Sep 6, 2018 · 1 min read
We actually have a pretty naive way of dealing with preemptibles being unavailable. If increasing the preemptible node pool size fails we don’t shrink the other node pool.
If you want to reverse direction all you have to do is switch NODE_POOL_FROM and NODE_POOL_TO.
I don’t think you necessarily need to run it in reverse though, since the cluster autoscaler will automatically scale a regular vm node pool if preemptibles are unavailable. It will do anything to get a place for pending pods. It might be too slow though for keeping all your applications in the green.