Considerations over current preemptible nodes approach
I've recently realized that this GKE provisioned example currently provisions preemptible nodes. It wasn't a conscious decision of mine back then.
Preemptible Virtual Machines are:
Affordable compute instances suitable for batch jobs and fault-tolerant workloads
They have some important limitations.
The great benefit that they bring is that they can be up to 80% cheaper than regular ones. In turn, you can get some errors every now and then, less performant results and lack of GCP SLA coverage, etc.
My thoughts regarding this are:
-
We're fostering users to create preemptible nodes with this example. While cheaper, it's probably safer to provision regular nodes since many users might not even know what this is about.
-
I wouldn't want to start adding much details to the docs as well, since I wouldn't want to get into the same situation with have with cluster management project where we start documenting a lot of technical bits of external tools.
-
Although, one point in favor of it is that:
Containers are naturally stateless and fault tolerant, making them a great fit for preemptible VMs!
Still, I think performance could be affected and errors can happen if services are requested during a preemptible VM restart window, which seems to happen at least daily.
@nagyv-gitlab @nicholasklick do you have thoughts what you prefer the example to have?
Changing it to regular nodes should be as easy as just setting this to false.