Engineering deep dives, building in public, Bitcoin sovereignty, and first-principles thinking.
We switched Datadog off and saved $40K a month. Then a $40K/month NAT gateway charge appeared in EC2-Other, and nobody noticed for one to two months. The second-order cost of self-hosting your telemetry.
I run Talos — immutable, API-managed Kubernetes OS — on bare metal. The install, the no-SSH model, the upgrade scheme, and what's genuinely painful.
Soft multi-tenancy on Kubernetes has a ceiling, and it arrives sooner than most teams expect. What the rungs above it actually are — per-tenant network separation, dedicated VMs, child clusters — and what each one really costs to run.
A line-by-line method for reading an observability bill: the four places waste hides, how to tell a decision from an accretion, and when the honest answer is to leave it alone.
Five failures found on my own infrastructure in one audit. Every one of them was reporting success. The common thread is that green status is a proxy, and proxies fail quietly.
Running multiple tenants on one Kubernetes cluster without a big ops team: namespaces, quotas, NetworkPolicy, RBAC, and when vCluster earns its weight.
A mini-PC home lab beats cloud on cost — payback in under 6 months, then a coffee a month to run. The honest home-lab-vs-cloud math, with sources.
The same principle that says own your compute says rent your durability. One rule, two opposite answers — and why a rented durability layer is the right home for your backups and the wrong shape for your steady workloads.
Bare metal Kubernetes can be 5x cheaper than AWS — but it's wrong for most teams. I run 6 Hetzner nodes; here's the honest tradeoff and decision framework.
Observability cost $80K/month at 1TB/day — $40K Datadog, $40K CloudWatch. A year later at 6× the data, the self-hosted LGTM stack I built ran $25K — vs a ~$225K SaaS projection. The teardown.
A real observability architecture at scale: the LGTM stack moving ~6 TB of telemetry a day across 15 departments — microservices mode, sharding, S3 tiering, retention, and cardinality control.
EKS, GKE and AKS charge ~$73/cluster/month for a control plane you could run yourself for free. Here is what that money actually buys, and when you should pay it.
The realistic minimum Grafana LGTM stack — Loki, Grafana, Tempo, Mimir, Alloy on one S3 bucket — for a small team that can't afford Datadog and doesn't need an enterprise build.
Cloud cost management isn't an accounting line at month-end. Treat the bill as a product metric — cost per request, per tenant — beside latency.
S3 cost optimization, step by step: storage classes, lifecycle policies, Intelligent-Tiering, multipart cleanup, retrieval fees, and the free VPC endpoint.
A decision guide to AWS Savings Plans vs Reserved Instances — when each wins, what to actually buy, and what the December 2025 Database Savings Plans launch changes.
AWS NAT Gateway pricing charges you twice — to exist and to move your bytes. How to spot the cost on your bill and the fixes that cut it to near zero.
A practitioner's cloud cost optimization audit for AWS, in order: data transfer, rightsizing, commitments, storage. Where the first 20-30% hides on a $50K bill.
Cloud cost optimization in four decisions: cut the waste, own past your break-even, observe cheaply, and don't over-staff. The whole method, with the receipts.
The Datadog alternative that ran 6 TB/day of observability for ~$25K/mo vs a ~$225K Datadog + CloudWatch projection. How the bill ballooned, what we changed, and the ~89% cut.
Six dedicated servers in a German data centre, and a multi-tenant Kubernetes platform on top of them. Most of what I run my life on now lives there. This is an honest account of why I built it, what it actually costs, and the situations in which you absolutely should not do the same.
A reusable framework for the build/buy/self-host call — real TCO including ops time, switching cost, core-vs-context, and the honest failure mode of each.
Python on Knative was costing me 5–7 seconds every cold start. On a small Kubernetes cluster, that latency was the difference between a usable AI agent and a broken one. So I rewrote four services in Go on three weekends. Here are the measured numbers, the unexpected wins, the unexpected pains, and when you absolutely should not do the same.
Everything you need to know about reducing your AWS bill by 20-40% — from quick wins to architectural changes.
Datadog was costing us $40K/month at 1 TB/day. We were growing to 6 TB. Instead of scaling the bill, we built our own platform that ran ~9× cheaper at production scale.
