Building an AI factory on Kubernetes

Cloudflare's new post walks through how they keep DNS cache from gobbling up RAM — a real problem when you're running thousands of zones and the cache grows until it chokes the rest of the system. The fix is a smarter eviction policy that bumps the least useful entries first instead of just dumping the oldest ones.

The same week, the CNCF published a how-to on standing up an AI inference stack on Kubernetes. They're treating GPU nodes like any other resource, scheduling models and workers across the cluster so nothing sits idle. The key insight is that AI workloads aren't fundamentally different from the apps we've been running on k8s for years — they just need more memory and faster GPUs.

Both posts are for the people who actually run the infrastructure: the engineers who stay up when something breaks, not the VCs pitching the next platform. They're practical, not promotional.

Why this matters for us: the same k8s clusters running our apps are now handling AI workloads too — and the cache bugs that slow down DNS also slow down inference, so the fixes here apply to the whole stack we depend on.