Production breaks in interesting ways.
I write some of those down.
Post-mortems, scaling stories, and the trade-offs nobody writes down.
Featured
Whose GPUs Are These? Building a Zero-Footprint GPU Extension for Freelens
Why desktop Kubernetes tools can't answer basic GPU questions, and how I built a Freelens extension that reads the exporters you already run — no DaemonSet, no port-forward, no Prometheus required — plus the MIG and power-attribution bugs I hit along the way.
More posts
All posts →MIG vs MPS in Kubernetes: Re-Carving a GPU Is Not a Pod Restart
What MIG and MPS actually give you on A100/H100, why the GPU Operator won't let you mix them on one node, how I ran one service on MPS anyway — and the three ways a MIG change took GPUs (and once, the whole node) down.
Cutting a Managed Prometheus Bill by 70% (While Adding Metrics)
How to attribute a Google Managed Prometheus bill to individual scrape jobs, turn off the GKE default collectors nothing reads, and add app metrics through measured keep-lists — plus the two cardinality bombs that would have multiplied the bill 11×.
Stop Port-Forwarding to 15672: A Cluster-Native RabbitMQ Console for Freelens
Why I built a RabbitMQ extension for Freelens, how it finds brokers and credentials and tunnels to the Management API from inside the desktop app, and what it takes to make dead-letter replay safe.
Networking Under the Hood: Diagnosing Latency and Packet Drops in the Linux Kernel
Tracing a packet's journey through the Linux kernel: ss, nstat, ethtool, tcpdump beyond the basics, and eBPF drop-reason tracing — plus a full walkthrough of a silent accept-queue overflow incident.