Your question is Distributed Training With Failures. Take a moment with it on the right.
Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).
How would you architect a distributed training system to handle multi-node failures?