Motivations

Rollback recovery - restoring a correct state using checkpoints.

In a distributed environment, this entails:

• collecting useful checkpoints during normal computation;
• designating appropriate checkpoints to restore computation when a failure occurs;
• removing those checkpoints that are deemed obsolete.

One unifying concept behind all three activities: "cause-effect" relationship preservation. If we remove the cause we must remove the effect too. (not really so but evocative)

A huge amount of excellent theoretical work on the subject, few experimental papers, a few partial implementations.

None of the approaches exploits the existence of a unifying concept: the three subproblems are usually treated independently.

We would like to divise a practically appealing design, that exploits the unifying concept.

Next: Directions