Development Patterns

Config-driven development

All tunable values live in config/*.yaml. Load via the Config class:

from utils.config import Config
config = Config()

# Dot-notation access
url = config.get('data.prometheus.url')
window_size = config.get('windowing.window_size', 20)  # with default
queries = config.get('data.metrics.queries', None)

Never hardcode values that could change between environments or experiments.

Train/inference parity checklist

When modifying data processing, ensure both pipelines stay in sync:

• • scripts/train.py -- training data path
• • scripts/inference.py -- inference data path + synthetic fallback
• • src/data/preprocessor.py -- shared preprocessor (saved/loaded via joblib)
• • config/data.yaml -- shared config source

Docker workflow

# Start dev stack
docker-compose --profile dev up -d

# Rebuild after code changes
docker-compose build anomaly-detection

# Run training in container
docker-compose run --rm anomaly-detection python scripts/train.py

# Check inference logs
docker logs tv-anomaly-detector --since 5m

# Inject anomaly for testing
curl -X POST http://localhost:8000/anomaly -H "Content-Type: application/json" -d '{"type": "latency_spike", "duration": 120}'

# Stop stack (data persists in prometheus_data volume)
docker-compose --profile dev down

Testing anomaly detection

The mock service at localhost:8000 supports anomaly injection:

Note: rate()[5m] smoothing means anomalies take ~5 min to fully appear in Prometheus and ~5 min to clear after stopping.

File naming conventions

• •Models: models/lstm_autoencoder.weights.h5, models/lstm_autoencoder_config.json
• •Preprocessor: models/preprocessor.joblib
• •Threshold: models/anomaly_threshold.npy
• •Config: config/{domain}.yaml (data, model, windowing, alerting)