[tjduigna]

Removes the Docker build/test machinery and runs CI directly on the GitHub runner.

Why

PRs into major_refactor didn't rebuild the base image (their dockerfiles/base/* matches the target), so CI tested them on the stale published plinder-base:v0.2.27. That image carries a 2024 lockfile: numpy 1.26, biotite 1.0, cloudpathlib 0.19, an old DuckDB, gemmi/mmcif/networkx, and no peppr/voronotalt. Example: #126 failed 50 tests + 16 errors that pass on #118. The lockfile also contradicted the refactor's own pyproject.toml.

Changes

• Removed: dockerfiles/ (base/main/test Dockerfiles, env.yml, the pinned requirements.txt), docker-compose.yml, flows/docker.py, flows/proc.py (only used by docker.py).
• PR test job:
  • micromamba from environment.yml provides the conda-only tools (openstructure, mmseqs2, foldseek).
  • Then pip install -r requirements_data.txt and pip install -e ".[data,test,eval,plots]".
  • Then PLINDER_OFFLINE=true python -m pytest -n auto -v (same command the Docker test service ran), and the coverage comment.
  • No more changed-files gating; the job always runs.
• main Package workflow: same environment setup, then version bump → tests → git tag → python -m build → PyPI publish → push tag → coverage comment. The tag is still created before the build, so setuptools_scm stamps the released version as before. The image build/promote/push steps are gone.
• version.py (repo root): the bumpversion major|minor|skip logic moved out of flows/docker.py unchanged. It returns v0.2.28 on the current history, identical to the old implementation.
• pyproject.toml: duckdb>=1.5. The refactor calls DuckDBPyConnection.to_arrow_table(), which doesn't exist before 1.5 (verified 1.1.3 → 1.4.0 lack it; 1.5.2 and 1.5.6 have it). Without the floor, pip keeps an older DuckDB if one is present, which is what broke the stale image.
• environment.yml: drop reduce (unused).
• tox -e py312-test: runs pytest directly with the same extras as CI.
• Docs: removed the Docker sections from contribution/development.md and contribution/release.md.

Notes

• CI now tests the dependency resolution users actually get from pip install, rather than a separate lockfile.
• No more images are published to ghcr.io/plinder-org/plinder. Existing tags stay. flows/data_ingest_report.py still names plinder:v0.1.1 for its Metaflow step; untouched.
• The main workflow can only be fully exercised on merge to main. The version-bump script was checked locally against the old implementation.

Follow-up: uv instead of pip, no requirements_data.txt

• All dependencies come from pyproject.toml. The dev dependency group is the complete CI environment: plinder[dev,eval,loader,plots], the git-only pdb-validation and mmpdb, networkit (a wheel is required via [tool.uv] no-build-package), tabulate and semver.
• PyTorch stays its own loader extra, mapped to the CPU-only index on Linux.
• CI: micromamba installs only the conda-only tools (openstructure, mmseqs2, foldseek), then uv sync --inexact installs everything else into that env. The release workflow uses uv build.
• No committed uv.lock (gitignored): CI resolves the latest versions allowed by pyproject.toml, which is what pip users get.
• pandas 3 support: patched alignment shards keep their original large_string schema, and the query test checks missing values with isna().

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
src/plinder/data/pipeline
tasks.py					2603
Project Total

This report was generated by python-coverage-comment-action