Retail Ops MLOps Lakehouse (M5)

A PowerShell-first, cross-platform MLOps mini-lakehouse for M5-style retail forecasting.

Highlights

• Reproducible end-to-end runner: ensure-dirs → train → eval
• Data quality gate + validation artifacts
• LaTeX-friendly outputs (CSV + .tex, PDF + PNG)
• CI on Windows + Ubuntu (Python 3.10)

---

Quickstart

1) Create and activate a virtual environment

Windows (PowerShell)

$ErrorActionPreference = "Stop"
Set-StrictMode -Version Latest

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -U pip
pip install -e ".[dev]"

Linux/macOS (bash)

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -U pip
pip install -e ".[dev]"

---

2) Run quality gate + tests

python -m pre_commit run --all-files
python -m pytest -q

---

3) Run the pipeline (recommended)

This runs the end-to-end baseline and writes artifacts to outputs/.

.\scripts
un_m5.ps1 -Force

The runner auto-detects the venv Python and works in both Windows PowerShell and pwsh (PowerShell 7).

---

What gets generated?

After a successful run you should see:

• Models: outputs/models/ (e.g. m5_ridge_baseline.joblib)
• Tables: outputs/tables/ (CSV + LaTeX .tex)
• Figures: outputs/figures/ (PDF + PNG)
• Reports: outputs/reports/ (JSON summaries per step)

Example:

outputs/
  models/
  tables/
  figures/
  reports/

---

CLI entrypoints (optional)

The PowerShell runner calls these commands (you can run them manually too):

python -m retail_ops_mlops.cli ensure-dirs
python -m retail_ops_mlops.cli train-m5 --force
python -m retail_ops_mlops.cli eval-m5 --force

---

Repo structure (high level)

configs/              # configuration (paths, logging, etc.)
data/                 # raw/interim/processed (gitignored as appropriate)
outputs/              # generated artifacts (gitignored; .gitkeep kept)
scripts/run_m5.ps1    # PowerShell-first pipeline runner
src/retail_ops_mlops/ # library + pipelines + CLI
tests/                # smoke tests (CI)

---

Troubleshooting

“pwsh: not found” (Linux)

On Ubuntu CI runners pwsh is available by default; locally you may need to install PowerShell 7. Alternatively, run the pipeline via the CLI entrypoints instead of the PS runner.

Windows execution policy blocks activation

If Activate.ps1 is blocked:

Set-ExecutionPolicy -Scope CurrentUser RemoteSigned

Where do I configure paths?

Edit configs/default.yaml. The runner and CLI read config from there.

---

License

MIT (update if you prefer another license).