python-dev-skill

Choose FastAPI for API-only services, async-first architecture, and microservices where performance matters. Choose Django when you need a built-in admin interface, full-stack templates, or batteries-included features like auth and ORM.

Use async for I/O-bound work like database queries, HTTP calls, and file operations. Use sync for CPU-bound computation, simple scripts, and legacy codebases. For mixed workloads, use an async orchestrator and offload CPU work to a thread pool.

uv is a Rust-based tool for pure Python projects offering 10-100x faster installs than pip with a pip-compatible workflow. Use pixi when you need conda packages, CUDA or GPU dependencies, or mixed conda and PyPI environments.

Use a src/ layout with pyproject.toml as the single configuration source, hatchling as the build backend, and define __all__ in __init__.py for explicit public APIs. Build with python -m build and publish via twine, testing on TestPyPI first.

Bare except clauses swallow all exceptions including system exits, making bugs invisible and debugging painful. Instead, define custom domain exception classes, raise them in service layers, and catch specific exceptions in the API layer to return structured error responses.

Yes, PyTorch Geometric supports large graphs through NeighborLoader for mini-batch neighbor sampling and GraphSAGE layers designed for inductive learning. It also supports heterogeneous graphs via HeteroData and to_hetero model conversion.