What We Talk About When We Talk About Distributed Systems
summary
Survey of consensus and replication literature that argues FLP impossibility means no deterministic consensus in a fully asynchronous model even with one crash, explains timing models (synchronous is unrealistic), compares Paxos/Zab/Raft tradeoffs, describes failure detectors (eventual accuracy vs perfect), cites TLA+/model checking and recommends QuickCheck testing for finding subtle bugs, names key books (Lynch, Birman/Isis) and warns that many implementations fail in real-world failure modes despite sounding correct.
AI-written summary · about
chapters (13)
- 0:00Introduction and context
- 2:44Defining distributed systems
- 4:49The challenge of paper shame
- 7:10Navigating literature and books
- 10:08Models, timing, and failures
- 13:24Types of system faults
- 17:01Liveness and safety properties
- 20:13Consensus in practice
- 23:47Atomic broadcast and grouping
- 26:41Failure detectors and trade-offs
- 30:26Consistency models
- 31:50Recommended reading and resources
- 36:53Conclusion and Q&A
related talks
27mDistributed Systems, Like It or Not
51mDo Not Reimplement Distributed Consensus Algorithms
44mDistributed systems and scalability rules
52mWhat We Talk About When We Talk About Distributed Systems
40mConsensus Algorithms in Distributed Systems
28mWhy Are Distributed Systems So Hard?
42mConsensus Algorithms in Distributed Systems (Theory and Practice)
48mDistributed Consensus Algorithms