Breakout Session
summary
Automates PostgreSQL failover via an orchestrated cluster state in Manatee using ZooKeeper for consensus, a separate client layer Moire, ZFS for efficient sends and three-node replication (three peers with primary, sync, async). Runs in three AZs, holds ~200M rows and several hundred GB per shard on PostgreSQL 9.2. Shows concrete tradeoffs: synchronous sync to avoid data loss, 20 minute rebuilds via zfs send, seconds‑scale typical failover but rare multi‑hour vacuum outages; everything is built and deployed with real operational lessons.
AI-written summary · about
chapters (18)
- 0:00Start
- 0:40Background
- 1:25The Problem
- 4:08Manta: storage architecture
- 5:40Metadata tier
- 7:40PostgreSQL for metadata
- 8:34Manatee cluster
- 10:09Unexpected constraint
- 10:42Primary fails
- 11:48Fault tolerance
- 12:16Cluster components
- 14:12Demo
- 21:36Moray: custom client
- 22:52Horizontal scalability
- 26:34Manatee/Moray for jobs
- 31:25Incident 2: queries hung
- 38:32PG locking
- 43:47Tools we built
related talks
55mHigh Availability with PostgreSQL and Pacemaker
36mDesigning Layered High‑Availability Architectures for PostgreSQL on a Budget
57mReliability features in Postgres Pro
41mEnsuring High Availability in Self-Managed PostgreSQL with Patroni
50mThe Immortal Postgres Web
51mPostgreSQL on ZFS
14mBuilding a better PostgreSQL together
13mBuilding reliable infrastructure with Postgres