The Week the Map Got Redrawn
The Week the Map Got Redrawn
Not every productive week produces something you can show off. Some weeks are spent redrawing the map: deciding where things belong, who may move them, and how anyone would notice if something stopped working. That was this week — unglamorous, and likely to pay for itself quietly for years.
The library finally puts itself in order
The largest finished job was a complete reorganisation of our internal documentation library. Years of good intentions had left competing top-level sections, near-duplicate branches, and paths that had drifted into odd shapes. All of it moved to one clean structure: a home for each place, each application, each operational procedure, each outside service, and the reusable templates.
That migration is complete and independently verified. The duplicate branches are gone. The library reports a consistent page count before and after, with only the planned additions appearing and nothing quietly lost. A sweep of every internal link turned up exactly one broken reference — an old spelling mismatch between two similar car pages, precisely the kind of tiny scar a real inventory leaves behind.
The useful side effect is that we now know what the library actually contains. A weekly record that used to live inside it now lives in version control, which is where a record with a date on it belongs.
Who gets a key, and who gets it alone
The second thread was identity. Single sign-on looks simple from outside and is treacherous underneath: it is easy to put every application behind one central gate, and just as easy to accidentally hand a dozen tools the same all-access pass. So the work was deliberately unglamorous — building an honest catalogue.
Every application was examined and given a verdict: does it genuinely support modern federated sign-in, does it only work behind a proxy, or is it a poor fit better left alone? That register is recorded, and it includes the uncomfortable answers. Some well-loved tools simply do not play nicely with centralised identity, and pretending otherwise would have been the expensive mistake.
Alongside it, the two systems that matter most — the secrets vault and the identity provider — were given separate, isolated homes. The crown jewels do not share a house with anything else. If the ordinary application host has a very bad day, the vault does not have that day with it.
Monitoring that reads like a sentence
Monitor names were tidied so each is a plain, human leaf name, while structure — which site, which category — now flows from the group tree into the dashboards automatically.
A synchroniser was built to walk the network inventory and reconcile every managed device into monitoring, with a dry-run default, no deletions, verified read-back after every change, and a refusal to silently overwrite an address that disagrees with what is already watched. When two sources disagree, the safe behaviour is to stop and say so. A small exporter was also finished, merged, and wired into the metrics pipeline, so the hierarchy is now structural data rather than a screenshot.
A reliable thing that wasn't
The week's most valuable finding was not an achievement at all. Two related problems surfaced from a single cause. A price-tracking service had stopped on the main container host. On its own that would be a minor annoyance — but the nightly backup preparation runs a database dump from that service, so its silence was dragging a backup job down with it. The job had been reporting trouble for two days, and the staging area it fills was empty.
A second, colder surprise: the public web tier fronting several sites was reported as stopped, its daily update and backup routines failing, while the outside world still received a perfectly normal-looking page. A page that loads is not proof that the thing behind it is healthy, and a green light is not a backup. Elsewhere the picture improved: a long-failing backup preparation job now succeeds, an off-site copy completed, and a community forum that had been dark came back.
Publishing, and a short note on speed
Support for richer formatting landed in the publishing pipeline, so longer reference pieces can finally be laid out properly instead of being flattened into paragraphs. A steady stream of work went out publicly: two substantial infrastructure essays, buying guides for local-model hardware, a lights-and-wiring build narrative, model benchmark notes, a themed resort review, and daily market briefs that run on their own schedule.
One piece of governance hardened too. Authors do not merge their own work. The weekly documentation report was written, validated and opened for review but deliberately left unmerged, precisely because the same person wrote it. A small friction that buys a large amount of trust.
What is still open, and next week
The price-tracking automation is the most interesting open item, and it is honest about why. The plumbing is built, deployed and tested, but has not yet produced one successful enrolment end to end. Its safety gates grew so strict — confirmed seller, confirmed stock, verified identity — that they have been rejecting every real candidate. The correction in flight is to simplify: an exact or close-enough match with a real price and a valid address is enough for a first version, while obvious accessories and different products still get skipped. Perfect is the enemy of done.
Next week should be a finishing week, not a starting one. Repair the backup preparation that is starved by a stopped service. Establish what is actually serving the public sites, then either repair the stopped tier or retire it deliberately. Give the market-brief routines a reproducible environment so a missing dependency stops being a recurring surprise. Prove one real enrolment in the price tracker. And let the new map settle.
The map is redrawn. Now it needs to be trusted — which only happens if the next few weeks leave it alone and let it be right.
