monitors.yml

name: monitors

# Runs the operational loop -- build, row-level monitors, arrival SLAs, and
# the dry-run ops agent -- on a schedule. This is the check that fires when
# every row-level quality check is silent because nothing new landed: an
# empty delta has no nulls, no duplicates, and nothing out of range.
#
# Every step below gates on a genuine failure (a breach or a crash), never
# on which forecast model won. The scheduled build runs reduced
# (N_TRANSACTIONS=20000, for a fast run), and reduced scale starves the spend
# features: at 5,000 transactions roughly 59% of training rows have every
# spend feature NULL, and 20,000 is still far short of the 250,000 the
# committed report was produced from. Model skill is therefore not a
# meaningful signal here -- see docs/forecast-report.md's small-sample
# caveat, and note that the model does not beat its baseline at full scale
# either, so a skill gate would be permanently red. `make forecast`
# already only fails when the pipeline produced no predictions at all, not
# on model comparison, and the same rule holds here: this workflow must
# never fail because the model "didn't win".
on:
  schedule:
    - cron: "0 6 * * 1"      # weekly, Mondays at 06:00 UTC
  workflow_dispatch:

jobs:
  run-monitors:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: astral-sh/setup-uv@v3
      - run: uv sync
      - name: Build the lakehouse
        run: N_TRANSACTIONS=20000 uv run make all
      - name: Data quality monitors
        run: uv run make monitor
      - name: Data arrival SLAs
        run: uv run make arrival
      - name: Ops agent (dry-run)
        run: uv run make agent