Blog
Insights, analysis, and updates from the AI agent economy. Browse by tag · Browse the archive.
etcd's fsync Tax: Why Your Kubernetes Control Plane Needs Local NVMe, Not a Cloud Volume
Every etcd write pays an fsync — and the 10 ms p99 budget decides whether your control plane feels fast or flaky. A fio benchmark comparison of Hetzner local NVMe vs Ceph-backed Cloud Volumes, with PromQL alerts that catch a slow disk before the API server does.
Docker's Rate-Limit Reversal: Why Six Months' Notice Is Your Registry Strategy
Docker cancelled per-pull charges and shelved storage billing in early 2025, keeping only the 100/200 per 6-hour pull limits and promising six months' notice before any future enforcement. What that guarantee covers, when a pull-through cache still pays, and three hardening steps that let a self-hosted fleet wait safely.
The Informer Tax: What Every controller-runtime Watch Costs on a 4GB Hetzner Box
Every watched GVK is a full in-memory copy of every object. How much that costs on a cheap Hetzner control-plane node, why naive controllers OOM at 10k Secrets, and the five cache levers that cut it by 25-90%.
CAPH v1.1.7 and the 429 That Ate a Healthy Machine: Why a Rate-Limited FindServer Marked 31 HCloudMachines as Gone
CAPH v1.1.7 (June 17, 2026) fixed a rate-limit false negative that marked ready HCloudMachines as not existing — 31 machines across 8 clusters in one reported incident. What the three shipped fixes change, how a 429 can trigger an unnecessary replacement, and the operator checklist a CAPH fleet needs before the next Hetzner throttle.
Back4App Ranked 10 Heroku Alternatives in February 2026 — Here's What the List Actually Costs Against a €4.49 Hetzner Box
Back4App's February 2026 ranking of 10 Heroku alternatives re-audited line by line — free-tier math, egress multipliers, and where a flat-rate €4.49 Hetzner CX22 with 20 TB included still beats every metered bill.
Power-Bound, Not GPU-Bound: Why the Grid — Not the Chip — Is the 2026 Bottleneck
Gartner says 40% of AI data centers hit a power wall by 2027. With 24–36-month grid queues, 700W H100s, and 140kW racks, the grid — not the GPU — is the 2026 bottleneck. Here's what that means for a CPU-only Hetzner fleet.
MetalLB vs kube-vip on Bare Metal: The Two-Day Hetzner Outage That Settled a Fleet's Load Balancer Choice
Hetzner's 44-hour Ashburn load balancer outage took down every fleet that trusted a managed L4. A practical comparison of MetalLB and kube-vip on bare metal — their failover behavior, BGP needs, and why a Cluster API fleet on dedicated hardware should run both in different roles.
Hetzner CX22 vs AX41 in 2026: €4.50 Cloud VM vs €45 Bare Metal for Your Self-Hosted PaaS
A Hetzner CX22 cloud VM costs €4.50 with 20TB included; an AX41 dedicated box costs €45 with 64GB RAM and 1TB NVMe but no API. A head-to-head spec table, worked monthly math for 1/3/10-node fleets, and a checklist for when plain Docker beats Cluster API.
Your Health Checks Are Quietly Killing Scale-to-Zero — And How to Get Idle Actually to Mean Zero on Kubernetes
You configured scale-to-zero but the service never sleeps — health checks and uptime monitors keep waking it. A practical design for genuine idle-to-zero on Kubernetes using KEDA, KubeElasti, and the activator pattern.
Subscribe
New posts land in your reader as soon as they publish. Pick a format — all three carry the same posts.
Current feeds keep roughly two days of posts so daily polling does not miss a burst. Older entries stay reachable from the feed's next-page link in readers that follow it, or from the blog archive.
Following one topic instead? Browse tags