A repeatable way to move a real Azure Terraform setup to azurerm 5: match the upgrade guide to your resources, fix what validate finds, plan to zero.
Badges
Badges come from GitHub once the author claims this feed.
Recent posts
every link opens the author’s post-
Upgrading the azurerm Terraform provider from 4 to 5
-
Backups without backup infrastructure
Azure Table Storage has no backup feature, so a small Go job, a write-once container and two narrow roles do it, for under a cent a month.
-
kubectl top node says 92%. A third of the node is free.
Why kubectl top node overstates memory on AKS, what working set counts, and the two numbers that really decide evictions and refused pods.
-
After the A record, part 5: secure after launch day
Weekly grouped Dependabot updates, SHA-pinned Actions, a security.txt that keeps itself fresh, a key vault that survives a mistake, and a checklist to re-run.
-
Who is using the memory on my one node?
My single 4 GiB node looked 76% full.
-
The cost figure that went stale overnight
Follow-up to the status page: the restart fix held, but Azure throttled the cost queries one night, and the page hid that its figure was nine hours old.
-
Scaling to zero on a 4 GiB node, with KEDA
Why the AKS KEDA add-on didn't fit one small node, a first try that took 148 seconds to add a pod, and the scale-to-zero demo it turned into.
-
After the A record, part 4: the edge does the bouncing
A free-plan rate limit, a WAF rule for scanners, Turnstile and Access stop abuse at Cloudflare's edge, and why I switched Bot Fight Mode off.
-
kalinowski.io 1.0: what made it a release
Five days and 95 releases after the first version, this site is 1.0. What 1.0 meant here, what it left out on purpose, and what changes now.
-
Deploy only what changed, four days later
Follow-up to the content-tagged pipeline: main still rebuilt the website after every pull request, and idle jobs billed a minute each.
-
Roll the dice: one request, 2 to 12 hops, one trace
What one press of Roll the dice does inside my cluster: one request hopping between five Go services, timed, traced, and broken on purpose.
-
We finally have time to fix the CLI
Agents make small internal tools cheap to try, which changes what a small platform team can build.
-
After the A record, part 3: headers the browser obeys
A Content Security Policy for a static site with third-party scripts, plus the smaller headers, and how I tested them all without breaking a page.
-
How much traffic can one small node take?
A ten-minute k6 test with 100 simulated visitors against a one-node AKS site behind a Cloudflare Tunnel: nothing failed, and the node barely noticed.
-
Free tiers, and the GitHub minutes I ran out of
GitHub Actions minutes ran out eight days early with nothing failing.
-
My website on Kubernetes, two weeks later
Follow-up to the architecture post: five services instead of one, monitoring, storage and backups, the same €58 a month, and a node that's getting full.
-
After the A record, part 2: HTTPS that can't be skipped
Redirects, HSTS and a TLS 1.2 minimum make HTTPS the only way into the site.
-
The scheduler's first night: a post that stayed queued
Follow-up to the Git-and-cron publishing post: the first real scheduled post went live five hours late.
-
After the A record, part 1: the DNS records nobody mentions
DNSSEC, CAA, one canonical hostname and DMARC for subdomains: the DNS work that comes after pointing a domain at a server, in Terraform, for €0.
-
A GitOps CMS: scheduled posts with Git and a cron job
Writing posts ahead and publishing them on a date, with a folder, a nightly GitHub Actions job and the existing GitOps pipeline.
-
A public backlog from private issues, no token to rotate
Private GitHub issues feed the public backlog through a workflow and GitOps.
-
Comments on a static site, part 2: what actually happened
Building the comment service: a 403 from azurerm behind a firewall, an edge cache I didn't need, Cloudflare Access, and a spam check that blocked my own tests.
-
Comments on a static site, part 1: the plan
I want readers to comment, and the feature to improve the platform rather than bolt on a widget.
-
An AI assistant for €0 a month, and a Worker who naps
This site answers questions about my work with a small model on Workers AI's free tier.
-
A typo fix shouldn't restart my API
My pipeline redeployed every service on every commit, harmless until one kept state in memory.
-
Is this Kubernetes website actually costing me €60 a month?
The estimate said about €60 a month.