Choosing a Prebuilt Image

​

Trust: three tiers

​

Python image variants

​

The Alpine trap

• Prebuilt binary wheels on PyPI (manylinux, targeting glibc) fall back to source builds on Alpine, requiring gcc / musl-dev and each package’s C headers in the image, so build time soars and the final image may not be smaller.
• numpy, scipy, pandas, PyTorch, TensorFlow and similar scientific / ML packages need self-compiling or unofficial packages on Alpine: not recommended.
• Alpine suits pure Python (no C-extension deps), a hard size requirement, and willingness to accept the compile cost in a multi-stage build.

​

Rules of thumb for picking a version

Need scientific packages (numpy / pandas / scipy / torch)?
  → Yes → python:3.X (full) or python:3.X-slim + manually install build-essential
  → No  → python:3.X-slim (default)
          └→ Hard size requirement + confirmed no C extensions?
                → Yes → python:3.X-alpine (accept self-compiling)

• Pin a major.minor tag (python:3.12-slim), not latest or python:3.
• Default to slim, adding build tools yourself when compiling native packages.
• bookworm = Debian 12, bullseye = Debian 11; use a codename tag (python:3.12-slim-bookworm) to lock the Debian version.

​

Next

• A Python app Dockerfile: writing a real Dockerfile with slim.
• Layer cache and best practices: pinning tags, non-root, and more.