1 - Do not run very large jobs on AWS spot instances only (specifically GPU instances) 2 - If you have to run something large, make sure it can be paused and r

1 - Do not run very large jobs on AWS spot instances only (specifically GPU instances) 2 - If you have to run something large, make sure it can be paused and resumed later. 3 - Run that job on the cheapest Spot/OnDemand instance you need. 4 - If the price is right on the other side of the spot/demand side, provision a new machine, checkpoint the job, and restart.