Loading page...Загружаем страницу...正在加载页面...
AtlasRepo

It Ran Every Morning and Still Broke: Failure Modes of a Free-Tier AI Job

Описание и применение

Автоматический ежедневный cron‑задание, которое собирает релиз‑ноты, обрабатывает их моделью ИИ и сохраняет сводку, часто завершалось без ошибок, но выдавалось неверное содержимое. В статье описаны пять типичных «плохих» сценариев и предложен скрипт‑проверка, который валидирует размер, схему, логику леджера, состояние и предупреждения об усечении.

Задача

A daily cron job that fetches release notes, summarizes them with an AI model, and writes the summary failed silently on a free tier, producing incorrect or empty outputs without raising errors.

Результат

The healthcheck script catches five common failure modes of free-tier AI jobs, turning silent failures into actionable alerts and ensuring reliable daily summaries.

Рабочий процесс

  • 1. Verify latest_summary.md is non-empty and >50 chars.
  • 2. Check that summary contains expected sections: Breaking, Migration, Affects.
  • 3. Confirm ledger.jsonl has a recent entry with today's date.
  • 4. Ensure state file (last_seen.json) matches the ledger's tag.
  • 5. Scan watch.log for any truncation warnings.

Доказательства

The cron log said exit 0 every morning. The summary file updated on schedule. The job was running. The output was wrong.

Источник

The fix was to validate the response body against an expected shape, not just the status code.

Источник

The healthcheck script verifies output size, schema, ledger entry, state consistency, and truncation warnings.

Источник

  • Дата записи: 2026-09-07T19:08:03.100Z