The Challenges of Distributing Postgres: A Citus Story

Set theory forms the basis for relational algebra and relational databases, and SQL is the lingua franca of modern RDBMS’s. Even with all the attention given to NoSQL in recent years, the lion share of database usage remains relational. But until recently, nearly all relational database solutions have been limited to the resources of a single node. Not anymore.

This talk is about my team’s journey tackling the challenges of distributing SQL. Specifically in the context of my favorite (open source) database: Postgres. I believe that too many developers spend too much time worrying about scaling their databases. So at Citus Data, we created an extension to Postgres that enables developers to scale out compute, memory, and storage by distributing queries across a cluster of nodes.

This talk describes the distributed systems challenges we faced at Citus in scaling out Postgres—and how we addressed them. I’ll talk about how we use PostgreSQL’s extension APIs to parallelize queries in a distributed cluster. I’ll cover the architecture of a distributed query planner and specifically how the join order planner has to choose between broadcast, co-located, and repartition joins in order to minimize network I/O. And if there’s time, I’ll walk through the dynamic executor logic that we built. The end result: a distributed database and a lot less time spent worrying about scale.

Transcript

1. The Challenges of Distributing Postgres: A Citus Story Ozgun Erdogan DataEngConf NYC | October 2017
2. Developers Love Postgres PostgreSQL MySQL MongoDB SQL Server + Oracle RDBMS: PostgreSQL, MySQL, Microsoft SQL Server, Oracle Ozgun Erdogan | DataEngConf NYC 2017
3. I love Postgres, too 3 Ozgun Erdogan | DataEngConf NYC 2017 Ozgun Erdogan CTO of Citus Data Distributed Systems Distributed Databases Formerly of Amazon Love drinking margaritas
5. Our mission at Citus Data 5 Ozgun Erdogan | DataEngConf NYC 2017 Make it so SaaS businesses never have to worry about scaling their database again
6. What is the Citus database? 1.Scales out PostgreSQL 2.Extension to PostgreSQL 3.Available in 3 Ways Ozgun Erdogan | DataEngConf NYC 2017 • Using sharding & replication • Query engine parallelizes SQL queries across many nodes • Using PostgreSQL extension APIs
7. Citus, Packaged Three Ways Ozgun Erdogan | DataEngConf NYC 2017 Open Source Enterprise Software Fully-Managed Database as a Service github.com/citusdata/citus
8. Simplified Citus Architecture
9. 3 Challenges Distributing Postgres 1. PostgreSQL and High Availability 2. To build new distributed database—or to fork? 3. Distributed transactions Ozgun Erdogan | DataEngConf NYC 2017
10. PostgreSQL & High Availability (HA) Designing for a Cloud-native world 1
11. Why is High Availability hard? PostgreSQL replication uses one primary & multiple secondary nodes. Two challenges: 1. Most Postgres clients aren’t smart. When the primary fails, they retry the same IP. 2. Postgres replicates entire state. This makes it resource intensive to reconstruct new nodes from a primary. Ozgun Erdogan | DataEngConf NYC 2017
12. Database Failures Should Be Transparent Ozgun Erdogan | DataEngConf NYC 2017
13. Database Failures Shouldn’t Be a Big Deal 1. PostgreSQL streaming replication to replicate from primary to secondary. Back up to S3. 2. Volume level replication to replicate to secondary’s volume. Back up to S3. 3. Incremental backups to S3. Reconstruct secondary nodes from S3. Ozgun Erdogan | DataEngConf NYC 2017 3 Methods for HA & Backups in Postgres
14. Postgres - Streaming Replication (1) Write-ahead logs (streaming repl.) Table foo Primary – PostgreSQL streaming repl. Table bar WAL logs Table foo Table bar WAL logs Secondary – PostgreSQL streaming repl. Monitoring Agents - streaming repl. setup & auto failover S3 / Blob Storage (Encrypted) Backup Process Ozgun Erdogan | DataEngConf NYC 2017
15. Postgres – AWS RDS & Azure (2) Postgres Primary Monitoring Agents (Auto node failover) Persistent Volume Postgres Standby S3 / Blob Storage (Encrypted) Table foo Table bar WAL logs Table foo Table bar WAL logs Backup process Backup Process Ozgun Erdogan | DataEngConf NYC 2017 Persistent Volume
16. Postgres – Reconstruct from WAL (3) Postgres Primary Monitoring Agents (Auto node failover) Persistent Volume Postgres Secondary Backup Process S3 / Blob Storage (Encrypted) Table foo Table bar WAL logs Persistent Volume Table foo Table bar WAL logs Backup process Ozgun Erdogan | DataEngConf NYC 2017
