What actually breaks when you autoscale GKE past 50 nodes

Draft notes, not published yet.

Most GKE autoscaling guides stop at the happy path: add a node pool, set min/max, done. The failures show up later, once a cluster is big enough that scheduling latency, IP exhaustion in the node subnet, and cold-start time on new nodes start competing with each other instead of being separate problems.

Full write-up coming once I’ve written up the specific failure modes cleanly enough to be useful instead of just a war story.