Skip to content
Trazum

1.56.0

Something that runs

The arc's question was whether alerting can be given without becoming a hosted service holding other teams' metrics. The answer is yes for the noticing, and no for the last hop — and both halves are now written down.

What shipped is not a runtime. It is three things that make the absence of a run visible: history names the calendar stretches no report covers, so a series with a hole in it stops reading like a shorter one; trazum pulse gives the outside view of a scheduled job, because the thing that would tell you a watcher stopped was the watcher; and docs/running.md is the reasoning, the recipes and the place the answer runs out.

Added#

docs/running.md — something that runs, and why it is not us. The arc asked whether alerting can be given without becoming a hosted service holding other teams' metrics. The answer is yes for the noticing, and this page is that answer with its reasoning, its recipes and the place it runs out.

What has to run and why the four jobs are separate — a single "run everything" command would hide which of them stopped. Recipes for cron, systemd timers, GitHub Actions and Windows Task Scheduler, each one a command this repository actually accepts. Where the credential lives on each platform, and why Trazum adds no new place for a key to sit.

And a section on where the answer runs out, because a page that only listed what works would be advertising: a laptop is not a scheduler; the last hop is still yours and Trazum cannot page you or know you are on holiday; a watcher can only judge what has been pulled; and nothing here watches the watcher's watcher, because the chain has to end at the thing you already trust to tell you when it breaks.

trazum pulse — did the things that are supposed to run, run? watch --once is built for a scheduler: a cron entry is the whole daemon, and its state file records each cycle precisely so a restart is honest about the stretch it did not watch. That file was read by exactly one thing, and that thing was the next cycle.

So nothing could tell you the watcher had stopped, because the thing that would tell you was the thing that stopped. A dead cron produces silence, and a watcher with nothing to report produces silence too. pulse is the outside view: the age of the last watch cycle, the age of the last pull into the store, and how far the stored measurements reach — with --max-stale-hours <n> to turn a run that stopped into a failing exit code.

It runs nothing and hosts nothing, which is the arc's question answered rather than dodged: something has to notice, and the something is already in your CI. A step running this on the schedule you already have turns a dead cron into a red build without Trazum holding anybody's metrics.

Three refusals: a first run that never happened is not late, because there is no cadence to be late against, so never-run never gates. Without a stated threshold nothing is judged at all — how stale is too stale is a policy. And how far the measurements reach is never judged by the same threshold: a store pulled ten minutes ago whose newest record stops two days back is a healthy cron in front of a provider that reports late.

And a whole class of flag defect is now guarded. --max-stale-hours 36 shipped, built, ran, printed a full report and gated on nothing: the flag was not in VALUE_FLAGS, so it parsed as a boolean and 36 became a positional argument. Nothing failed anywhere. A test now derives the rule from the source — anything read with stringFlag or numberFlag takes a value by definition — and it is proved against the exact line that shipped.

A series with a hole in it is not a shorter series, and history now says which it is looking at. The first chapter of arc 1.56 — something that runs — and it turns out most of that arc is not a runtime at all. A scheduled job that stopped three weeks ago does not announce itself; it just stops adding points, and a series missing its last three weeks and a series that only ever covered a fortnight are the same document until somebody states the difference.

unmeasured[] names every stretch of calendar time between the first report's start and the last one's end that no report covers — with the days, the instants, and the reports either side. It is arithmetic on the spans and never an inference about a schedule: this module has no idea how often anybody meant to run anything, and guessing a cadence in order to call a gap late would be the tool deciding what somebody's routine is.

And the caveat travels on the finding. A run is consecutive reports and read as consecutive time: four rising periods with an unmeasured fortnight between the second and the third was reported as "climbing for four periods", when the climb may have reversed and come back inside the hole. Each run now carries unmeasuredDays, and the terminal prints it on the run's own line — a caveat one section away arrives after the reader has already formed the sentence.

overlappingReports[] names two reports covering some of the same days. history never sums across periods, but a reader with the document in a spreadsheet will, and two exports over the same fortnight count it twice. Named and never merged: which of the two is the better measurement is not knowable from here.

A gap shorter than a whole day is not reported — that is the seam between two adjacent exports, not a stretch nobody measured — and the day count is floored rather than rounded, so a gap this tool calls three days is at least three days.