Citus Cluster: our zero-downtime migration story
summary
Algolia migrates a 5TB analytics Citus cluster (spread over 20 machines) from Citus Cloud to Azure Hyperscale with goals of zero downtime, stability, and no cost increase. Preparation includes a step-by-step playbook, heavy monitoring, and Microsoft creating Hyperscale followers. Writers and reads are tested; pg tools (like pg_stat_statements and EXPLAIN ANALYZE) find a JIT caused 2x regression on write-heavy rollup jobs and a libc-triggered index corruption fixed by disabling JIT and reindexing. Each cluster cutover takes ~40 minutes and succeeds with minimal user impact.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 2:09About Algolia, the analytics dashboard, & the dashboard architecture
- 3:56Why did Algolia need to migrate?
- 4:57Migration goals
- 7:00Preparing for the migration
- 18:57Executing the migration
- 20:48Migration timeline
- 21:28Key takeaways
- 24:30Q&A with Matthieu, Antoine, & Claire
related talks
24mCitus from the Customer eyes
47mSmooth Sailing...
25mSelf Hosted Citus using Cloud Computing
16mOnline Scale-Out and Cloud Data Ingestion
24mLessons learned: Migrating from AWS-Hosted PostgreSQL RDS to Self-Hosted Citus
52mScaling a Saas application beyond a single Postgres with Citus
02h 02mAzure Cosmos DB: Liftoff
08mFoundations of Distributed PostgreSQL with Citus