Beginner's Guide to Partitioning vs. Sharding in Postgres | Claire Giordano | PGConf EU 2023

This slide deck is from Claire Giordano's talk at PGConfEU 2023 in Prague, about partitioning and sharding in Postgres. Partitioning has come a long way since the Postgres 10 days, as has sharding via the Citus extension. If you need to scale your Postgres, your friends may recommend you look into partitioning and/or sharding. But what’s the difference between these two approaches?

This beginner’s guide is for those who want to start with the basics and advance from there. You’ll learn what partitioning and sharding are in Postgres; how they’re different; as well as why (and when) these capabilities can help improve query performance. You’ll learn why some people call partitioning invaluable and others call sharding a lifesaver.

But there are gotchas and they’re not for every use case. As David Rowley says, “Understanding why partitioning helps is key to making the decision about whether to partition, what to partition on, and the type and number of partitions.” Similarly, understanding why sharding helps is key to deciding whether sharding is right for you. The concept of data locality will be touched on as well. And for those who prefer to learn by example and not just theory, yes, there will be examples.

Transcript

1. CLAIRE GIORDANO @clairegiordano@hachyderm.io • @clairegiordano • @AzureDBPostgres • @citusdata Beginner's Guide to Partitioning vs. Sharding in Postgres
2. @clairegiordano @clairegiordano@hachyderm.io Engineer Dev Manager PM Marketer Writer Open-source champion Community Lead
3. 3 Quite a lot of Postgres work @ Microsoft
4. 4 Quite a lot of Postgres work @ Microsoft aka.ms/blog-pg-at-microsoft
5. Inspiration for this talk came from a #PGSQLPhriday community blog-fest back in August 2023 Hosted by Tomasz Gintowt
6. So I wrote a thing for PGSQL Phriday “Understanding Partitioning and Sharding in Postgres and Citus” aka.ms/blog-partitioning-sharding-claire
7. Which led to this talk!
8. 8 What is sharding? Partitioning + sharding together How partitioning & sharding are different Why partitioning helps query performance When partitioning helps query performance Why sharding helps query performance When sharding helps query performance This Beginner Guide… What is Postgres partitioning?
9. 9 What is sharding? Partitioning + sharding together How partitioning & sharding are different Why partitioning helps query performance When partitioning helps query performance Why sharding helps query performance When sharding helps query performance Part 1 What is Postgres partitioning?
10. What is Postgres table partitioning Let’s say this huge Redwood Tree is like a big Postgres table...
11. What is Postgres table partitioning What if we could split this tree trunk into many smaller logs, & only work with the ones we need?
12. • Splits large tables into many smaller tables (partitions) • Built-in, native feature in Postgres • Needs to be declared up-front when you first CREATE TABLE
13. So many improvements to Postgres partitioning in last 5-6 years Shout-out to @Brandur for this blog post... aka.ms/partitioning-2022-brandur
14. 14 What are the benefits of Postgres Partitioning? Can improve query performance Improves performance of bulk deletes /if DROP PARTITION Improves performance of autovacuum /in so many cases!! Enables you to optimize storage costs (cheap vs. expensive)
15. In Postgres, trying to remove old rows from a large, hot table is flitting with disaster. A long running query must iterate through and mark each one as dead, and even then nothing is reclaimed until an equally expensive vacuum runs through and frees space, and only when it's allowed to after rows are no longer visible to any other query in the system, “ — Brandur whether they're making use of the large table or not. Each row removal land in the WAL, resulting in significant amplification. But with partitions, deletion becomes a simple DROP TABLE. It executes instantly, and with negligible costs (partitioning has other benefits too). The trade-off is maintenance." From Brandur.org Title of post: “Partitioning in Postgres, 2022 Edition”
16. 16 Declarative Postgres partitioning has 3 partitioning methods PARTITION BY RANGE PARTITION BY LIST PARTITION BY HASH Time-series & IOT Countries, company divisions,… When partitioning by range & by list do not work...
