request: flag to lower 15 instance type requirement for spot-to-spot consolidation?

[drawnwren]

Description

What problem are you trying to solve?
I'm running karpenter with gpu workloads and aws often doesn't even offer 15 instance types with the gpu that a workload requires. I'd still like to try spotToSpot with these workloads because of the huge cost savings associated with spot gpu instances currently. Would possibly configuring the 15 instance type limit make sense in this case?

[jonathan-innis]

Would possibly configuring the 15 instance type limit make sense in this case

I definitely think that we'd be open to make this configurable. What value would you like to set it to? The biggest reason that we went with 15 is -- we did some analysis work and found that this is around the right number of instances to ensure that you have enough flexibility in your launch request that you won't get an immediate spot interruption after the launch. This is obviously more of a heuristic, but you're definitely in uncharted waters in terms of your propensity to potentially get re-interrupted on a consolidation if you try to lower this number

[jonathan-innis]

@drawnwren Have you considered forking this (just for PoC/testing) and see what kind of performance you get?

[jonathan-innis]

/triage accepted