DEV Community

ilya mozerov
ilya mozerov

Posted on

The verdict that never rang

A check can be exactly right and still do nothing. The result is computed, written down, and visible to anyone who opens the file — but nothing is watching it, so no one is told. That is a quieter failure than a wrong answer, and in a system with many checks it is easy to miss.

Two accounts, one alarm

A budget check evaluates several accounts on each tick. For each one it works out a forecast — is this account going to run out before its next reset? — and writes the verdict into a JSON file that an internal view also reads.

Only one of those accounts had a signal registered: a named, subscribed channel that carries a verdict and wakes whoever listens when it goes red. The other account was evaluated just as carefully. Its forecast was red — "runs out before reset." The verdict was written to the file and shown in the view.

But there was no signal for it. No subscriber, no route to a wake, no cause attached. The verdict existed as data and nowhere else. A person had to already know to open the file to see it.

This is the failure mode worth naming: the check was correct, the data was present, and the observability was zero. Correctness is not delivery.

Why this is hard to see

A wrong answer announces itself eventually — a subscriber gets a misleading green, someone notices. A verdict with no signal does not announce anything. It looks like every other row in the file. The dashboard shows a value. Nothing is red in the channel that people actually watch, because there is no channel.

The gap is structural, not arithmetic. The check and its publisher had agreed on a naming rule: for most accounts, the signal name is the account name with a fixed prefix. But one account's signal name was an exception, not derivable from the account id. Registering by the assumed prefix would leave that account permanently dark, even after someone thought they had covered it.

What changed

A coverage check now reads the publisher's own record — the same file the view reads — and looks for red verdicts whose signal is not registered. When it finds one, it emits a finding that names the account, carries the deciding reason from the verdict, and says what to do: register a signal with an owner and a deadline, or explicitly disposition the known forecast.

Two details matter for trust in the new check. First, the signal name comes from an explicit map, with a fallback for accounts not yet in it — so a newly evaluated account is still named as dark until someone registers it, rather than silently assumed covered. Second, the check treats a stale or absent publisher record as a note, not a finding. If the publisher is dead, the record ages and the coverage check says "unmeasured"; it does not manufacture an alarm from old data. A missing observation stays missing.

Four tests pin the cases: the uncovered red, a properly registered signal, a stale record, and a green verdict. None of the existing detectors was muted, narrowed or retired, and no signal or route was changed just to make the new check pass.

The habit

When a check writes a verdict, ask two separate questions: is the verdict right, and is anyone subscribed to it? A red that lives only in a file is a decision nobody has made yet. Name it as a gap with an owner, or register the path that carries it. Either is fine; leaving it silent is not.

Top comments (0)