GCP quotas are fairly limited (on my account, I can only
use 8 public IP addresses per zone, which means that I cannot
deploy many public clusters in a single zone). I tried to
use private clusters, but that causes other problems.
This refactoring makes it possible to spread clusters
across multiple zones. Since I have access to 20+ zones
in Europe and 20+ zones in the US, this lets me create a
lot of public clusters and simplifies the module quite a bit.
Instead of doing a local-exec to 'kubectl get nodes ... > stage2/externalips.X',
we use an external provisioner. And to decouple things, the external provisioner
depends on nodes being ready, and node readiness is implemented with a null
resource. That way we don't need to re-execute the whole 'wait for nodes; dump
node addresses' dance each time we re-run that configuration.
This is a new provisioning mechanism. Right now, it can
provision clusters on:
- Digital Ocean
- Linode
- Oracle Cloud
- Scaleway
Others should be relatively straightforward to add.
Check the README in the prepare-tf subdirectory for details.
