"Operating Datomic at Scale"
summary
Nubank runs Datomic across ~3,000 databases for ~75 million customers, with >300 DBs holding >10 billion datoms; they use io-stats plus query-stats to find and reorder costly query clauses, cutting unnecessary work. Tuning the memory-index threshold reduces index jobs from ~20–30/day to ~4/day and halves write storage costs; switching JVM GC to ZGC shrinks stop‑the‑world pauses, raises transactor throughput ~33%, and recovers millions of potential transactions lost to GC.
AI-written summary · about
chapters (14)
- 0:00Start
- 2:06What scale?
- 3:30Order of clauses
- 8:21Enabling io-stats
- 10:09Identify bottlenecks
- 10:59Optimize cache hits
- 13:26Indexing job
- 14:11Memory index
- 14:41Requesting indexing
- 16:04Cost savings
- 17:04Memory is not free!
- 19:55The stakeholder problem
- 20:38The technical problems
- 23:28Remember!
related talks
28mDatomic at Nubank, Simplicity Scaled
51mHow Nubank leverages Datomic's database as an immutable value to run at scale
40mDatomic: The Most Innovative DB You've Never Heard Of
41mExploring four hidden superpowers of Datomic
51mDatomic
37m11 insights after 11 years with the functional database Datomic
40mThe power of Datomic database
41mReal-World Datomic: An Experience Report