How to specify Python version range in environment.yml file?

Question

Does it make sense to specify range of allowed Python versions in environment.yml file? I got this idea while reading the Google's Biq Query documentation

Supported Python Versions
Python >= 3.7, < 3.11

If this makes sense then what is the right syntax to specify the range in the environment.yml file?

Answer 1

Recommendation: Prefer exact version, not range

While there is nothing logically incorrect with specifying a version range for Python, it has the downside of defining a large solution space, which can lead to slow solving. For practical environments, I would recommend specifying the version for python through the minor version, e.g., python=3.9.

Note that this behavior mostly pertains to central packages that define variants of other packages, like python, r-base, or cudatoolkit. For most other packages, the impact is not as drastic.

Benchmarking

Here is a simple environment for basic data analysis, with and without the Python version specified.

so-py39.yaml

name: so-py39
channels:
  - conda-forge
  - nodefaults
dependencies:
  - python ==3.9
  - ipykernel
  - numba
  - pandas
  - scikit-learn
  - scipy

so-py3x.yaml

name: so-py3x
channels:
  - conda-forge
  - nodefaults
dependencies:
  - python >=3.7,<4.0
  - ipykernel
  - numba
  - pandas
  - scikit-learn
  - scipy

Conda

First, we can use the regular conda command.

Command

command time conda env create -dqn foo -f [file]

Results

Timing creating environment: so-py39.yaml
       20.51 real        19.57 user         0.94 sys
       20.73 real        19.84 user         0.97 sys
       19.43 real        18.66 user         0.95 sys
       19.22 real        18.36 user         0.92 sys
       19.34 real        18.48 user         0.94 sys
       19.08 real        18.16 user         0.94 sys

Timing creating environment: so-py3x.yaml
       30.53 real        29.56 user         1.00 sys
       29.21 real        28.21 user         1.08 sys
       31.13 real        29.77 user         1.07 sys
       29.93 real        28.46 user         0.99 sys
       30.53 real        29.43 user         0.98 sys
       28.60 real        27.68 user         1.03 sys

That is, solving for the environment with a range takes ~10s (~50%) longer.

Mamba

We can also test solving the environment with Mamba.

Command

command time mamba env create -dqn foo -f [file]

Results

Timing creating environment: so-py39.yaml
        3.30 real         2.79 user         0.49 sys
        3.36 real         2.84 user         0.51 sys
        3.25 real         2.74 user         0.49 sys
        3.34 real         2.82 user         0.51 sys
        3.29 real         2.78 user         0.51 sys
        3.24 real         2.74 user         0.48 sys

Timing creating environment: so-py3x.yaml
        3.27 real         2.79 user         0.47 sys
        3.26 real         2.78 user         0.46 sys
        3.33 real         2.83 user         0.48 sys
        3.28 real         2.79 user         0.47 sys
        3.31 real         2.81 user         0.49 sys
        3.29 real         2.81 user         0.47 sys

This indicates that when using Mamba the difference in solve time is neglible.