A transient failure makes automatic reduction skip a run for good

Aperta
  • Qualcosa non va?
#68 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
5/5
Tempo stimato
Più di una settimana
Idoneità per principianti
25/100
Tipo di issue
Funzionalità
Chiarezza
Da chiarire
Stato di attività
Attiva
Stack tecnologico
python
Ambito
backend, data-engineering

Direzione di ricerca

Read rules.py at TriggerLoop._requests, automatic-reduction.md under “The trigger loop,” and the linked requirements question in docs/requirements/data.md. Compare the failure cases and prior art in docs/developer/prior-art/; the issue leaves three design questions open. Done requires a decided approach for distinguishing or retrying failures and dataset readiness, with the related behavior and documentation specified.

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

Descrizione

The trigger loop counts any record under a rule's label as handled, including a failed one (TriggerLoop._requests in rules.py; automatic-reduction.md, "The trigger loop"). That avoids retrying a dataset that fails at every step. But failed covers three cases that a driver cannot tell apart, since a failed record holds only str(error):

  • the workflow raised, or reported a failure;
  • the framework could not run it: a dataset could not be read, or a write failed;
  • the process died without a Python exception (out-of-memory kill, segfault), which the service must capture from the console and the batch system.

So a transient failure loses a run for good: a slow or missing /ess mount, or a dataset source that lists a run before its file is complete. Mantid's reflectometry autoprocessing hit exactly the second case in 2024 (mantidproject/mantid#36860); they fixed it by listing only runs already in the data cache, not by retrying.

Prior art, in docs/developer/prior-art/ (#67):

  • AiiDA ends a process Finished with a declared exit code, Excepted, or Killed; restart handlers key on exit codes. Transient transport failures are retried with backoff and then paused for a person, not failed.
  • Snakemake on Slurm takes OUT_OF_MEMORY, TIMEOUT, and NODE_FAIL from the batch system, not from the job.
  • Mantid never retries failed rows automatically; a user selects them and processes again.

Questions:

  1. Should a failed record say which kind of failure it was, so that an application can decide retries (relevant to #33, where rules move into applications)?
  2. Should the service retry framework failures, such as an unreadable dataset, before it fails a record (#27)?
  3. Should the dataset source promise to list a dataset only once its file is complete? This ties to the Open question in docs/requirements/data.md on how long after the file writer finishes the catalogue lists a run.
Lingua principale
Python
Stelle
0
Fork
1
Merge medio
9h 23m
PR unite (30g)
65

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 scipp/essapps

Tutte le issue di scipp/essapps

Issue simili

Altre issue su Python

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.