Practical Partitioning in Production with Postgres

Practical
Partitioning in
Production with
Postgres
Jimmy Angelakos
Senior PostgreSQL Architect
Postgres Vision 2021-06-23

© Copyright EnterpriseDB Corporation, 2021. All rights reserved.
2
We’ll be looking at:
• Intro to Partitioning in PostgreSQL
• Why?
• How?
• Practical Example

Introduction to
Partitioning in
PostgreSQL

4
• RDBMS context: division of a table into distinct independent tables
• Horizontal partitioning (by row) – different rows in different tables
• Why?
– Easier to manage
– Performance
What is partitioning?

5
• Has had partitioning for quite some time now PG 8.1 (2005)
…
– Inheritance-based
– Why haven’t I heard of this before?
– It’s not great tbh...
• Declarative Partitioning: PG 10 (2017)
– Massive improvement
Partitioning in PostgreSQL
HISTORY

6
CREATE TABLE cust (id INT, signup DATE)
PARTITION BY RANGE (signup);
CREATE TABLE cust_2020
PARTITION OF cust FOR VALUES FROM
('2020-01-01') TO ('2021-01-01');
• Partitions may be partitioned
themselves (sub-partitioning)
Declarative Partitioning
( PG 10+ )
Specification of: By declaring a table (DDL):
• Partitioning method
• Partition key
– Column(s) or expression(s)
– Value determines data routing
• Partition boundaries

8
• Database size: unlimited ✅
• Tables per database: 1.4 billion ✅
• Table size: 32 TB 😐
– Default block size: 8192 bytes
• Rows per table: depends
– As many as can fit onto 4.2 billion blocks 😐
PostgreSQL limits
(Hard limits, hard to reach)

9
• Disk size limitations
– You can put partitions on different tablespaces
• Performance
– Partition pruning
– Table scans
– Index scans
– Hidden pitfalls of very large tables*
What partitioning can help with (i)
(Very large tables)

10
• Maintenance
– Deletions (some filesystems are bad at deleting large numbers of files)
🤭
– DROP TABLE cust_2020;
– ALTER TABLE cust DETACH PARTITION cust_2020;
• VACUUM
– Bloat
– Freezing → xid wraparound
What partitioning can help with (ii)
(Very large tables)

11
• Magic bullet
– No substitute for rational database design
• Sharding
– Not about putting part of the data on different nodes
• Performance tuning
– Unless you have one of the mentioned issues
What partitioning is not

13
• Get your calculator out
– Data ingestion rate (both rows and size in bytes)
– Projected increases (e.g. 25 locations projected to be 200 by end of year)
– Data retention requirements
• Will inform choice of partitioning method and key
• For instance: 1440 measurements/day from each of 1000 sensors – extrapolate per year
• Keep checking if this is valid and be prepared to revise
Dimensioning
Plan ahead!

14
• Range: For key column(s) e.g. ranges of dates, identifiers, etc.
– Lower end: inclusive, upper end: exclusive
• List: Explicit key values stated for each partition
• Hash (PG 11+): If you have a column with values close to unique
– Define Modulus ( & remainder ) for number of almost-evenly-sized partitions
Partitioning method
Dimensioning usually makes this clearer

15
• Analysis
– Determine main keys used for retrieval from queries
– Proper key selection enables partition pruning
– Can use multiple columns for higher granularity (more partitions)
• Desirable
– High enough cardinality (range of values) for the number of partitions needed
– A column that doesn’t change often, to avoid moving rows among partitions
Partition Key selection
Choose wisely - know your data!

16
• Simply put, partitions are partitioned tables themselves. Plan ahead!
• CREATE TABLE transactions ( , location_code
… TEXT, tstamp TIMESTAMPTZ)
PARTITION BY RANGE (tstamp);
• CREATE TABLE transactions_2021_06
PARTITION OF transactions FOR VALUES FROM ('2021-06-01') TO ('2021-07-01')
PARTITION BY HASH (location_code);
• CREATE TABLE transactions_2021_06_p1
PARTITION OF transactions_2021_06 FOR VALUES WITH (MODULUS 4, REMAINDER 0);
Sub-partitioning

17
Partitioning by multiple columns
• CREATE TABLE transactions ( , location_code
… TEXT, tstamp TIMESTAMPTZ)
PARTITION BY RANGE (tstamp, location_code);
• CREATE TABLE transactions_2021_06_a PARTITION OF transactions
FOR VALUES FROM ('2021-06-01', 'AAA') TO ('2021-07-01', 'AZZ');
• CREATE TABLE transactions_2021_06_b PARTITION OF transactions
FOR VALUES FROM ('2021-06-01', 'BAA') TO ('2021-07-01', 'BZZ');
ERROR: partition "transactions_2021_06_b" would overlap partition
"transactions_2021_06_a"
• Because tstamp '2021-06-01' can only go in the first partition!
Be careful!

