Evolving Transactions in Apache Cassandra
summary
Cassandra reduces LWT phases by merging Paxos prepare and read, cutting geo latency from four to two wide‑area round trips for writes and to one wide‑area round trip for reads when uncontended, implements a deterministic simulator plus a linearizability verifier, and runs in production with ~1.5 trillion read-only and 0.25 trillion write transactions to date. Proposal adds leaderless, multi-shard consensus (Accord/tempo ideas) to guarantee strict serializability with one‑round latency in normal conditions and resilient fast paths that adapt electorates under failures.
AI-written summary · about
chapters (9)
- 0:00Introduction to Cassandra transactions
- 3:03Lightweight transactions and Paxos
- 8:42Performance improvements in Paxos
- 13:14General purpose distributed transactions
- 19:33Leaderless consensus protocols
- 28:47Ideal transaction properties
- 34:25Failure properties and recovery quorums
- 37:12Future of Cassandra transactions
- 46:46Q&A session
related talks
30mApache Cassandra: What’s New and What’s Next
34mThe Accord Consensus Protocol: ACID transactions globally at any scale
34mCassandra on ACID: This Changes Everything
48mThe Accord Consensus Protocol, how to do ACID transactions at scale
46mData Replication with Apache Accord
53mApache Cassandra. Inception - all you need to know
45mCassandra Architecture: Understanding, Scaling and Optimizing
37mGeo-Replicated Transactions in 1.5RTT