mlflow/mlflow · error · MlflowException

Served PyFunc Model is missing server process ID.

Error message

Served PyFunc Model is missing server process ID.

What it means

`_ServedPyFuncModel.pid` returns the OS process ID of the scoring-server subprocess, stored in `_server_pid`. If the server process was never started or its PID was never recorded (None), accessing `.pid` raises MlflowException rather than returning None.

Source

Thrown at mlflow/pyfunc/__init__.py:1244

            data: Model input data.
            params: Additional parameters to pass to the model for inference.

        Returns:
            Model predictions.
        """
        if "params" in inspect.signature(self._client.invoke).parameters:
            result = self._client.invoke(data, params=params).get_predictions()
        else:
            _log_warning_if_params_not_in_predict_signature(_logger, params)
            result = self._client.invoke(data).get_predictions()
        if isinstance(result, pandas.DataFrame):
            result = result[result.columns[0]]
        return result

    @property
    def pid(self):
        if self._server_pid is None:
            raise MlflowException("Served PyFunc Model is missing server process ID.")
        return self._server_pid

    @property
    def env_manager(self):
        return self._env_manager

    @env_manager.setter
    def env_manager(self, value):
        self._env_manager = value


def _load_model_or_server(
    model_uri: str, env_manager: str, model_config: dict[str, Any] | None = None
):
    """
    Load a model with env restoration. If a non-local ``env_manager`` is specified, prepare an
    independent Python environment with the training time dependencies of the specified model
    installed and start a MLflow Model Scoring Server process with that model in that environment.

View on GitHub (pinned to 6a27f2decc)

Solutions

1. Only access `.pid` while the served model is running and healthy (e.g. after a successful predict call)
2. Check `_server_pid is not None` before reading `.pid`
3. If the server failed to launch, first diagnose the launch failure (see the 'failed to launch' error) and re-invoke load_model with env_manager
4. Re-create the served model via `mlflow.pyfunc.load_model(model_uri, env_manager='uv')`

Example fix

// before
pid = served.pid  # raises if server missing
// after
if served._server_pid is not None:
    pid = served.pid

if served_model._server_pid is None:
    raise RuntimeError("scoring server is not running; cannot get pid")

def server_is_running(served) -> bool:
    import psutil
    return served._server_pid is not None and psutil.pid_exists(served._server_pid)

try:
    pid = served.pid
except MlflowException:
    pid = None  # server not launched or already terminated
    restart_server()

When it happens

Trigger: Accessing `.pid` on a `_ServedPyFuncModel` that was constructed without a running server process — e.g. the server failed to launch, was already torn down, or the object was created directly with server_pid=None.

Common situations: Inspecting `.pid` after a failed model-server launch; accessing `.pid` on a served-model object obtained via `mlflow.pyfunc.load_model(..., env_manager=...)` after the server exited or before it started; race with server shutdown.

Related errors

• INVALID_PARAMETER_VALUE
• Unsupported input type: {type(data)}. It must be one of [str
• This container only supports models with the PyFunc flavors.
• Failed to install serving dependencies into the model enviro
• Failed to install mlflow into the model environment.

AI-assisted analysis of mlflow/mlflow@6a27f2decc (2026-08-29). Data as JSON: /api/errors/9b8196e7e952057c. Report an issue: GitHub.