Crash-stop failure model
Crash-stop is a process-failure model in which a process follows its algorithm until it halts and then takes no further steps for the rest of the modeled execution.
In Chandra and Toueg's model, the crashed-process set can only grow, and a process does not recover. Correct processes never crash during that execution. Failure detection is modeled separately and may be mistaken; a crash assumption alone supplies neither an accurate detector nor a detection deadline. [1]
Constructed example: observer A sends a probe at tick 5 and times out at tick 10. B may have crashed at tick 7, or its reply may simply arrive at tick 12. A has the same local history at the deadline in both cases. Even a reply at tick 12 can have been sent before B crashed. These are illustrative logical ticks, not measured latency or a failure probability. Executable traces include a timely-reply control. [4]
The historical fail-stop processor additionally specifies failure-status detection and accessible stable storage. Its recovery discussion allows restarting a computation on repaired hardware; that is not permission for a crashed process to resume inside this crash-stop execution. [2] [3]
Storage durability, corruption behavior and fencing of a suspected writer need explicit contracts beyond this process model. A restart requires a recovery model or an explicitly identified new process incarnation, with its state and membership rules defined. [1]
Relations
Limitations
- The finite traces illustrate an observation ambiguity. They do not prove a general impossibility theorem, implement consensus or evaluate a real failure detector. [4]
Citations
- T. D. Chandra and S. Toueg. Unreliable Failure Detectors for Reliable Distributed Systems. JACM 43(2), 225-267, 1996; sections 1 and 2.1-2.4.
- Schlichting and Schneider. Fail-Stop Processors: An Approach to Designing Fault-Tolerant Computing Systems. TOCS 1983, sections 2 and 3.1.
- F. B. Schneider. Byzantine Generals in Action: Implementing Fail-Stop Processors. TOCS 2(2), 145-154, 1984; sections 1-2.
- Hesela: constructed observer traces for crash, delay and an in-flight reply; executable example, not a hardware experiment.
Sources & caveats
- Uses the no-recovery failure pattern of Chandra and Toueg (1996), section 2.1.
- Hesela's deterministic traces are constructed examples with explicit message ordering; no measured population or empirical failure rate is asserted.