Dora Noda
Software Engineer
2336 posts · View all authors
Railway's 8-Hour Outage Was an Architecture Problem, Not a Bug
Railway's May 2026 outage wasn't just a GCP problem — it exposed why multi-cloud doesn't save you when your control plane's routing table lives in one vendor's account.
Railway Killed Its Free Tier for a One-Time $5 Credit: What a Real Hobby App Costs on Railway vs a Hetzner Box
Railway's free tier died in 2023 and never came back. Here's the line-by-line recompute of what a real web+worker+Postgres hobby app costs on Railway's metered Hobby plan versus a flat €5.49/month Hetzner box.
Render's Buildpacks Aren't 75% Faster Than Dockerfiles — Here's What Its 2026 Changelog Actually Says
The '75% faster' Render buildpacks claim traces back to no real Render source. Here's what its 2026 changelog actually documents, why Railway rebuilt Nixpacks around the same caching shift, and how to get the same layer caching self-hosted with Paketo, kpack, and Railpack.
Render's April 23, 2026 Repricing: Flat Fees Land, but Included Egress Drops From 100GB to 5GB
Render's April 2026 repricing swapped per-seat fees for flat ones but slashed included egress 95% on Hobby and Pro. Here's the line-by-line bill recompute at real traffic levels, plus what the same stack costs on a Hetzner box.
The Self-Hosted GPU Breakeven Point: Why ~50M Tokens/Day Is Where Owning H100s Beats Renting Inference APIs
A dedicated H100 breaks even against typical hosted LLM API pricing at roughly 50 million tokens a day of sustained use, but the real number depends entirely on which API tier you're comparing against and how well a platform can bin-pack tenant traffic to keep the GPU busy.
Short-Lived Credentials for AI Agents on Kubernetes: Designing Out the Long-Lived-Secret Failure Mode
Why plain Kubernetes Secrets are the failure mode behind 2026's AI-agent credential incidents, and the concrete Vault/CSI/MCP architecture that replaces them with short-lived, per-task tokens.
How SNCF Keeps 200+ Kubernetes Clusters in Sync With Zero Drift — and Why the Same Loop Works on Three Hetzner Boxes
SNCF, France's national railway, updates 200+ Kubernetes clusters every month with zero drift using Cluster API. Here's exactly how that reconciliation loop works — and why the same mechanism runs unmodified on three self-hosted Hetzner boxes.
Talos Linux Meets Cluster API: The Immutable-OS Alternative to Ubuntu+kubeadm
Talos Linux strips the Kubernetes node OS down to a dozen binaries, no SSH, and a single mTLS API — and it plugs into Cluster API through its own bootstrap and control-plane providers, right alongside CAPH on Hetzner.
Vercel's Active CPU Billing Saves 93% on a Real AI Agent — Here's the Exact Math
Vercel claims Active CPU pricing cuts costs up to 95% across 45 billion weekly requests. A worked example on a realistic AI-agent function gets 93% — and shows the exact invocation volume where the discounted bill catches back up to a flat-rate box you already own.