Sharded & Distributed Are Not the Same:
summary
Claims that Citus is not ACID for multi-shard transactions and that two-phase commit only provides atomic commit (not isolation) drive the talk; proof comes from code-level reasoning, TPCC runs and a BenchBase fork with Java-thread fixes. Tuned PostgreSQL on three 128‑core servers reaches ~200k tpmC in fault‑tolerant config, exposing a single‑threaded synchronous‑replication bottleneck and checkpoint latency spikes. Distributed systems show distributed transactions under ~50 ms in many cases despite added RTTs.
AI-written summary · about
chapters (16)
- 0:00Introduction and speaker intro
- 1:49Myths about YDB and Yugabyte
- 2:58Sharding vs distributed systems
- 4:46ACID properties and isolation levels
- 6:50Citus not ACID compliant
- 8:16Two-phase commit misconceptions
- 9:22Why transactions are expensive
- 11:34Replication depth and concerns
- 15:06Citus limitations with multi-shard
- 15:49When PostgreSQL isn't enough
- 19:00TPC-C benchmark explained
- 21:51PostgreSQL tuning for TPC-C
- 25:28Distributed systems vs PostgreSQL
- 27:13TPC-C benchmark results
- 28:38Conclusion and recommendations
- 30:01Q&A on ACID compliance
related talks
01h 00mDistributed transactions implementation trade-offs
44mDistributed Postgres with Citus
31mSharding Without the Complexity
31mKEYNOTE: The Distributed PostgreSQL Problem & How Citus Solves it
49mThe Paradigm Shift - Understanding the Move To Distributed Databases
47mBeing transactional in 2019 - what does this actually mean?
08mFoundations of Distributed PostgreSQL with Citus
38mPostgreSQL as a distributed computing platform