17. WHO DOES THIS? PRIMARY BENEFITS Streaming Replication (local / ephemeral disk) On-prem Manual EC2 Simple to set up Direct I/O: High I/O & large storage Disk Mirroring RDS Azure Preview Works for MySQL and PostgreSQL Data durability in cloud environments Reconstruct from WAL Heroku Citus Data Enables Fork and PITR Node reconstruction in background (Data durability in cloud environments) How do these approaches compare? 17 Ozgun Erdogan | DataEngConf NYC 2017
18. Summary • In PostgreSQL, a database node’s state gets replicated in its entirety. The replication can be set up in three ways. • Reconstructing a secondary node from S3 makes bringing up or shooting down nodes easy. • When you shard your database, the state you need to replicate per node becomes smaller. Ozgun Erdogan | DataEngConf NYC 2017
19. PostgreSQL has a huge ecosystem. How do you keep up with it? 2
20. 3 ways to build a distributed database 1. Build a distributed database from scratch 2. Middleware sharding (mimic the parser) 3. Fork your favorite database (like PostgreSQL) Ozgun Erdogan | DataEngConf NYC 2017
21. Example Transaction Block Ozgun Erdogan | DataEngConf NYC 2017
22. Postgres Features, Tools & Frameworks • PostgreSQL manual (US Letter) • Clients for diff programming languages • ORMs, libraries, GUIs • Tools (dump, restore, analyze) • New features Ozgun Erdogan | DataEngConf NYC 2017
23. At First, Forked PostgreSQL with Style Ozgun Erdogan | DataEngConf NYC 2017
24. Two Stage Query Optimization 1. Plan to minimize network I/O 2. Nodes talk to each other using SQL over libpq 3. Learned to cooperate with planner / executor bit by bit (Volcano style executor) Ozgun Erdogan | DataEngConf NYC 2017
25. Citus Architecture (Simplified) 25 SELECT avg(revenue) FROM sales Coordinator SELECT sum(revenue), count(revenue) FROM table_1001 SELECT sum … FROM table_1003 Worker node 1 Table metadata Table_1001 Table_1003 SELECT sum … FROM table_1002 SELECT sum … FROM table_1004 Worker node 2 Table_1002 Table_1004 Worker node N . . . . . . Each node PostgreSQL with Citus installed 1 shard = 1 PostgreSQL table Ozgun Erdogan | DataEngConf NYC 2017
26. Unfork Citus using Extension APIs CREATE EXTENSION citus; • System catalogs – Distributed metadata • Planner hook – Insert, Update, Delete, Select • Executor hook – Insert, Update, Delete, Select • Utility hook – Alter Table, Create Index, Vacuum, etc. • Transaction & resources handling – file descriptors, etc. • Background worker process – Maintenance processes (distributed deadlock detection, task tracker, etc.) • Logical decoding – Online data migrations Ozgun Erdogan | DataEngConf NYC 2017
27. PostgreSQL has transactions. How to handle distributed transactions 3
28. BEGIN INSERT UPDATE SELECT COMMIT ROLLBACK
29. Consistency in Distributed Databases 1. 2PC: All participating nodes need to be up 2. Paxos: Achieves consensus with quorum 3. Raft: More understandable alternative to Paxos Ozgun Erdogan | DataEngConf NYC 2017
30. Concurrency in Distributed Databases Ozgun Erdogan | DataEngConf NYC 2017
31. Locks Locks
32. What is a Lock? • Protects against concurrent modifications. • Locks are released at the end of a transaction. Deadlocks
33. Transactions Block on 1st Conflicting Lock What is a lock? Protects against concurrent modifications Locks released at end of transaction BEGIN; UPDATE data SET y = 2 WHERE x = 1; <obtained lock on rows with x = 1> COMMIT; <all locks released> BEGIN; UPDATE data SET y = 5 WHERE x = 1; <waiting for lock on rows with x = 1> <obtained lock on rows with x = 1> COMMIT;
34. Transactions and Concurrency • Transactions that don’t modify the same row can run concurrently. Transactions block on 1st lock that conflicts BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; COMMIT; <all locks released> BEGIN; UPDATE data SET y = y + 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; <waiting for lock on rows with x = 1> <obtained lock on rows with x = 1> COMMIT;
35. Transactions and Concurrency • Transactions that don’t modify the same row can run concurrently. Transactions block on 1st lock that conflicts BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; COMMIT; <all locks released> BEGIN; UPDATE data SET y = y + 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; <waiting for lock on rows with x = 1> <obtained lock on rows with x = 1> COMMIT; (Distributed) deadlock! BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; UPDATE data SET y = y + 1 WHERE x = 2; BEGIN; UPDATE data SET y = y - 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; But what if they start blocking each other?
