Is there any optimization we can do when using preemptible instances running jobs.

[zwzhang0107]

/kind question

Question:

Most cloud providers have provide preemptible instance with better cost efficiency(e.g. Alibaba Cloud ECI).

In old days before k8s, spark added better handling for node shutdown, for better use of EC2 spot node. Recently, there are also some discussions in ASPLOS'24.

I'm wondering is there any work (maybe already done?) to make jobs better use of preemptible instances. I will be honored to join if there is any plan.