0082 — Retire the CI metrics store; the run page is the dashboard

Status: Accepted Date: 2026-08-09

Context

ADR-0047 bought durable CI observability with a git branch as the database: a collector workflow fired on every completed run of six tracked workflows, swept the Actions API, folded in artifact-borne records, and pushed NDJSON plus a rendered DASHBOARD.md and site/summary.json to an orphan ci-metrics branch. It worked — 953 bot commits and ~3.2 MB of records in under a month.

That is also the problem. A high-frequency bot branch sits in every git branch -a, every fetch, and every branch picker, in a repo whose day-to-day branch list is how work gets navigated. The machinery is disproportionate to its use: the questions it answers (is check getting slower, what did that bump cost) get asked rarely, and when they are asked, the run page and a local cargo clean && just timings answer them. The per-run collector, the single-writer concurrency rule, the idempotent sweep, the fork-artifact trust boundary, and the schema validation are all cost paid continuously for a lookup that happens occasionally.

Decision

Delete the store and everything that exists to feed it:

• the ci-metrics branch and the ci-metrics workflow that was its sole writer;
• the observatory workflow (weekly cold dev/release --timings builds, binary size with bloat attribution, dependency stats);
• scripts/ci/{collect-runs,render-dashboard,parse-timings,binary-size}.sh and the ci-metrics-* artifact uploads in ci.yml and stress.yml;
• just ci-report.

Keep the free half: scripts/ci/timed.sh still wraps each cargo phase and scripts/ci/summarize-job.sh still renders phase timings, cache hit/miss, target-dir size, and slowest tests into the run’s step summary. Nothing outlives the run, so summarize-job.sh no longer emits a machine-readable record and the scratch directory is target/lane-signal, not target/ci-metrics.

A hosted dashboard is not ruled out — it is moved off this repo. If https://phux.phall.io/ci returns, the site owns the collection and the storage, and phux stays a repo that builds and tests itself. Nothing in this decision blocks that; it removes the assumption that the data must first pass through a phux branch.

Why

The store’s own cost model beat it. Every tracked run paid a collector run, and the collector’s whole design — single writer, sweep-and-diff, schema-validate untrusted artifact input — exists to make a git branch behave like an append-only database it was never meant to be. Deleting it removes one workflow, four scripts, three artifact uploads, two weekly cold ARM builds, and a branch, and costs a trend line that was consulted about as often as it was regenerated.

The step summary survives because it is the part with no marginal cost: it runs inside a lane that was already running, needs no store, no token, and no second workflow, and it answers the question at the moment someone actually has it — while looking at the run.

Tradeoffs

Historical trend is gone, not archived: the ~3.2 MB of NDJSON dies with the branch, and CI wall times are only observable for as long as GitHub retains runs. Cold-build timelines, binary-size attribution, and dependency-count drift are no longer sampled on a schedule, so a slow regression across many small bumps will be noticed later, and by a human running just timings rather than by a chart. https://phux.phall.io/ci loses its data source and comes down with this change.

Alternatives

Move the store to the site’s edge (R2/KV behind the demo Worker, ingested from CI): keeps full fidelity and full history, but trades a branch for a storage binding, an ingest route, and a push secret in this repo — more machinery than the retired dashboard justified.

Rebuild the rollup from the Actions API at site build time: no store and no branch, but artifact-borne detail needs a cross-repo token, and it keeps a scheduled job alive to serve the same rarely-asked question.

Keep the branch and prune it: the branch noise in the branch list is the complaint, and pruning shards does not fix that.