Dora Noda
Software Engineer
2304 posts · View all authors
What a Git-Push PaaS Should Actually Clone: Shallow vs Partial Clones and What Checkout Time Adds to Every Build
Shallow clones are 4-6x faster and the industry default, but they break versioning tools and make every later fetch 25x more expensive. Benchmarks, real PaaS defaults from Render to Heroku, and the break-even point where blobless partial clones win.
Field Is Immutable: Why helm upgrade Dies at the API Server, and Which Escape Hatch Actually Scales
A community Helm chart renames one label and your upgrade dies in the API server's validation path. Here is the full table of immutable Kubernetes fields with their exact error strings, the five ways out priced against each other, and the fleet size where a platform operator finally beats the runbook.
Hetzner Is Building an LLM Inference API: What €889/Month of GPU and a Free Experimental Endpoint Do to the Self-Hosted AI Math
Hetzner is testing an OpenAI-compatible inference API alongside its €889/month RTX PRO 6000 GEX131 — what the experiment changes for self-hosted AI rent-vs-own math, EU sovereignty, and running inference as a regular tenant workload.
Hetzner Unfroze the 50-Second Load Balancer Idle Timeout — Your Build Logs Will Still Die at 15 Seconds
Hetzner made its 50-second load balancer idle timeout configurable in April 2026, but on a Kubernetes stack it is rarely the rung that kills your log stream. The full timeout ladder — with real defaults, the idle-vs-total distinction, and the heartbeat math that survives all of it.
Hetzner Removed the EC2-Compatible Metadata Routes on August 1 — Every Bootstrap Script Still Curling /latest/ Is Now Broken
On August 1, 2026 Hetzner removed the EC2-compatible /latest/ and /2009-04-04/ metadata routes, breaking every bootstrap script that inherited the AWS curl idiom. The full route-by-route migration table, a repo audit command, and the CAPH preKubeadmCommands diff that fixes it.
Kubernetes Finally Has a Checkpoint/Restore Working Group. Here's What Pausing a Mid-Task Agent Sandbox Actually Saves
A line-by-line cost model for pausing idle AI agent sandboxes on Kubernetes: PVC hibernation already captures 67% of the savings, CRIU checkpoint/restore is worth a further 59%, and today's kubelet checkpoint API can't deliver either.
Your MCP Server Lost Its Confirmation Prompt on July 28: Rebuilding the Delete Gate on input_required
The 2026-07-28 MCP spec deleted server-initiated requests, and with them the only way a destructive tool could stop and ask a human. Here is the replacement on the wire: a delete_service gate built on resultType input_required, the signed requestState ticket that survives a round-robin load balancer, and the once-only guarantee the protocol makes your problem.
90% Off Is the Break-Even: Preview-Environment Cost Math on Hardware You Already Bought
Preview-environment guides all say the same thing: put it on spot capacity for up to 90% off. Priced against a Hetzner box you own, 88.9% is exactly the break-even — so here is the real cost table, the duty cycle where renting still wins, and the three knobs that move the number when the meter is gone.
How the controller-runtime Cache Actually Works: What PaaS Builders Should Know Before Their First OOMKill
Your operator's Get() never hits the API server — it silently caches every object of that Kind in the cluster. How controller-runtime's shared-informer cache works, why it OOMKills control planes at fleet scale, and the exact selectors, transforms, and DisableFor config that shrink a PaaS control plane from gigabytes to a few hundred MB.