Crash-recovery model
The crash-recovery model allows a process to stop and later resume under an explicit contract for which state survives the crash.
Unlike crash-stop, recovery permits the same process to return. In the 1998 model, identities are fixed, recovery is locally recognized, volatile state is lost, and explicitly saved stable state can survive. [1]
An unstable process crashes and recovers infinitely often; a finite restart trace cannot establish that whole-execution classification. Fair-loss links do not promise delivery of each message. [1]
Stable-state writes and message sends are separate actions. Recovery therefore needs an ordering contract, not merely persistent media. [1]
Raft's currentTerm, votedFor and log are persistent state; relevant changes precede RPC responses. Recovered participation still requires the protocol's rules. [2]
The author's errata qualify lastApplied: its lifetime must match the state machine's. A durable state machine needs durable applied progress; treating the index as always volatile is unsafe. [3]
A valid local file checksum need not establish freshness. Paxos Made Live explicitly excludes old-state rollback from that check. Its separately ordered database-checksum comparison addresses replica divergence. Neither check alone restores voting authority after lost protocol state. [5]
Relations
- related
- Stable storage, Consensus, Raft, Crash-stop failure model, Paxos
Citations
- Aguilera, Chen and Toueg. Failure Detection and Consensus in the Crash-Recovery Model. DISC 1998, pp. 231-245, sections 2.1-2.4.
- Ongaro and Ousterhout. In Search of an Understandable Consensus Algorithm, extended version (2014), Figure 2 and sections 5.5-5.6.
- Diego Ongaro. Consensus: Bridging Theory and Practice, author errata for Figure 3.1 (Chapter 3), repository revision 0c10ea1b53d73bf6d163bdde8e4141a1cbd8f36a.
- Chandra, Griesemer and Redstone. Paxos Made Live. PODC 2007, sections 5.1 and 6.2, footnotes 2-3; DOI 10.1145/1281100.1281103.
Further reading
Sources & caveats
- Primary-source model synthesis, not a protocol implementation, failure experiment or proof of the paper's consensus bounds.