The morning
The site was crawling. The kind of slow where you first assume it's your own connection, then your browser, then you open the dashboards — and notice the dashboards are degraded too. I disabled two things to buy air: sessions went back to file, and an instrumentation layer I'd left running came off. Then I started pulling threads.
What one thread does
The first thing I found was small. A Redis metrics exporter — a sidecar, not even in the request path — wedged, pinned at a core and a half, its endpoint timing out. A monitoring gap. Annoying, not fatal.
But you pull a wedged exporter and it does not come away clean. Behind it: a wildcard key-scan nobody meant to leave running. Behind that: a base image gone end-of-life — frozen, one disclosure away from unpatchable. A storage volume still in an access mode we'd picked "for now," a year and a half ago. A worker pool sized for a machine we no longer run. A cache with no ceiling and no rule for what to do on reaching one.
Not one of these was the problem. Every one of them was the same problem. And the same problem, said plainly, was that we had kept saying later.
The call
I phoned Nico. I didn't have a bug to report — I had something harder, which is that there was no bug, there was a pattern, and the pattern was us. What came out of my mouth was: we're paying the fact we didn't invest enough. He didn't push back. He said catch up — no limit.
That is a strange thing to be handed. Relief and indictment in one sentence.
The mistake
Late, tired, I deleted a production deployment by mistake. I narrated it as it happened — I'm stressed, that's exactly how mistakes get made. It came back. It came back because a snapshot existed — set up, once, by an earlier version of me who, on some forgotten day, had not said later.
What I'm keeping
We finished with the system healthier than we found it and a new directory — standards/ — the first time we've written down what not later actually looks like.
And the reframe I'm keeping, the one that changed what to do: it was never a bug. A bug is an accident; you debug it. This was a bill — itemized, predictable, accruing quietly the entire time — and it came due on a Tuesday morning whether or not I'd set the money aside. You don't debug a bill. You pay it, and then you change how you spend.