It seems to me most of their challenges could be solved by replacing gke with self-managed k8s cluster specifically tuned for the use case. But then it sounded they had pretty barebones stack if they managed to pull off migration in a month
If self-managed Nomad was easier for them to configure than a managed K8s distribution, then a self-managed (possibly with custom patches) system probably would have been even worse.
There’s not a lot of knobs you can tune with gke. For example you can only choose out of two linux node image families, the control plane is not really tunable etc
Not sure nomad would support an api/real-time workload as well as k8s. But certainly a fine option for batch heavy workload