End-to-end reference implementation of a capacity planning system that:

1. clusters workload patterns from telemetry (heterogeneous DB fleet)
2. predicts baseline and peak compute needs via quantile boosted-tree models (P50/P95)
3. adds explainability + drift-aware retraining to reduce overprovisioning and prevent slowdowns

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

python scripts/generate_synthetic.py --days 90 --n_dbs 120
python scripts/train.py
python scripts/predict.py --input data/raw/telemetry.csv --output reports/predictions.csv
python scripts/drift_check.py --maybe_retrain

See docs/ for overview + schema + extension ideas.

MIT

See: docs/design_decisions.md

After training, generate SHAP explanations:

python scripts/explain_shap.py --quantile 0.95

Artifacts: reports/shap/

Generate plots:

python scripts/plot_metrics.py

Artifacts: reports/plots/

These images are generated after running:

python scripts/train.py
python scripts/plot_metrics.py
python scripts/explain_shap.py --quantile 0.95