Spot recovery strategy

[yuval-airis]

I saw that two StrategyExecutor for spot recovery are defined (FailoverStrategyExecutor and EagerFailoverStrategyExecutor for the sky.Resources class.

The former tries to failover within the same cluster, and the latter prefers to failover to a machine in another cluster.

Can I also set some more complicated method of picking the right cluster? (e.g. the cheapest one / nearest one) Would that be achieved by implementing an executor of my own?

[concretevitamin]

The former tries to failover within the same cluster, and the latter prefers to failover to a machine in another cluster. Can I also set some more complicated method of picking the right cluster? (e.g. the cheapest one / nearest one)

Hey @yuval-airis. To clarify:

• FailoverStrategyExecutor does have a chance to launch a new cluster elsewhere.
• Both strategies use sky.launch() as a kernel to relaunch, and this function already optimizes for the cheapest and most available location (zone, region, cloud) within a given search space.

The two strategies differ in whether to back off from a preempted region quickly.

Yes, you could implement a new StrategyExecutor and put custom logic in recover(). Curious what policy are you thinking of?