Hey!

I have a suggestion/feature request for the autoscaler.
Would it be possible to define multiple "worker templates", tagged with custom resources? So you can e.g. define a worker template for workers with GPU, so you don't have to pay for a GPU if there is no GPU needed.
In the same way it would enable using conflicting requirements in different parts of your code. You could run for example tensorflow 1.15 and 2.0 next to each other, but in different workers, by defining multiple worker manifests (using different docker images) and tagging them with a 'custom resource': tensorflow2.0 or tensorflow1.15... So when Ray autoscales, it checks which resources are depleted and then spawns the right worker type...?

Thanks in advance for considering it!
Pieterjan