We have that 'issue' with many of our Spark workloads where there isn't any of our desired capacity available as spot, but we have a baseline reserved up front

We have that 'issue' with many of our Spark workloads where there isn't any of our desired capacity available as spot, but we have a baseline reserved up front instances for anything realtime anyway so with a bit of planning it's a non-issue. It does cost money, but then again, so does not running certain processes. The trick becomes calculating the intersection at which point the costs outweigh the benefits, and that calculation applies everywhere.