[squid-protocol]

What this does

The CI hours bars show how each day was found. The estimates ran high, so measured and estimated days shouldn't look alike:

Grey	Meaning
Dark	Measured: running-job readings cover at least 75% of the day
Mid	The stratified sample's estimate
Light	The old small sample, least reliable

The tooltip gives the method, and the legend lists only the ones on the chart. Today's measured bar isn't faded, because a faded dark bar reads as an estimate.

A new limits panel, "Runners in use while work waits". It shows each day's median running jobs when 10+ runs are queued, against the runner limit, with the peak and the share of time at the limit in the tooltip. A long queue should fill the pool, so a line under the limit is a ceiling worth explaining.

occupancy_backfill.jsonl adds historical readings. main() reads it alongside occupancy.jsonl. It's a separate file so the live sampler can't overwrite it, and each line is marked "source": "churn-tick".

The backfill data

The file comes from churn's per-tick job counts, published separately to the dashboard branch. Against the sampler, the two sources agree: 299 paired readings, an average difference of 0.13 jobs. Ticks that read fewer runs than were in progress are left out, leaving 1,657 readings. They count GitGalaxy only, while other repos added a few runner-hours a day at most on the days we can compare.

Day	Estimate	Measured before	Measured with backfill
Oct 25	468 h	38% coverage, unused	305 h
Oct 27	332 h	224 h	229 h
Oct 28	414 h	278 h	289 h
Oct 29	400 h	72% coverage, unused	272 h
Oct 30	294 h	323 h	317 h

Oct 24 and 26 stay estimates, at 31% and 74% coverage. Nothing recorded running jobs before Oct 24, so earlier days can't be corrected and stay marked as estimates.

Median running while work waits, Oct 24 to 30: 14, 13, 15, 16, 15, 13, 14, at the limit about 1-4% of the time.

Do the tests fail when something is broken?

Sabotage	Result
Measured days treated as estimates	2 failed
Quiet readings counted as waiting	1 failed
Backfill file not read	passed at first, so I added an end-to-end test through main(); it now fails

Checks

All tests in this repo pass; the count is in the PR's pytest check. Rendered from live data and checked in Chrome.

🤖 Generated with Claude Code