684 posts tagged with "Self-Hosting"
Running your own PaaS and infrastructure on machines you own
Kubernetes 1.36's CRI List Streaming Doesn't Cut Kubelet Memory — It Removes a Hard Container-Count Wall Bin-Packed Nodes Actually Hit
KEP-5825's own text admits CRI List Streaming does not reduce kubelet memory usage. The real fix is a hard 11,000-container / 14,000-pod gRPC ceiling that's already bricked production nodes — and it matters more on a fixed Hetzner box than an autoscaling cloud fleet.
Kubernetes' Partitionable GPU Devices (KEP-4815): The Real Cost Math Behind Slicing One GPU Into 7 Schedulable Sandboxes
Kubernetes' Partitionable Devices KEP lets one GPU serve multiple independently-scheduled sandboxes instead of one tenant. Here's the actual cost math on 7-way MIG slicing versus whole-card allocation, and what a Cluster API GPU node needs to expose those slices.
Kubernetes 1.34 Ships KYAML: What a Safer YAML Subset Does (and Doesn't) Fix for a Manifest-Generating PaaS
Kubernetes 1.34 shipped KYAML to kill YAML's Norway-problem bugs, but the KEP is explicit that it's a kubectl output format, not a server-side change. Here's what that actually means for a platform generating Cluster API and tenant manifests programmatically.
Kubernetes SIG Storage's Next Move Isn't Another Snapshot Feature: What Volume Health Monitoring and Mutable PV Affinity Actually Buy a Self-Hosted Fleet
Kubernetes SIG Storage's Volume Health Monitoring and Mutable PV Affinity don't need Hetzner's still-missing CSI snapshot support to matter — but only one of them actually works on a self-hosted fleet today. Here's the verified breakdown of both KEPs, straight from the spec text and the driver source.
Railway vs. Fly.io vs. Owning the Box: The Cost Comparison No PaaS Vendor Will Run For You
Northflank's Railway-vs-Fly.io teardown never asks what the same workload costs on an owned Hetzner box — here's the post-price-hike math, tier by tier, including the one case where per-second billing is supposed to win.
Your Node Buildpack's --ignore-scripts Flag Didn't Stop Phantom Gyp
A June 2026 npm worm called Phantom Gyp bypassed --ignore-scripts entirely by hiding in binding.gyp instead of a lifecycle script — here's the exact mechanism, why it worked, and what a git-push buildpack needs to do instead of trusting one flag.
Pulumi Neo Is GA and Reads Your Terraform State Directly — Here's Why bex's Deploy API Won't
Pulumi Neo reads live Terraform and CloudFormation state directly and executes changes inside it. bex's MCP server calls a fixed, enumerable set of verbs instead. Here's the concrete case for why an agent that can create and destroy production infrastructure needs to know which shape it's actually operating under.
Railway's Free Tier Never Came Back: What a One-Time $5 Trial Credit Actually Buys a New Team in 2026
Railway's one-time $5 trial credit sounds generous until you run the actual per-minute billing math — here's exactly how many days it lasts for a real web app plus database, and what the same workload costs on an owned Hetzner box instead.
Railway's Reliability Problem: Five Incidents Since November 2025, and What an Outage Actually Costs a Migrating Team
Railway has logged five platform incidents since November 2025, including an 8-hour GCP account suspension. A worked cost model — SLA credits, hot-standby insurance, and the actual incident bill — for a team deciding whether to migrate.