17. 17 Tracking monthly concert revenue for Taylor Swift Step 1: Decide on partition method Step 2: Decide on partition key Step 3: Create a partitioned table Step 4: Create partitions
18. 18 Let’s create a partitioned table CREATE TABLE concert_revenue ( city_id int not null, sale_date date not null, merch_sales int, ticket_sales int ) PARTITION BY RANGE (sale_date);
19. 19 Let’s create our first partition for June 2023 CREATE TABLE concert_revenue_cy2023m06 PARTITION OF concert_revenue FOR VALUES FROM ('2023-06-01') TO ('2023-07-01'); TIP: lower bound is INCLUSIVE TIP: upper bound (2023-07-01) is NOT inclusive
20. 20 Let’s create 3 partitions for Jun / Jul / Aug 2023 CREATE TABLE concert_revenue_cy2023m06 PARTITION OF concert_revenue FOR VALUES FROM ('2023-06-01') TO ('2023-07-01'); CREATE TABLE concert_revenue_cy2023m07 PARTITION OF concert_revenue FOR VALUES FROM ('2023-07-01') TO ('2023-08-01'); CREATE TABLE concert_revenue_cy2023m08 PARTITION OF concert_revenue FOR VALUES FROM ('2023-08-01') TO ('2023-09-01');
21. 21 Later, you can delete older data—e.g. June 2023 DROP TABLE concert_revenue_cy2023m06;
22. Partitioning is not magic! ...You need to maintain these partitions pg_partman with pg_cron
23. 23 What is sharding? Partitioning + sharding together How partitioning & sharding are different Why partitioning helps query performance When partitioning helps query performance Why sharding helps query performance When sharding helps query performance Part 2 What is Postgres partitioning?
24. What is sharding? 24 Splitting Postgres tables into smaller tables (“shards”)
25. What is sharding? 25 Splitting Postgres tables into smaller tables (“shards”) Distributing shards horizontally across multiple nodes
26. “A rose by any other name is still a rose” Other terms people use for Sharding 26 Horizontal Sharding Scaling out Horizontally Distributed Postgres
27. Back to previous metaphor If you have one big ginormous tree trunk & you want to split it into smaller pieces & distribute it across multiple parts of forest
29. 29 3 common ways to shard Postgres 2 1 Partitions + postgres_fdw Manual sharding sometimes called “sharding at application layer” Create hash-partitioned Postgres table in which every partition is a foreign table using postgres_fdw 3 Citus Extension to Postgres Open source, & also a managed service
30. Bulk of sharding work done by... 2 1 3 DBA Citus Dev App Dev Partitions + postgres_fdw Manual sharding Citus
31. In today’s talk, we’ll focus on sharding with Citus extension to Postgres
32. 32 Diagram of sharding Postgres with Citus Postgres Single Node Distributed Citus cluster with shards (smaller tables) on each node in cluster
33. 33 Citus extension = LOT more than sharding aka.ms/blog-citus-technical-readme
34. Pic from yesterday’s Distributed PostgreSQL talk at PGConfEU (by Marco Slot)
35. 35 Benefit of sharding Postgres with Citus: >>>> Get more Performance & Scale than you can eek out on a single node Can improve query performance @ scale //due to all the memory, CPU, & disk from multi-node cluster Enables your application to grow & scale // not just now, but in future Improves performance of autovacuum //since it runs in parallel across all the nodes in the cluster
36. 36 How to distribute tables with Citus for row-based sharding SELECT create_distributed_table( 'table_name', ‘sharding_key'); N.B. “sharding key” is often called a “distribution column”
37. 37 New to Citus 12.0, a 2nd way to shard: schema-based sharding SELECT citus_schema_distribute( 'name');
38. 38 What is sharding? Partitioning + sharding together How partitioning & sharding are different Why partitioning helps query performance When partitioning helps query performance Why sharding helps query performance When sharding helps query performance Part 3 What is Postgres partitioning?
