Files
go-weatherstation/docs/rain_model_runbook.md
2026-03-12 19:55:51 +11:00

142 lines
3.5 KiB
Markdown

# Rain Model Runbook
Operational guide for training, evaluating, deploying, monitoring, and rolling back the rain model.
## 1) One-time Setup
Apply monitoring views:
```sh
docker compose exec -T timescaledb \
psql -U postgres -d micrometeo \
-f /docker-entrypoint-initdb.d/002_rain_monitoring_views.sql
```
## 2) Train + Evaluate
Recommended evaluation run (includes validation-only tuning, calibration comparison, naive baselines, and walk-forward folds):
```sh
python scripts/train_rain_model.py \
--site "home" \
--start "2026-02-01T00:00:00Z" \
--end "2026-03-03T23:55:00Z" \
--feature-set "extended" \
--model-family "auto" \
--forecast-model "ecmwf" \
--tune-hyperparameters \
--max-hyperparam-trials 12 \
--calibration-methods "none,sigmoid,isotonic" \
--walk-forward-folds 4 \
--model-version "rain-auto-v1-extended" \
--out "models/rain_model.pkl" \
--report-out "models/rain_model_report.json" \
--model-card-out "models/model_card_{model_version}.md" \
--dataset-out "models/datasets/rain_dataset_{model_version}_{feature_set}.csv"
```
Review in report:
- `candidate_models[*].hyperparameter_tuning`
- `candidate_models[*].calibration_comparison`
- `naive_baselines_test`
- `walk_forward_backtest`
## 3) Deploy
1. Promote the selected artifact path to the inference worker (`RAIN_MODEL_PATH` or CLI `--model-path`).
2. Run one dry-run inference:
```sh
python scripts/predict_rain_model.py \
--site home \
--model-path "models/rain_model.pkl" \
--model-name "rain_next_1h" \
--dry-run
```
3. Run live inference:
```sh
python scripts/predict_rain_model.py \
--site home \
--model-path "models/rain_model.pkl" \
--model-name "rain_next_1h"
```
## 4) Rollback
1. Identify the last known-good model artifact in `models/`.
2. Point deployment to that artifact (worker env `RAIN_MODEL_PATH` or manual inference path).
3. Re-run inference command and verify writes in `predictions_rain_1h`.
4. Keep the failed artifact/report for postmortem.
## 5) Monitoring
### Feature drift
```sql
SELECT *
FROM rain_feature_drift_daily
WHERE site = 'home'
ORDER BY day DESC
LIMIT 30;
```
Alert heuristic: any absolute z-score > 3 for 2+ consecutive days.
### Prediction drift
```sql
SELECT *
FROM rain_prediction_drift_daily
WHERE site = 'home'
ORDER BY day DESC
LIMIT 30;
```
Alert heuristic: `predicted_positive_rate` shifts by > 2x relative to trailing 14-day median.
### Calibration/performance drift
```sql
SELECT *
FROM rain_calibration_drift_daily
WHERE site = 'home'
ORDER BY day DESC
LIMIT 30;
```
Alert heuristic: sustained Brier-score increase > 25% from trailing 30-day average.
## 6) Pipeline Failure Alerts
Use the health-check script in cron, systemd timer, or your alerting scheduler:
```sh
python scripts/check_rain_pipeline_health.py \
--site home \
--model-name rain_next_1h \
--max-ws90-age 20m \
--max-baro-age 30m \
--max-forecast-age 3h \
--max-prediction-age 30m \
--max-pending-eval-age 3h \
--max-pending-eval-rows 200
```
The script exits non-zero on failure, so it can directly drive alerting.
## 7) Continuous Worker Defaults
`docker-compose.yml` provides these controls for `rainml`:
- `RAIN_TUNE_HYPERPARAMETERS`
- `RAIN_MAX_HYPERPARAM_TRIALS`
- `RAIN_CALIBRATION_METHODS`
- `RAIN_WALK_FORWARD_FOLDS`
- `RAIN_ALLOW_EMPTY_DATA`
- `RAIN_MODEL_CARD_PATH`
Recommended production defaults:
- Enable tuning daily or weekly (`RAIN_TUNE_HYPERPARAMETERS=true`)
- Keep walk-forward folds `0` in continuous mode, run fold backtests in scheduled evaluation jobs