Code That Can't Fail
summary
Explains how durable execution makes a function guaranteed to resume after crashes, redeploys or long sleeps by persisting each step; implementation uses Temporal with workers, gRPC server, and Cassandra for a write‑heavy history plus a transactional outbox to move tasks. Describes a single‑partition batch for atomic state+task+timeout, a tombstone problem (10,000 inserts → 99,999 tombstones causing ~300ms scans and 10‑day GC), and a custom write‑ahead log that drops P90 start latency from ~90ms to average ~6ms and aims for a ~10ms P90.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 3:28Durable execution matters
- 6:36How durable execution works
- 8:00Temporal architecture
- 10:45Cassandra in durable execution
- 13:58Cassandra anti-patterns and fixes
- 15:00Cassandra learnings
- 16:01Optimizing Cassandra with WAL
- 17:24Recap and key takeaways
related talks
26mMicroservices Madness: Application And Infrastructure Automation Using Temporal
55mCode the future, now
31mBeing Creative: Tales from the Trenches
41mFixin the Hard Bits of Event Processing with Restate & Kafka
55mCode the future, now
38mFree Your Mind, Your App Will Follow
36mTask Failed Successfully
12mRequirements [Citation needed]