Strip symbol tables from the compiled extensions in the Docker image

[Flix6x]

Description

Strips the symbol tables from the compiled extensions in the image. Most wheels ship them unstripped; nothing uses them at runtime.

Shares the changelog entry introduced by #2438 rather than adding one of its own, since both shrink the image through the builder stage. #2440 is stacked on this and widens the same entry again.

What changed

One RUN in the builder stage, after the existing rm -rf of sktime's stray payloads:

RUN command -v strip > /dev/null || { \
        echo "strip not found: the builder stage needs binutils" >&2; exit 1; \
    }; \
    find "${VIRTUAL_ENV}" \( -name '*.so' -o -name '*.so.*' \) -type f \
    -exec strip --strip-unneeded {} + 2>/dev/null || true

Placement matters. Because the runtime stage only does COPY --from=builder ${VIRTUAL_ENV}, the unstripped copies never reach a layer that ships — stripping in the runtime stage instead would grow the image, since the fat layer would still sit underneath. strip --strip-unneeded keeps everything dynamic linking needs, which is why the extensions stay loadable.

The explicit strip check came out of review: binutils only reaches this stage incidentally, as a transitive of the gcc install above. Without the check, the trailing || true would swallow strip: not found and turn the whole step into a no-op, growing the image back by ~130 MB with nothing failing and CI staying green. The || true stays on the find, where it is wanted — tolerating individual files strip cannot handle.

Measured

image	size
main (with #2438)	1.52 GB
plus this PR	1.39 GB
	−130 MB

Against main before #2438 the same step was worth 132.8 MB. The 2.8 MB difference is symbol tables in dev-only packages that --no-dev already removes, mostly mypy's mypyc extensions — the two changes overlap slightly, so the numbers should not be added.

Where it comes from, measured across all 726 shared objects belonging to locked distributions:

distribution	saved	% of its .so bytes
openturns	53.2 MB	22.6%
llvmlite	18.8 MB	11.2%
fonttools	10.0 MB	88.9%
scipy	9.3 MB	11.4%
kiwisolver, sqlalchemy, pillow, numba, numpy, cryptography, pyyaml, scikit-learn, greenlet	~28 MB	7–96%

How to test

docker build -t fm-strip .
docker run --rm fm-strip flexmeasures --version

Smoke-tested on the built image: the CLI starts (and stops at the missing SQLALCHEMY_DATABASE_URI, as expected without a database), and flexmeasures, timely_beliefs, openturns, scipy, numpy, pandas, sklearn, lightgbm, darts and statsmodels all import. openturns got the closest look, being both the biggest saver and a C++/SWIG library with a large ABI surface: Normal, UserDefined, NormalCopula, IndependentCopula, JointDistribution and Mixture all behave identically stripped.

The guard was checked on /bin/sh, the builder's shell, in both directions: with strip present it exits 0 and strips; with strip off the PATH it prints the message and exits 1.

Further Improvements

The bigger remaining items are about what is in the image rather than how it is packed: darts and its chain (shap/pyod → numba → llvmlite) is ~310 MB, openturns ~250 MB, vl-convert-python 88 MB. Ranked with their user-facing costs in #2437.

Related Items

Part of #2437. #2440 stacks on this.

---

Sign-off

• I agree to contribute to the project under Apache 2 License.
• To the best of my knowledge, the proposed patch is not based on code under GPL or another incompatible license.

🤖 Generated with Claude Code

[read-the-docs-community]

Documentation build overview

📚 flexmeasures | 🛠️ Build #34238009 | 📁 Comparing 2554df2 against latest (3272084)

  🔍 Preview build  

2 files changed

± changelog.html
± api/v3_0.html

[Copilot]

Pull request overview

This PR reduces the size of the FlexMeasures Docker image by stripping symbol tables from compiled Python extension modules during the builder stage, so only the stripped artifacts are copied into the runtime image. It also updates the changelog entry (stacked on #2438) to reflect the additional image-size reduction.

Changes:

• Strip symbol tables from all *.so / *.so.* files inside ${VIRTUAL_ENV} during the builder stage to reduce final runtime image size.
• Expand the existing Infrastructure / Support changelog entry (from #2438) to include symbol-table stripping and link both PRs.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
Dockerfile	Add builder-stage strip --strip-unneeded pass over shared objects to shrink the runtime image
documentation/changelog.rst	Expand the existing Docker image shrinking entry to include symbol-table stripping and link PR #2439

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.