great for both new and existing tables, with zero downtime and minimal app changes.
🍊 Battle-tested at Instacart
Run:
gem install pgslice
-
Ensure the table you want to partition has been created. We’ll refer to this as
<table>
. -
Specify your database credentials
export PGSLICE_URL=postgres:https://localhost/myapp_development
- Create an intermediate table
pgslice prep <table> <column> <period>
Period can be day
or month
.
This creates a table named <table>_intermediate
with the appropriate trigger for partitioning.
- Add partitions
pgslice add_partitions <table> --intermediate --past 3 --future 3
This creates child tables that inherit from the intermediate table.
Use the --past
and --future
options to control the number of partitions.
- Optional, for tables with data - Fill the partitions in batches with data from the original table
pgslice fill <table>
Use the --batch-size
and --sleep
options to control the speed.
To sync data across different databases, check out pgsync.
- Swap the intermediate table with the original table
pgslice swap <table>
The original table is renamed <table>_retired
and the intermediate table is renamed <table>
.
- Fill the rest (rows inserted between the first fill and the swap)
pgslice fill <table> --swapped
- Archive and drop the original table
To add partitions, use:
pgslice add_partitions <table> --future 3
Add this as a cron job to create a new partition each day or month.
# day
0 0 * * * pgslice add_partitions <table> --future 3 --url ...
# month
0 0 1 * * pgslice add_partitions <table> --future 3 --url ...
Add a monitor to ensure partitions are being created.
SELECT 1 FROM
pg_catalog.pg_class c
INNER JOIN
pg_catalog.pg_namespace n ON n.oid = c.relnamespace
WHERE
c.relkind = 'r' AND
n.nspname = 'public' AND
c.relname = '<table>_' || to_char(NOW() + INTERVAL '3 days', 'YYYYMMDD')
-- for months, use to_char(NOW() + INTERVAL '3 months', 'YYYYMM')
To undo prep (which will delete partitions), use:
pgslice unprep <table>
To undo swap, use:
pgslice unswap <table>
pgslice prints the SQL commands that were executed on the server. To print without executing, use the --dry-run
option.
$ pgslice prep locations created_at day
BEGIN;
CREATE TABLE locations_intermediate (LIKE locations INCLUDING ALL);
CREATE FUNCTION locations_insert_trigger()
RETURNS trigger AS $$
BEGIN
EXECUTE 'INSERT INTO locations_' || to_char(NEW.created_at, 'YYYYMMDD') || ' VALUES ($1.*)' USING NEW;
RETURN NULL;
END;
$$ LANGUAGE plpgsql;
CREATE TRIGGER locations_insert_trigger
BEFORE INSERT ON locations_intermediate
FOR EACH ROW EXECUTE PROCEDURE locations_insert_trigger();
COMMIT;
$ pgslice add_partitions locations --intermediate --past 1 --future 1
BEGIN;
CREATE TABLE locations_20160423
(CHECK (created_at >= '2016-04-23'::date AND created_at < '2016-04-24'::date))
INHERITS (locations_intermediate);
ALTER TABLE locations_20160423 ADD PRIMARY KEY (id);
CREATE INDEX ON locations_20160423 USING btree (shopper_id);
CREATE TABLE locations_20160424
(CHECK (created_at >= '2016-04-24'::date AND created_at < '2016-04-25'::date))
INHERITS (locations_intermediate);
ALTER TABLE locations_20160424 ADD PRIMARY KEY (id);
CREATE INDEX ON locations_20160424 USING btree (shopper_id);
CREATE TABLE locations_20160425
(CHECK (created_at >= '2016-04-25'::date AND created_at < '2016-04-26'::date))
INHERITS (locations_intermediate);
ALTER TABLE locations_20160425 ADD PRIMARY KEY (id);
CREATE INDEX ON locations_20160425 USING btree (shopper_id);
COMMIT;
$ pgslice fill locations
/* 1 of 3 */
INSERT INTO locations_intermediate (id, latitude, longitude, created_at)
SELECT id, latitude, longitude, created_at FROM locations
WHERE id > 0 AND id <= 10000 AND created_at >= '2016-04-23'::date AND created_at < '2016-04-26'::date
/* 2 of 3 */
INSERT INTO locations_intermediate (id, latitude, longitude, created_at)
SELECT id, latitude, longitude, created_at FROM locations
WHERE id > 10000 AND id <= 20000 AND created_at >= '2016-04-23'::date AND created_at < '2016-04-26'::date
/* 3 of 3 */
INSERT INTO locations_intermediate (id, latitude, longitude, created_at)
SELECT id, latitude, longitude, created_at FROM locations
WHERE id > 20000 AND id <= 30000 AND created_at >= '2016-04-23'::date AND created_at < '2016-04-26'::date
$ pgslice swap locations
BEGIN;
ALTER TABLE locations RENAME TO locations_retired;
ALTER TABLE locations_intermediate RENAME TO locations;
ALTER SEQUENCE locations_id_seq OWNED BY locations.id;
COMMIT;
$ pgslice add_partitions locations --future 2
BEGIN;
CREATE TABLE locations_20160426
(CHECK (created_at >= '2016-04-26'::date AND created_at < '2016-04-27'::date))
INHERITS (locations);
ALTER TABLE locations_20160426 ADD PRIMARY KEY (id);
CREATE INDEX ON locations_20160426 USING btree (shopper_id);
COMMIT;
This set up allows you to read and write with the original table name with no knowledge it’s partitioned. However, there are a few things to be aware of.
When possible, queries should include the column you partition on to limit the number of partitions the database needs to check. For instance, if you partition on created_at
, try to include it in queries:
SELECT * FROM
some_table
WHERE
some_column = 123 AND
-- for performance only
created_at >= '2016-01-01' AND created_at < '2016-01-02'
For this to be effective, ensure constraint_exclusion
is set to partition
(default value) or on
.
SHOW constraint_exclusion;
If you use INSERT
statements with a RETURNING
clause (as frameworks like Rails do), you’ll no longer receive the id of the newly inserted record back. If you need this, you can either:
- Insert directly into the partition
- Get the value after the insert with
SELECT CURRVAL('sequence_name')
You can also use pgslice to reduce the size of a table without partitioning by creating a new table, filling it with a subset of records, and swapping it in.
pgslice prep <table> --no-partition
pgslice fill <table> --start 1000 # starting primary key
pgslice swap <table>
Run:
gem install pgslice
To use master, run:
gem install specific_install
gem specific_install ankane/pgslice
- Command to sync index changes with partitions
- Disable indexing for faster
fill
- ETA for
fill
Everyone is encouraged to help improve this project. Here are a few ways you can help:
- Report bugs
- Fix bugs and submit pull requests
- Write, clarify, or fix documentation
- Suggest or add new features