OOP and Software Architecture for Data Scientists

Data Science Retreat · 2 days · 16 hours

This course teaches OOP and software engineering/architecture fundamentals through the lens of data science. We are not going to build abstract toy examples, but to apply every concept to real ML problems: feature pipelines, training loops, config management, and production hardening.

---

The Through-Line

We start with a broken notebook (session 00). We end with production-grade code (session 14). Every concept in between is a tool that fixes something wrong with that original code.

---

Structure

Day 1 - Writing Code That Doesn't Rot

#	Notebook	Time	What you learn
00	Warmup: What's Wrong With This Code?	30 min	Diagnose a realistic messy notebook
01	Intro to OOP	90 min	Classes, objects, __init__, methods, properties
02	Inheritance	60 min	Subclasses, method overriding, super(), MRO
03	Docstrings and Type Hints	75 min	Google-style docstrings, type annotations, DS-specific patterns, Callable, TypeAlias, TYPE_CHECKING
04	Special Methods	60 min	__str__, __len__, __getitem__, context managers
05	Encapsulation, Abstraction & Layering	90 min	_protected attributes, ABCs, data leakage, layered architecture
06	Polymorphism, Duck Typing & Protocols	75 min	Polymorphic pipelines, duck typing, the tension with type hints, structural typing with Protocol

Day 2 - Code That Works in Production

#	Notebook	Time	What you learn
07	UML Class Diagrams	30 min	Sketch first, code second
08	Dataclasses and Pydantic	60 min	Structured config, validation, @classmethod factories
09	Project Structure	60 min	src layout, pyproject.toml, __init__.py
10	Logging	60 min	Log levels, handlers, JSON logs, loguru
11	Handling Exceptions	60 min	Custom hierarchies, propagation, when NOT to catch
12	Testing with pytest	90 min	Fixtures, parametrize, ML-specific tests, mocking
13	Capstone: Refactoring	120 min	Transform the warm-up spaghetti into production code
14	Git, DVC, and MLflow	60 min	Reproducible experiments end-to-end

---

Setup

This repo uses uv for dependency management and VS Code as the IDE.

# 1. Clone
git clone <repo-url>
cd oop_software_arch_DSR

# 2. Create the venv and install all dependencies (reads uv.lock for exact pins)
uv sync --extra dev

# 3. Open in VS Code — it will detect .venv automatically
code .

VS Code will prompt you to install the recommended extensions on first open (Jupyter, Ruff, mypy, Mermaid preview). Accept them all.

Selecting the kernel in a notebook: Open any .ipynb → click the kernel picker (top right) → choose the interpreter at .venv/bin/python (it should have the same name as the project). VS Code remembers this per-workspace.

Running tests:

uv run pytest

Adding a new dependency:

uv add <package>          # runtime
uv add --dev <package>    # dev only

Python 3.13 required (pinned in .python-version).

---

Prerequisites

• Python basics (functions, lists, dicts)
• pandas and sklearn at a "I've used them before" level
• Git basics (if not: go to learngitbranching.js.org first)

---

What's Not Here

• Deep dive into design patterns (Factory, Observer, Strategy) — that's a separate course
• Concurrency and async — out of scope for batch ML
• API deployment (FastAPI) — mentioned in the capstone as a "next step" option