Pushing Kubernetes to 1M nodes isn’t just hardware—it's architectural judo.
Networking flips to exclusive IPv6. Less chatter, more breathing room.
etcd hits a wall. Write throughput stalls at scale, so they swap it out.
Enter mem_etcd, a Rust-built replacement pushing over 1M buffered writes per second.
The scheduler gets chopped up. A sharded design splits work between Relay and worker schedulers. They ditch kube-apiserver’s watch latency, wiring in webhook events instead.
