It's one of the heaviest dependencies, and I think it's responsible for bringing in cudatoolkit, cudnn, and mkl
$ grep '"size":' ${CONDA_PREFIX}/conda-meta/*.json | sort -k3rn | column -t | sed 's|/home/marcogorelli/mambaforge/envs/pandas-dev/conda-meta||g' | head -n 20
/cudatoolkit-11.7.0-hd8887f6_10.json: "size": 871987469,
/cudnn-8.4.1.50-hed8a83a_0.json: "size": 648438607,
/pytorch-1.12.1-cuda112py38hd94e077_201.json: "size": 514342178,
/mkl-2022.1.0-h84fe81f_915.json: "size": 209326825,
/nccl-2.14.3.1-h0800d71_0.json: "size": 152235202,
/magma-2.5.4-h6103c52_2.json: "size": 93350293,
/qt-main-5.15.6-hc525480_0.json: "size": 64435513,
/gcc_impl_linux-64-10.4.0-h7ee1905_16.json: "size": 48931715,
/pillow-9.2.0-py38ha3b2c9c_2.json: "size": 47380059,
/libllvm14-14.0.6-he0ac6c6_0.json: "size": 36954351,
/sysroot_linux-64-2.12-he073ed8_15.json: "size": 32940552,
/arrow-cpp-9.0.0-py38hc370d79_10_cpu.json: "size": 32731227,
/pandoc-2.19.2-h32600fe_1.json: "size": 31408180,
/libllvm11-11.1.0-hf817b99_3.json: "size": 30536060,
/scipy-1.9.3-py38h8ce737c_2.json: "size": 27570711,
/python-3.8.13-h582c2e5_0_cpython.json: "size": 26366309,
/libdb-6.2.32-h9c3ff4c_0.json: "size": 24409456,
/mypy-0.981-py38h0a891b7_0.json: "size": 16946789,
/icu-70.1-h27087fc_0.json: "size": 14191488,
/bokeh-2.4.3-pyhd8ed1ab_3.json: "size": 13940985,
and as far as I can tell it's only used in a single ASV benchmark:
https://github.com/pandas-dev/pandas/blob/aa85f02acb7bcc1ebd05bcbf8d94a3e9f70bb95c/asv_bench/benchmarks/frame_ctor.py#L208-L222
In which case, it feels a bit wasteful to have everyone install it. We could noticeably cut down the environment size by reducing it - not all contributors have the fastest internet connections so this'd make a difference