39. 39 Can you Partition & Shard together? Single PG node Partitions on a single Postgres node Sharded Postgres partitions on a distributed 4-node Citus cluster Node 1 Node 2 Node 3 Node 4
40. 40 These two Citus UDFs simplify partition management create_time_partitions(table_name regclass, partition_interval interval, end_at timestamp with time zone, start_from timestamp with time zone DEFAULT now()) #1—Create as many partitions as necessary for given time range
41. 41 These 2 Citus UDFs simplify partition management drop_old_time_partitions(table_name regclass, older_than timestamp with time zone) #2—Drop all partitions older than given timestamp
42. 42 What is sharding? Partitioning + sharding together How partitioning & sharding are different Why partitioning helps query performance When partitioning helps query performance Why sharding helps query performance When sharding helps query performance Part 4 What is Postgres partitioning?
43. Let’s look at a comparison table! How partitioning & sharding are different
44. Postgres native Partitioning Sharding with Citus extension Partitioning + Sharding Combination 2 1 3 0 Aspects or attributes of these technologies
45. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
46. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
47. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
48. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
49. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
50. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
51. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
52. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
53. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
54. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
55. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
56. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
57. Attribute Partitioning Citus Sharding Partition + Shard Combo Built into Postgres Extension to Postgres Single node Multi-node Drop old data quickly Parallel, distributed SQL/DDL/DML Partition/Shard Pruning Parallel autovacuum Better index cache hit ratios Automatic maintenance Time series apps (e.g. IOT) Multi-tenant SaaS apps
58. 58 What is sharding? Partitioning + sharding together How partitioning & sharding are different Why partitioning helps query performance When partitioning helps query performance Why sharding helps query performance When sharding helps query performance Part 5 What is Postgres partitioning?
59. Why partitioning can help improve query performance Partition Pruning
60. Why partitioning can help improve query performance Data Locality
61. Why partitioning can help improve query performance Partition Pruning Data Locality &
62. 62 What is sharding? Partitioning + sharding together How partitioning & sharding are different Why partitioning helps query performance When partitioning helps query performance Why sharding helps query performance When sharding helps query performance Part 6 What is Postgres partitioning?
63. When partitioning helps improve query performance Just like when you bake a cake, need essential ingredients to get the right result
64. When partitioning helps improve query performance Only when you use a WHERE clause that prunes (excludes) partitions from the query
65. 65 Remember that Taylor Swift concert example... Tracking monthly concert revenues... What if we wanted to know what the August revenue was from ticket sales + merch sales?
66. 66 This EXPLAIN query is missing WHERE clause EXPLAIN (costs off) SELECT count(*) FROM concert_revenue;
67. 67 So it has to scan all 3 of the partitions  EXPLAIN (costs off) SELECT count(*) FROM concert_revenue; QUERY PLAN --------------------------------------------------------------------- Aggregate -> Append -> Seq Scan on concert_revenue_cy2023m06 concert_revenue_1 -> Seq Scan on concert_revenue_cy2023m07 concert_revenue_2 -> Seq Scan on concert_revenue_cy2023m08 concert_revenue_3 (5 rows)