18
• Automatic creation of partitions
– Create in advance
– Use a cronjob
• Imperative merging/splitting of partitions
– Move rows manually
• Sharding to different nodes
– You may have to configure FDW manually
What Postgres does not do
core

20
• Is your table too large to handle?
• Can partitioning help?
• What if it’s in constant use?
Partitioning a live production system

21
• OLTP workload, transactions keep flowing in
– Table keeps increasing in size
• VACUUM never ends
– Has been running for a full month already…
• Queries are getting slower
– Not just because of sheer number of rows...
The situation
Huge 20 TB table

22
• Postgres has 1GB segment size
– Can only be changed at
compilation time
– 20 TB table = 20000 segments
(files on disk)
• Why is this a problem?
– md.c →
* Hidden performance pitfall (i)
For VERY large tables

23
●
This loops 20000 times every time you
want to access a table page
– Linked list of segments
●
Code from PG 9.6
●
It has been heavily optimised recently
(caching, etc).
●
Still needs to run a lot of times
* Hidden performance pitfall (ii)

24
• Need to partition the huge table
– Dimensioning
– Partition method
– Partition key
• Make sure we’re on the latest version (PG 13)
– Get latest features & performance enhancements
So what do we do?
Next steps

25
• Dimensioning
– One partition per month will be about 30GB of data, so acceptable size
• Method, Key
– Candidate key is transaction date, which we can partition by range
– Check that there are no data errors (e.g. dates in the future when they shouldn’t be)
• Partition sizes don’t have to be equal
– We can partition older, less often accessed data by year
What is our table like?
It holds daily transaction totals for each point of sales

26
• Lock the table totally (ACCESS EXCLUSIVE) or prevent writes
– People will start yelling, and they will be right
• Cause excessive load on the system (e.g. I/O) or cause excessive disk space usage
– Can’t copy whole 20 TB table into empty partitioned table
– See above about yelling
• Present an inconsistent or incomplete view of the data
Problems
What things you cannot do in production

27
• Rename the huge table and its indices
• Create an empty partitioned table with the old huge table’s name
• Create the required indices on the new partitioned table
– They will be created automatically for each new partition
• Create first new partition for new incoming data
• Attach the old table as a partition of the new table so it can be used normally*
• Move data out of the old table incrementally at our own pace
The plan
Take it step by step

28
-- Do this all in one transaction
BEGIN;
ALTER TABLE dailytotals RENAME TO dailytotals_legacy;
ALTER INDEX dailytotals_batchid RENAME TO dailytotals_legacy_batchid;
ALTER INDEX …
…
Rename the huge table and its indices

29
CREATE TABLE dailytotals (
totalid BIGINT NOT NULL DEFAULT nextval('dailytotals_totalid_seq')
, totaldate DATE NOT NULL
, totalsum BIGINT
…
, batchid BIGINT NOT NULL
)
PARTITION BY RANGE (totaldate);
CREATE INDEX dailytotals_batchid ON dailytotals (batchid);
…
Create empty partitioned table & indices

30
CREATE TABLE dailytotals_202106
PARTITION OF dailytotals
FOR VALUES FROM ('2021-06-01') TO ('2021-07-01');
Create partition for new incoming data

31
DO $$
DECLARE earliest DATE;
DECLARE latest DATE;
BEGIN
-- Set boundaries
SELECT min(totaldate) INTO earliest FROM dailytotals_legacy;
latest := '2021-06-01'::DATE;
Attach old table as a partition (i)

32
-- HACK HACK HACK (only because we know and trust our data)
ALTER TABLE dailytotals_legacy
ADD CONSTRAINT dailytotals_legacy_totaldate
CHECK (totaldate >= earliest AND totaldate < latest)
NOT VALID;
-- You should not touch pg_catalog directly 😕
UPDATE pg_constraint
SET convalidated = true
WHERE conname = 'dailytotals_legacy_totaldate';
Attach old table as a partition (ii)

33
ALTER TABLE dailytotals
ATTACH PARTITION dailytotals_legacy
FOR VALUES FROM (earliest) TO (latest);
END;
$$ LANGUAGE PLPGSQL;
COMMIT;
Attach old table as a partition (iii)

34
• For instance, during quiet hours for the system, in scheduled batch jobs, etc.
WITH rows AS (
DELETE FROM dailytotals_legacy d
WHERE (totaldate >= '2020-01-01' AND totaldate < '2021-01-01')
RETURNING d.* )
INSERT INTO dailytotals SELECT * FROM rows;
• In the same transaction: DETACH the old table, perform the move, reATTACH with changed
boundaries. Rinse and repeat!
• Make sure the target partition exists!
Move data from old table at our own pace

35
• PG11: DEFAULT partition, UPDATE on partition key, HASH method, PKs, FKs, Indexes, Triggers
• PG12: Performance (pruning, COPY), FK references for partitioned tables, ordered scans
• PG13: Logical replication for partitioned tables, improved performance (JOINs, pruning)
• (Soon) PG14: REINDEX CONCURRENTLY, DETACH CONCURRENTLY, faster UPDATE/DELETE
Partitioning improvements
Make sure you’re on the latest release so you have them!

36
• Know your data!
• Upgrade – be on the latest release!
• Partition before you get in deep water!
• Find me on Twitter: @vyruss
To conclude...

Practical Partitioning in Production with Postgres

More Related Content

What's hot (20)

Similar to Practical Partitioning in Production with Postgres (20)

More from Jimmy Angelakos (9)

Recently uploaded (20)

Practical Partitioning in Production with Postgres