Design Patterns for Checkpoint-Based Rollback Recovery

Checkpoint-based rollback recovery is a very popular category of fault tolerance techniques, which are based on a simple idea: save the system state during error-free portions of the system execution; when an error occurs, use the saved state to rollback the system to a recent consistent state. This way, after an error occurs the system does not have to start its execution from the beginning, which would result in longer execution times or even failure of the system (e.g. when the I/O events that drove the execution of the system are not reproducible). This paper presents three design patterns that capture the most widely used methods for checkpoint-based rollback recovery. The Independent Checkpoint pattern describes the method where constituent components of a system take checkpoints without synchronizing with each other. Synchronization will take place after the occurrence of an error when a consistent system state must be re-established from the partial system states found in the checkpoints. The Coordinated Checkpoint pattern describes the method where constituent components of a system take checkpoints after synchronizing with each other. In this case, no synchronization is required during the re-establishment of a consistent system state after the occurrence of an error. Finally, the Communication-Induced Checkpoint pattern describes the methods where the synchronization of the checkpointing is triggered by communication events. This method combines the benefits of the previous two in terms of time and space overhead incurred to the system execution.