venvmux 1.0.0

Run Python functions in persistent, warm subprocesses inside isolated virtual environments.

venvmux

Run Python functions in persistent, warm subprocesses inside isolated virtual environments.

Motivation

• Microservice-like boundaries for Python code on a single machine, without deploying a service.
• Cleanly separate libraries with diverging dependencies by running them in dedicated virtual environments.
• Keep worker processes warm to avoid repeated heavy imports and initialization costs.
• Simple, local orchestration: your application calls functions; workers execute them in isolated subprocesses via JSONL over stdio.
• Great for: side-by-side versions (e.g., numpy v1 and v2), incremental upgrades, and safer experimentation.

Features

• Stdlib-only core, optional uv support if installed
• Pyproject-first dependency resolution, fallbacks to requirements or inline packages
• JSONL protocol over stdio; simple Pool API with timeouts and restarts
• Context manager, async and batch calls; inflight concurrency control
• Autoscaling workers by default (workers="auto"), fixed-size workers optional

Limits and guarantees

• Trusted code only; no sandboxing or privilege dropping.
• Return values are sent as JSONL. venvmux applies a best-effort encoder so your worker functions don’t need decorators or library changes:
  • Dataclasses → dict
  • Enums → value (fallback: name)
  • datetime/date → ISO-8601 string
  • Decimal/UUID/Path → str
  • bytes → { "__bytes__": true, "b64": "..." } (base64)
  • set/tuple → list
  • Optional: numpy.ndarray → list, pandas.DataFrame/Series → dict (if installed)
  • Duck-typed: .model_dump() / .dict() / __json__() if available
  • Unknown objects → repr(obj)
• Large results: optionally spill to a temp file to avoid large JSON frames.

Install

• Requires Python 3.11+ (CPython). Supported platforms: macOS, Linux, Windows.
• Normal install:

  pip install venvmux
  

• With speed extra (optional; enables faster serialization via orjson if available):

  pip install "venvmux[speed]"
  

  • Behavior: when Pool(serializer="orjson") is set, workers attempt to use orjson; if not installed, they gracefully fall back to stdlib json.
• Editable for local dev:

  uv pip install -e .
  

  or with pip:

  python -m pip install -e .
  

• Editable with dev extras (contributors; installs pytest/ruff/mypy/pre-commit tooling):

  uv pip install -e .[dev]
  # or
  python -m pip install -e .[dev]
  

  • You can combine extras, e.g. .[dev,speed] to install both tooling and the speed extra.

Quickstart (multiple environments)

• Create worker modules that expose plain Python functions (no decorators required)

# file: yourpkg_v1/entry.py (module loaded by the worker process)
def compute(n: int) -> dict:
    return {"version": "v1", "n": n}

# file: yourpkg_v2/entry.py (module loaded by the worker process)
def compute(n: int) -> dict:
    return {"version": "v2", "n": n}

• Call those functions from your application using named environments

# file: main.py (your application)
from venvmux import EnvSpec, Pool

pool = Pool.from_envs(
    {
        "v1": (EnvSpec(workers=1), "yourpkg_v1.entry"),
        "v2": (EnvSpec(workers=2), "yourpkg_v2.entry"),
    }
)

pkg_v1 = pool.venv("v1")
pkg_v2 = pool.venv("v2")

print(pkg_v1.compute(n=1))
print(pkg_v2.compute(n=2))
pool.close()

Notes

• The worker_module (here yourpkg_v1.entry and yourpkg_v2.entry) must be importable in the worker process.
  • If your project has a pyproject.toml, venvmux will (by default) build an isolated venv and install your project there.
  • For local development, you can reuse the current interpreter: EnvSpec(python=sys.executable) and ensure your packages are on PYTHONPATH or use worker_paths.

    import sys
    from venvmux import EnvSpec
    
    EnvSpec(python=sys.executable, worker_paths=["/abs/path/to/src"])  # example
    

Lifecycle

• Use pool.close() to fully stop all environments and background threads; close() is an alias for stop().
• To stop only a single environment without tearing down the pool, call pool.stop_env("env_name").
• To gracefully restart all workers for one environment while keeping sizing, use pool.reload_env("env_name").
• pool.stop(name="env_name") is also supported to stop a specific env.

