Replace virtualenv with python -m venv in the virtualenv env_manager path

Created by Claude Code for handoff.

Patch release: No

Summary

MLflow uses the third-party virtualenv package to create Python virtual environments when env_manager="virtualenv" is specified. The only virtualenv-specific feature used is the --python flag (to target a different Python interpreter than the one running the command). This can be replaced with python -m venv from the standard library by invoking venv directly from the target Python binary, allowing the virtualenv dependency to be removed entirely.

Note: The default env_manager is already "uv" — the virtualenv path is the legacy/fallback option. This change only affects the env_manager="virtualenv" code path.

Codebase Analysis

Core change — environment creation command

• mlflow/utils/virtualenv.py:268-280 — The env_manager == em.VIRTUALENV branch in _create_virtualenv():

  # Current: uses sys.executable to run virtualenv, --python to target a different interpreter
  env_creation_cmd = [sys.executable, "-m", "virtualenv", "--python", python_bin_path, env_dir]
  
  # New: invoke venv directly from the target Python
  env_creation_cmd = [python_bin_path, "-m", "venv", env_dir]

  This is the only place where virtualenv is actually invoked as a tool.

Validation functions to update

• mlflow/utils/virtualenv.py:68-72 — _is_virtualenv_available(): Checks shutil.which("virtualenv"). Can be removed or replaced with a check that python_bin_path -m venv works (or just removed since venv is part of stdlib).
• mlflow/utils/virtualenv.py:75-84 — _validate_virtualenv_is_available(): Error message tells users to pip install virtualenv. Should be updated or removed. Consider keeping a validation that checks the target Python has venv module available (some Debian/Ubuntu system Pythons don't ship it without python3-venv apt package). This only matters for system Python — pyenv-built Python always includes venv.
• mlflow/utils/virtualenv.py:419-421 — _get_or_create_virtualenv() calls both _validate_pyenv_is_available() and _validate_virtualenv_is_available() when env_manager == "virtualenv". Remove the virtualenv validation call.

Dependency removal

• pyproject.toml:84 — "virtualenv" is listed in the sagemaker optional extras group. Remove it.
• dev/pyproject.py:340 — "virtualenv" is listed as a dev dependency. Remove it.

Docker image update

• mlflow/models/docker_utils.py:18-39 — The SETUP_PYENV_AND_VIRTUALENV string contains RUN pip install virtualenv (line 38). Remove that line. Consider renaming the constant to SETUP_PYENV since virtualenv is no longer installed.

Files that import/reference virtualenv but DON'T need code changes

These files use function names like _get_virtualenv_name, _get_virtualenv_activate_cmd, etc. — these are internal function names that refer to the broader concept of "virtual environments" in MLflow, not the virtualenv package specifically. Renaming them is optional and not required for correctness:

• mlflow/pyfunc/__init__.py — Imports _get_virtualenv_name, uses EnvType.VIRTUALENV (this is a config key in MLmodel files, not the package name)
• mlflow/pyfunc/backend.py — Imports _get_or_create_virtualenv
• mlflow/models/container/__init__.py — Imports _get_or_create_virtualenv
• mlflow/projects/backend/local.py — Imports multiple virtualenv utility functions
• mlflow/server/jobs/utils.py — Imports virtualenv utility functions (but only uses the uv path, no change needed)
• mlflow/utils/env_manager.py:6 — VIRTUALENV = "virtualenv" constant. This is a user-facing env_manager name, not the package. No change needed.
• mlflow/utils/cli_args.py — Help text references "virtualenv" as an env_manager option. Update docs to clarify it uses venv internally.

Tests to update

• tests/pyfunc/test_virtualenv.py:26-27 — Skip condition checks _is_virtualenv_available(). Update to no longer require virtualenv binary.
• tests/pyfunc/test_virtualenv.py:18-20 — Imports _is_virtualenv_available. Update import if function is removed.
• tests/projects/test_virtualenv_projects.py — Uses _create_virtualenv spy. No functional change needed but verify tests pass.

Success Criteria

• _create_virtualenv() uses [python_bin_path, "-m", "venv", env_dir] instead of [sys.executable, "-m", "virtualenv", "--python", python_bin_path, env_dir] when env_manager="virtualenv"
• _is_virtualenv_available() and _validate_virtualenv_is_available() are removed or replaced
• virtualenv dependency removed from pyproject.toml and dev/pyproject.py
• Docker setup in docker_utils.py no longer installs virtualenv via pip
• CLI help text in cli_args.py updated to mention venv instead of virtualenv where appropriate
• Existing tests in tests/pyfunc/test_virtualenv.py and `...