1732 posts tagged with "Self-Hosting"
Running your own PaaS and infrastructure on machines you own
Railway vs Render vs Fly.io Re-Benchmarked: What the July 2026 Numbers and Render's New Billing Lines Mean for the Self-Host Math
July 2026's three-way PaaS benchmarks still hold in October — but Render's metered bandwidth and per-domain billing lines redraw the bill: worked team math, multi-tenant domain costs, and the self-hosted column that beats every hosted row.
Railway vs Render vs Fly.io for Solo Developers in 2026: The Hidden-Cost Anatomy
Two side projects — one always-on, one bursty — priced line by line on Railway, Render, and Fly.io: the bandwidth, IPv4, and volume lines that flip the ranking, and the workload count where a flat-rate Hetzner box undercuts all three.
Your PaaS Bumped the Default Runtime Again: Render's Bun 1.4.3 and the Case for Pinning Build Images You Control
Render's default Bun version moved three times in three months — 1.3.14 in July, 1.4.2 in September, 1.4.3 on October 12 — while existing services keep the version from their creation date. Why that forks unpinned fleets across a real behavioral boundary, and how to pin the runtime in four places.
Render's $100/mo Dedicated Outbound IPs vs a $14 Self-Hosted NAT Tier: What Stable Egress Identity Really Costs
Render's May 2026 dedicated outbound IPs cost $100/month for three static addresses. This teardown prices the self-hosted equivalent — an HA NAT gateway pair with floating IPs and per-tenant SNAT at ~$14/month in infrastructure — then adds the fully-loaded labor math, the honest failover story, and a decision guide for enterprise migrators.
Rivet Wants to Be Kubernetes 2.0: What a PaaS Without Kubernetes Gives Up on Machine Two
Rivet pitches itself as Kubernetes 2.0 — a trivially self-hostable orchestrator with a small core. A concrete inventory of what it drops (machine provisioning, upgrades, networking, bin-packing) versus what a Cluster API fleet already solves, and why a self-hosted PaaS still needs Kubernetes once the second machine arrives.
Rootless Kubelet Hits Beta in Kubernetes 1.37: What Unprivileged Nodes Buy a Multi-Tenant Fleet
Kubernetes 1.37 promotes rootless nodes to beta, so the kubelet and runtimes can run unprivileged and cap five years of container breakouts — but only on cgroup v2 with local storage and a short CNI list. Here is the evidence and the adoption checklist.
Self-Hosted Search TCO in 2026: Algolia's Per-Query Meter vs Typesense or Meilisearch on a Hetzner Box
A 250,000-SKU store doing 1M searches a month pays Algolia about $555 a month — or $82–145K over three years self-hosted on Hetzner instead of $192–360K. This teardown compares the meter against Typesense and Meilisearch, prices the engineering labor both sides hide, and shows where Postgres full-text search wins outright.
Self-Hosting gpt-oss-20B on a Single 12GB GPU: What OpenAI's Open Weights Cost to Serve vs the API
Serving gpt-oss-20B on a single 12GB GPU costs about $30 a month all-in and breaks even against o3-mini-class API pricing near 15 million tokens — the full VRAM math and worked self-host vs API costing for both model sizes.
Talos Linux Grows a Hypervisor: VMs Without KubeVirt, Containers Without Kubernetes
Sidero Labs is building a native hypervisor and Kubernetes-less containers into Talos Linux, with an alpha demo at TalosCon 2026 and GA in December — here is when to choose Talos VMs over KubeVirt, what single-node edge containers are good for, and what it means for a Cluster-API fleet.