Context manager and async/batch

from venvmux import EnvSpec, Pool

with Pool.from_envs({"env": (EnvSpec(workers=2), "yourpkg_v2.entry")}) as pool:
    # async
    fut = pool.call_async("compute", {"n": 21}, env="env")
    print(fut.result())

    # batch
    results = pool.venv("env").map("compute", [{"n": i} for i in range(3)], max_workers=2)
    print(results)

Concurrency

• EnvSpec.inflight limits concurrent in-flight calls per worker process.

Workers and autoscaling

• By default, EnvSpec.workers = "auto" and the pool autoscales:
  • Starts with a small number of workers (min 1 by default).
  • Grows when all current workers are saturated (up to max_workers, default: max(2, CPU count)).
  • Scales down after inactivity (scale_down_idle_s, default: 120s) back to at least min_workers.
• Tuning knobs on EnvSpec (optional):
  • min_workers: minimum warmed workers when using auto
  • max_workers: maximum workers when using auto
  • autogrow: enable/disable growth even if workers="auto"
  • scale_down_idle_s: idle time before scaling down
  • saturation_window_len: number of recent checks before growing (default 5)
  • scale_up_cooldown_s: minimum seconds between scale-ups (default 2.0)
• Fixed-size mode: set an integer, e.g. EnvSpec(workers=4), to always run exactly 4 workers (autogrow disabled).

Autoscaling knobs (selected)

• min_workers/max_workers: bounds for auto mode.
• inflight: concurrent in-flight calls per worker process.
• autogrow: enable/disable growth when workers="auto".
• scale_down_idle_s: idle time before shrinking back down.
• saturation_window_len: number of recent checks to consider.
• scale_up_cooldown_s: minimum seconds between scale-ups.

Autoscaling: quick reference

• Growth: requires sustained saturation (recent majority of checks show all workers at inflight limit).
• Shrink: after scale_down_idle_s of inactivity, oldest workers stop until min_workers remain.

Payload size

• You can set Pool(max_payload_bytes=...) to spill large payloads to a temp file automatically.

Result size

• You can set Pool(max_result_bytes=...) to spill large results to a temp file automatically. The worker returns a small manifest; the pool reads and returns the actual data transparently.

Serializer

• Default serializer is stdlib json. For faster serialization, set Pool(serializer="orjson"). The pool passes VENV_MUX_SERIALIZER=orjson to workers; they gracefully fall back to json if orjson is unavailable.

Logging

• Control via env vars:
  • VENV_MUX_LOG=DEBUG (or INFO/WARN/ERROR)
  • VENV_MUX_LOG_JSON=1 to emit JSON logs
  • VENV_MUX_LOG_FORMAT to customize text logs

Errors

• Calls raise subclasses of CallError with a structured code and optional traceback:
  • NotFoundError: function not found in worker module
  • InvalidJsonError: worker received invalid JSON input
  • RemoteExceptionError: remote function raised; includes traceback

Protocol compatibility

• venvmux uses a simple JSONL protocol with a shared PROTOCOL_VERSION. The pool validates the worker's reported version during startup and raises a clear error if there is a mismatch.

Convenience: preload multiple environments

See Quickstart above for the preferred multi-env usage pattern.

Environment resolution

• Priority: pyproject.toml (uv.lock + uv sync if available) > requirements files > inline packages
• Venv path: ~/.venvmux/envs/{hash}/
• Hash inputs: interpreter/version, files content, inline packages, platform/arch
• Also includes PIP_INDEX_URL, PIP_EXTRA_INDEX_URL, and UV_INDEX if set

Home and cache

• Default home: ~/.venvmux (envs under ~/.venvmux/envs/{hash})
• Override via Pool(home="/path/to/.venvmux") or environment variable VENV_MUX_HOME
• Cleanup: stop the pool, then you can safely remove stale env directories under the home

Worker import paths

• If your worker module is not importable by default, set EnvSpec(worker_paths=["/abs/path/to/src"]) to extend the worker's PYTHONPATH.
• Security note: worker paths are trusted; only point to directories you control.

Public API

• EnvSpec
• Pool
• VenvHandle

Security

• For trusted code only; no sandboxing.

Troubleshooting

• If uv not available, the library uses pip.
• Set VENV_MUX_LOG=DEBUG for verbose logs.
• If using custom indices, note hashing includes index URLs; changing them creates a new env.

Tests

• Unit tests mock installers and rely on the local interpreter.

Examples

• See examples/two_envs_demo.py
• See examples/typed_remote_demo.py