36. Transactions and Concurrency • Transactions that don’t modify the same row can run concurrently. Transactions block on 1st lock that conflicts BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; COMMIT; <all locks released> BEGIN; UPDATE data SET y = y + 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; <waiting for lock on rows with x = 1> <obtained lock on rows with x = 1> COMMIT; (Distributed) deadlock! BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; UPDATE data SET y = y + 1 WHERE x = 2; BEGIN; UPDATE data SET y = y - 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; But what if they start blocking each other? Deadlock detection in PostgreSQL Deadlock detection builds a graph of processes that are waiting for each other.
37. Transactions and Concurrency • Transactions that don’t modify the same row can run concurrently. Transactions block on 1st lock that conflicts BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; COMMIT; <all locks released> BEGIN; UPDATE data SET y = y + 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; <waiting for lock on rows with x = 1> <obtained lock on rows with x = 1> COMMIT; (Distributed) deadlock! BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; UPDATE data SET y = y + 1 WHERE x = 2; BEGIN; UPDATE data SET y = y - 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; But what if they start blocking each other? Deadlock detection in PostgreSQL Deadlock detection builds a graph of processes that are waiting for each other. Deadlock detection in PostgreSQL Transactions are cancelled until the cycle is gone
38. Transactions and Concurrency • Transactions that don’t modify the same row can run concurrently. Transactions block on 1st lock that conflicts BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; COMMIT; <all locks released> BEGIN; UPDATE data SET y = y + 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; <waiting for lock on rows with x = 1> <obtained lock on rows with x = 1> COMMIT; (Distributed) deadlock! BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; UPDATE data SET y = y + 1 WHERE x = 2; BEGIN; UPDATE data SET y = y - 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; But what if they start blocking each other? Deadlock detection in PostgreSQL Deadlock detection builds a graph of processes that are waiting for each other. Deadlock detection in PostgreSQL Transactions are cancelled until the cycle is gone Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus Citus delegates transactions to nodes Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus Citus delegates transactions to nodes
39. Transactions and Concurrency • Transactions that don’t modify the same row can run concurrently. Transactions block on 1st lock that conflicts BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; COMMIT; <all locks released> BEGIN; UPDATE data SET y = y + 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; <waiting for lock on rows with x = 1> <obtained lock on rows with x = 1> COMMIT; (Distributed) deadlock! BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; UPDATE data SET y = y + 1 WHERE x = 2; BEGIN; UPDATE data SET y = y - 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; But what if they start blocking each other? Deadlock detection in PostgreSQL Deadlock detection builds a graph of processes that are waiting for each other. Deadlock detection in PostgreSQL Transactions are cancelled until the cycle is gone Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus Citus delegates transactions to nodes Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus Citus delegates transactions to nodes Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus PostgreSQL’s deadlock detector still works
40. Transactions and Concurrency • Transactions that don’t modify the same row can run concurrently. Transactions block on 1st lock that conflicts BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; COMMIT; <all locks released> BEGIN; UPDATE data SET y = y + 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; <waiting for lock on rows with x = 1> <obtained lock on rows with x = 1> COMMIT; (Distributed) deadlock! BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; UPDATE data SET y = y + 1 WHERE x = 2; BEGIN; UPDATE data SET y = y - 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; But what if they start blocking each other? Deadlock detection in PostgreSQL Deadlock detection builds a graph of processes that are waiting for each other. Deadlock detection in PostgreSQL Transactions are cancelled until the cycle is gone Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus Citus delegates transactions to nodes Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus Citus delegates transactions to nodes Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus PostgreSQL’s deadlock detector still works Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus When deadlocks span across node, PostgreSQL cannot help us Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus When deadlocks span across node, PostgreSQL cannot help us
