Fix(google/gke): require node_count or autoscaling on node pools

What does this MR do and why?

A minimal node_pools entry inherited node_count = 0 with autoscaling unset, producing a node pool with no nodes. Nothing could be scheduled, so terraform apply waited on the operator Deployment rollout until it failed with an opaque context deadline exceeded.

This adds a validation to node_pools that rejects a pool configuring neither, with a message naming the actual problem.

The issue offers "default node_count to 1, and/or add a validation". I went with validation only, because changing the default would regress autoscaling users.

main.tf passes node_count = each.value.node_count > 0 ? each.value.node_count : null. Zero is the sentinel that hands sizing to the autoscaling block. Defaulting to 1 would send a node_count alongside autoscaling for every pool that configures only autoscaling.

It also keeps README.md untouched: terraform-docs does not render validation blocks, but it does render defaults.

One thing worth a close look: the test fixture was an instance of the same bug.

In gke_test.go, vpc and autoscaling were nested inside node_pools rather than beside it. Terraform discards attributes absent from an object type constraint, so each silently became a node pool with everything defaulted — node_count = 0, autoscaling = null. TestGKE asserted all three, so the suite has been planning three zero-node pools.

Both now sit at the top level, matching the module's own vpc and autoscaling variables at variables.tf:157 and :79, and TestGKE expects the one real pool. I am reading that nesting as accidental — please confirm, since it is the part of this MR that changes existing expectations.

Differences

Before: node_pools = { default = {} } planned successfully and created a pool with zero nodes. The failure surfaced much later as context deadline exceeded while waiting on the operator rollout, pointing at the operator rather than the config.

After: the same config fails immediately at variable validation, naming the pool requirement directly. Pools with node_count > 0, or with autoscaling configured, are unaffected.

How to set up and validate locally

Variable validation is evaluated before the provider needs credentials, so these two can be checked without GCP access:

go test ./modules/google/gke/ -run 'TestGKEPlanErrors/node-pool'
terraform fmt -check modules/google/gke/

node-pool-without-nodes-or-autoscaling trips the new rule; node-pool-with-autoscaling-and-no-node-count does not.

Full plan assertions (TestGKE) need real GCP credentials and fail locally with could not find default credentials, so those run in CI.

Related to

• Fixes #234 (closed)