68. 68 And doesn’t benefit from partition pruning  EXPLAIN (costs off) SELECT count(*) FROM concert_revenue; QUERY PLAN --------------------------------------------------------------------- Aggregate -> Append -> Seq Scan on concert_revenue_cy2023m06 concert_revenue_1 -> Seq Scan on concert_revenue_cy2023m07 concert_revenue_2 -> Seq Scan on concert_revenue_cy2023m08 concert_revenue_3 (5 rows)
69. 69 This EXPLAIN query includes a WHERE clause EXPLAIN (costs off) SELECT count(*) FROM concert_revenue WHERE sale_date BETWEEN '2023-08-01' AND '2023-08-31';
70. 70 And so it only needs to scan the Aug partition EXPLAIN (costs off) SELECT count(*) FROM concert_revenue WHERE sale_date BETWEEN '2023-08-01' AND '2023-08-31’; QUERY PLAN --------------------------------------------------------------------- ---------------------- Aggregate -> Seq Scan on concert_revenue_cy2023m08 concert_revenue Filter: ((sale_date >= '2023-08-01'::date) AND (sale_date <= '2023-08-31'::date)) (3 rows)
71. 71 And so it only needs to scan the Aug partition EXPLAIN (costs off) SELECT count(*) FROM concert_revenue WHERE sale_date BETWEEN '2023-08-01' AND '2023-08-31’; QUERY PLAN --------------------------------------------------------------------- ---------------------- Aggregate -> Seq Scan on concert_revenue_cy2023m08 concert_revenue Filter: ((sale_date >= '2023-08-01'::date) AND (sale_date <= '2023-08-31'::date)) (3 rows)
72. 72 What is sharding? Partitioning + sharding together How partitioning & sharding are different Why partitioning helps query performance When partitioning helps query performance Why sharding helps query performance When sharding helps query performance Part 7 What is Postgres partitioning?
73. Why sharding can help query performance Let’s say this huge Redwood Tree is like a big Postgres table... More CPU, Memory, Disk
74. Why sharding can help query performance Let’s say this huge Redwood Tree is like a big Postgres table... More CPU, Memory, Disk
75. Why sharding can help query performance Let’s say this huge Redwood Tree is like a big Postgres table... Shard Pruning Logic
76. Why sharding can help query performance Co-Located Tables
77. Why sharding can help query performance Let’s say this huge Redwood Tree is like a big Postgres table... Shard Pruning Logic Co-Located Tables More CPU, Memory, Disk
78. 78 What is sharding? Partitioning + sharding together How partitioning & sharding are different Why partitioning helps query performance When partitioning helps query performance Why sharding helps query performance When sharding helps query performance Part 8 What is Postgres partitioning?
79. When sharding helps improve query performance Let’s say this huge Redwood Tree is like a big Postgres table... CPU Memory Disk CPU Memory Disk CPU Memory Disk CPU Memory Disk CPU Memory Disk CPU Memory Disk CPU Memory Disk + + + + + When your application needs more cpu, memory, or disk than is possible on a single Postgres node Common example: “Cache hit ratio” has been going down—leading to slower query response times
80. So what are the takeaways?
81. 81 Takeaways about PG Partitioning & Sharding 1. Partitioning & Sharding can be “Invaluable” and “Lifesavers” 3. Planning required 2. You don’t have to pick between Partitioning & Sharding 4. Keys Matter
82. 82 Takeaways about PG Partitioning & Sharding 1. Partitioning & Sharding can be “Invaluable” and “Lifesavers” 3. Planning required 2. You don’t have to pick between Partitioning & Sharding 4. Keys Matter
83. 83 Takeaways about PG Partitioning & Sharding 1. Partitioning & Sharding can be “Invaluable” and “Lifesavers” 3. Planning required 2. You don’t have to pick between Partitioning & Sharding 4. Keys Matter
84. 84 Takeaways about PG Partitioning & Sharding 1. Partitioning & Sharding can be “Invaluable” and “Lifesavers” 3. Planning required 2. You don’t have to pick between Partitioning & Sharding 4. Keys Matter
85. People to thank for inspiration &or reviews Daniel Gustafsson David Rowley Isaac Alves Jelte Fennema-Nio Marco Slot Melanie Plageman Robert Treat Ryan Booz Thomas Munro Umur Cubukcu Charles Feddersen
86. aka.ms/PathToCitusCon A podcast for developers who love Postgres
87. @clairegiordano @clairegiordano@hachyderm.io aka.ms/PathToCitusCon z