Transactor Performance Theory & Practice
summary
Gives a concrete performance model: queuing + the working‑set model, measurable metrics like apply msec and TX‑MC, and practical tools — add an NVMe val cache to cut IO and double throughput from ~15k to ~38k tx/min in a 5,000 TPS saturation test, switch to sequential IDs (squids) to reduce object‑cache needs and reach ~54k/min (~950 TPS) and billions of datoms, and a new datomic hints / prefetch system that moves segment discovery to peers and overlap IO while queued, substantially lowering transactor utilization.
AI-written summary · about
chapters (9)
- 0:00Transactor performance intro
- 1:06Datomic architecture
- 3:57Queuing theory matters
- 10:45Queuing theory in systems
- 14:03Working set model
- 16:5510B data speedrun
- 25:54Improving transactor performance
- 33:00Recent transactor improvements
- 37:39Transaction life cycle
related talks
15mTake Out the TraChe: Maximizing (Tra)nsactional Ca(che) Hit Rate
41mTransactions: myths, surprises and opportunities
54mThe Top 8 Reasons for Your Transaction Performance Problems
33mCassandra Performance Tuning Like You've Been Doing It for Ten Years
36mThe Journey to Achieve 1000x TPS for a Real-World Core Banking Application on TiDB
47mLoad Testing Storage Performance for SQL Server
19mFast Distributed Transactions for RDMA-based Disaggregated Memory
49m10ms Response with Transactions, Big Data, Flexible Logic and OpenSource