41. Transactions and Concurrency • Transactions that don’t modify the same row can run concurrently. Transactions block on 1st lock that conflicts BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; COMMIT; <all locks released> BEGIN; UPDATE data SET y = y + 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; <waiting for lock on rows with x = 1> <obtained lock on rows with x = 1> COMMIT; (Distributed) deadlock! BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; UPDATE data SET y = y + 1 WHERE x = 2; BEGIN; UPDATE data SET y = y - 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; But what if they start blocking each other? Deadlock detection in PostgreSQL Deadlock detection builds a graph of processes that are waiting for each other. Deadlock detection in PostgreSQL Transactions are cancelled until the cycle is gone Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus Citus delegates transactions to nodes Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus Citus delegates transactions to nodes Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus PostgreSQL’s deadlock detector still works Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus When deadlocks span across node, PostgreSQL cannot help us Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus When deadlocks span across node, PostgreSQL cannot help us Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlock detection in Citus 7 Citus 7 adds distributed deadlock detection
42. Transactions and Concurrency • Transactions that don’t modify the same row can run concurrently. Transactions block on 1st lock that conflicts BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; COMMIT; <all locks released> BEGIN; UPDATE data SET y = y + 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; <waiting for lock on rows with x = 1> <obtained lock on rows with x = 1> COMMIT; (Distributed) deadlock! BEGIN; UPDATE data SET y = y - 1 WHERE x = 1; UPDATE data SET y = y + 1 WHERE x = 2; BEGIN; UPDATE data SET y = y - 1 WHERE x = 2; UPDATE data SET y = y + 1 WHERE x = 1; But what if they start blocking each other? Deadlock detection in PostgreSQL Deadlock detection builds a graph of processes that are waiting for each other. Deadlock detection in PostgreSQL Transactions are cancelled until the cycle is gone Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus Citus delegates transactions to nodes Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus Citus delegates transactions to nodes Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus PostgreSQL’s deadlock detector still works Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus When deadlocks span across node, PostgreSQL cannot help us Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlocks in Citus When deadlocks span across node, PostgreSQL cannot help us Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlock detection in Citus 7 Citus 7 adds distributed deadlock detection Firstname Lastname | Citus Data | Meeting Name | Month Year Deadlock detection in Citus 7 Citus 7 adds distributed deadlock detection.
43. Distributed transactions are… a complex topic • Most articles on distributed transactions focus on data consistency. • Data consistency is only one side of the coin. If you’re using a relational database, your application benefits from another key feature: deadlock detection. • https://www.citusdata.com/blog/2017/08/31/databases -and-distributed-deadlocks-a-faq Ozgun Erdogan | DataEngConf NYC 2017
44. So now what? We talked about 3 challenges distributing Postgres… 1. PostgreSQL, Replication, High Availability 2. Tradeoffs in different approaches to building a distributed database—and how we chose PostgreSQL’s extension APIs 3. Distributed deadlock detection & distributed transactions Ozgun Erdogan | DataEngConf NYC 2017
45. 45 “SQL is hard, not impossible, to scale”
46. © 2017 Citus Data. All right reserved. ozgun@citusdata.com Questions? @citusdata Ozgun Erdogan www.citusdata.com