Labelled corpus of 90 days of ci.yml failures, as evidence for #2177

Aperta
#2,647 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 2 giorni

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
5/5
Tempo stimato
Più di una settimana
Idoneità per principianti
35/100
Tipo di issue
Funzionalità
Chiarezza
Abbastanza chiara
Stato di attività
Attiva
Stack tecnologico
github, github-actions, rust

Direzione di ricerca

Read the linked corpus repository, especially tool/harvest.sh, tool/classify.py, tool/label.py, and tool/labels.yaml, then review #2177 and #2523 for project context. The issue asks whether to build a project hosted dashboard that refreshes the analysis; confirm the project's interest and agree on scope before implementation. Done means an accepted proposal defining where it runs, what it displays, and how the scheduled refresh works.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

For the flake tracking in #2177 I wanted to know how much of the CI failure volume is flaky, and from which causes. So I classified every failed ci.yml job from 13 August to 10 November 2026: 2638 failed jobs across 890 runs, grouped by log signature and labelled flake, real or unclear.

The corpus, the per-job labels and the scripts are here: https://github.com/andrewdunndev/bootc-ci-corpus

824 of those jobs are the two required-checks gates, which mirror failures in other jobs. The other 1814 split like this:

Label Jobs Share Largest causes (jobs)
real 1206 66% grub2-efi-cc absent from the ELN repos (282); PR-specific test failures (225); one PR's VM launch failures (176)
flake 530 29% quay blob EOF and 502/504 (134); dnf mirror 404 or timeout (82); guest reboot lost in tmt plans (66); runner lost (32)
unclear 78 4% not enough in the log to say
  • Labels are rule-assigned. The rules are in tool/labels.yaml: a default label per signature, sub-cause rules that match the log tail, and one override. A non-gate job whose same sha passed on another attempt is labelled flake (261 of the 530 flakes).
  • The grub-cc window and that one PR inflate the real share. Without them, real is 55% of 1356 jobs.
  • A libvirt session socket failure at VM launch is still occurring (5 jobs from 1 to 10 November, 24 in all) and isn't in #2177. Neither are the quay blob EOF failures (#2177 has the 502 case).

What it's for

Evidence for #2177: which flakes cost the most jobs, which are fixed, and which still occur. labels.jsonl also works as an answer key for the ci-triage analyzer, and for the post-merge triage in #2523.

How it was labelled

AI agents wrote the rules from an earlier hand-labelled pass. Every label comes from labels.yaml, and the reason field in labels.jsonl says what set it: the signature default, a log-tail or PR rule, or the same-sha override. Two hand sub-causes have no rule and stay at signature level; corpus.md lists them.

Rerun or extend

tool/harvest.sh . raw work   # needs gh, jq, python3; GET requests only
python3 -I tool/classify.py jobs.jsonl tool/signatures.yaml assignments.jsonl
python3 -I tool/label.py jobs.jsonl assignments.jsonl tool/labels.yaml labels.jsonl corpus.csv corpus.md

Same inputs give byte-identical outputs. The harvest keeps the last 150 lines of each job log; those tails are not in the repository.

Comments wanted

  • Rules that look wrong, most likely in vm-launch-failed and tmt-plans-failed-uncategorised, where one signature holds both flake and real jobs.
  • Whether the libvirt socket and quay blob EOF failures belong in #2177.
  • Whether per-job signatures are the right key for scoring ci-triage.
  • Known gaps: cancelled jobs (141) aren't harvested.

@cgwalters, would this kind of analysis be a welcome contribution as part of the project itself, on a dashboard that updates on its own? The harvest and labels already rerun from scratch, so a scheduled job could keep the numbers behind #2177 current. I'd be glad to build it if so, or keep it outside the project if you'd rather.

Lingua principale
Rust
Stelle
2.3k
Fork
230
Merge medio
2g 11h
PR unite (30g)
42

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di bootc-dev/bootc

Tutte le issue di bootc-dev/bootc

Issue simili

Altre issue su Rust

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.