Zeta-Telemetry

Machine-generated telemetry and heartbeat data for Zeta.

Nothing here is human-authored, and nothing here is source. If you are looking for documentation, it is in Zeta under docs/. This repository exists so that docs/ in Zeta can go back to meaning documents a person wrote.

Why this is a separate repository

Measured on Zeta's main over 24h (2026-08-29): 101.7 MiB of new blob content, of which 100.2 MiB — 98.5% — was telemetry. Real engineering work was 1.5 MiB. File count told the same story: 54,649 files, +857/day, 80% of the new ones machine-generated.

Git cannot forget. A commit's tree is immutable, so "delete the old telemetry" inside Zeta would mean rewriting Zeta's history — a force-push that invalidates every clone and worktree, for data nobody reads after a month.

Here, history is disposable by design. Rolling data off is an ordinary commit; and if the pack ever needs to be genuinely reclaimed, this repository can be re-rooted or re-created without touching a single line of Zeta's history.

Layout — the shape IS the retention policy

by-day/YYYY/MM/DD/     MEASUREMENTS — rotates at 30 days
corpus/                THE RECORD — never rotates
schema/                what the records mean

Two roots because there are two kinds of data with opposite retention needs, and a single policy over both is wrong in one direction or the other.

There is deliberately no current/ directory. See "Reading the latest" below — an always-current file committed to git does not grow, but its history does, forever, which is the exact cost this repository exists to avoid.

by-day/ is the whole retention mechanism. Rolling off a month is:

git rm -r --quiet by-day/2026/07 && git commit -m "roll: drop 2026-07"

One directory, one commit, no per-file logic, no manifest to keep in step. This is the property being bought — and it is why a dated record must never be written anywhere except under by-day/. A file that escapes the date partition is a file the rollover cannot find.

Everything measured lands in by-day/. Nothing is overwritten in place.

corpus/ is the exception that proves the rule: it is append-only and permanent. It holds the PR review archives — review threads paired with the commits that resolved them — which are training data and are not reconstructible from the API once threads are edited or deleted. See corpus/README.md.

The counter-intuitive part, measured: the permanent corpus costs 0.43 MiB/day, while the derivable index pointing at it churned 65.35 MiB/day. The valuable half is nearly free; the expensive half is the disposable one. A retention policy that treated both as "PR data" would delete the wrong one.

Reading this data from Zeta's GitHub Page

Both hosts send access-control-allow-origin: * (verified 2026-08-29), so a static page can fetch from here directly with no token and no proxy:

Reading the latest — and why it is not a file in this repo

publish-latest.yml assembles the newest record of each kind and publishes it as a GitHub Pages artifact. Actions uploads that artifact; it is never committed, so republishing it every few minutes costs zero git history. That is the whole point: a committed current/tick-latest.json would be a new blob on every write, forever.

const res  = await fetch(
  "https://lucent-financial-group.github.io/Zeta-Telemetry/latest/tick-latest.json");
const tick = await res.json();

*.github.io sends access-control-allow-origin: * (verified 2026-08-29, 600s cache), so Zeta's Page can read it with no token and no proxy. _meta.json beside it carries generated_at and commit so a consumer can distinguish fresh from stale instead of inferring it.

Historical records are still readable straight from git when a consumer wants a specific day:

https://raw.githubusercontent.com/Lucent-Financial-Group/Zeta-Telemetry/main/by-day/2026/08/29/<record>.json

Release assets were considered and rejected for the browser path: they are also zero-history, but the redirect chain to release-assets.githubusercontent.com sends no ACAO header, so page JavaScript cannot read them cross-origin. They remain usable from CI and scripts.

This only works because this repository is public. A private one would require a credential the page cannot safely hold. That is a constraint on the design, not a preference.

Retention

by-day/ — rolling 30 days. Measurements whose value decays.

corpus/ — forever. The record. A rollover job must never be able to reach it; see corpus/README.md.

Nothing else in this repository accumulates, so nothing else needs a policy.