build(deps): make polars, dask and tqdm optional, drop dead dependencies

[olivier-lacroix]

Linopy currently requires some unused dependencies, and some which are used only in a select number of methods. This PR removes unused dependencies, and make some others optional

Note

The following content was generated by AI.

What

The declared dependency list had 11 entries. Instrumenting the test suite showed which ones the code actually reaches.

Dependency	Measured use	Outcome
toolz	never imported; a hard requirement of dask	removed, arrives with the dask extra
numexpr	0 calls across 8054 tests	removed
dask	never imported; reached only through xarray's .chunk()	linopy[dask]
tqdm	6 progress bars in the LP writer	linopy[progress]
deprecation	one decorator; the other 30 deprecation sites call warnings.warn	removed
polars	to_polars() and the LP writer	linopy[polars]
bottleneck	286 nanmax, 284 nanmin, 16 push calls	stays required

Core dependencies: numpy, scipy, xarray, pandas, packaging, bottleneck. pandas is declared explicitly, since 18 modules import it and it previously arrived only through xarray.

Why bottleneck stays

xarray's duck_array_ops._push imports bottleneck directly, with no numpy fallback. Variable.ffill and Variable.bfill go through it, so a core-only install raised ModuleNotFoundError from both. A core-only venv confirmed it, and the no-extras CI job now calls ffill and bfill so this cannot regress silently.

The nanmin/nanmax calls come from xarray's nanops behind .min()/.max() on NaN-carrying label and coefficient arrays, as used by Constraints.coefficientrange and the objective range.

numexpr stays out: linopy never calls pd.eval or Dataset.query, and xarray's accel extra does not list it.

How polars became optional

polars is loaded through a placeholder that raises ImportError with an install hint on first use. import linopy, model building, direct-API solves, MPS export and Variable.ffill all work without it. The default Model.solve() writes an LP file and still needs polars; to_lp_file raises a message naming linopy[polars] and pointing at io_api="direct".

get_printers_scalar drops polars for numpy, which removes polars from the direct-API solve path shared by the Highs, Gurobi, Cplex and Xpress backends.

New extras: polars, dask, progress. The dev and docs extras pull them in through a self-reference, so the extra list lives in one place.

Tests

• New tests in test/test_io.py cover the tqdm fallback, the LP guard with polars hidden from sys.modules, the placeholder, and the dunder probe that pytest's doctest collection makes.
• A new CI job installs the wheel with no extras and smoke-tests the core path.
• Generated LP files and printed variable and constraint names are byte-identical to master.

How the usage was measured, and the full suite results

A pytest plugin wrapped every public callable of bottleneck and numexpr before collection and counted calls, recording the first-call stack for each. Running it over the whole suite gave 586 bottleneck calls and 0 numexpr calls.

Stack traces pointed at:

• xarray/computation/nanops.py nanmin/nanmax, reached from constraints.py:2627 coefficientrange
• xarray/core/duck_array_ops.py:892 _push, reached from xarray/core/dataarray.py:3858 bfill

[codspeed]

Merging this PR will not alter performance

✅ 181 untouched benchmarks
⏩ 181 skipped benchmarks¹

---

_{Comparing olivier-lacroix:dep-trim (5194de8) with master (2d6aaf9)}

Footnotes

1. 181 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