import ctypes import json import logging import os import pathlib import shlex import signal import subprocess import sys import warnings from pathlib import Path from mlflow import pyfunc from mlflow.exceptions import MlflowException from mlflow.models import FlavorBackend, Model, docker_utils from mlflow.models.docker_utils import PYTHON_SLIM_BASE_IMAGE, UBUNTU_BASE_IMAGE from mlflow.pyfunc import ( ENV, _extract_conda_env, _mlflow_pyfunc_backend_predict, scoring_server, ) from mlflow.tracking.artifact_utils import _download_artifact_from_uri from mlflow.utils import env_manager as em from mlflow.utils.conda import get_conda_bin_executable, get_or_create_conda_env from mlflow.utils.environment import Environment, _get_pip_install_mlflow, _PythonEnv from mlflow.utils.file_utils import ( TempDir, get_or_create_nfs_tmp_dir, get_or_create_tmp_dir, path_to_local_file_uri, ) from mlflow.utils.model_utils import _get_all_flavor_configurations from mlflow.utils.nfs_on_spark import get_nfs_cache_root_dir from mlflow.utils.os import is_windows from mlflow.utils.process import ShellCommandException, cache_return_value_per_process from mlflow.utils.string_utils import quote from mlflow.utils.virtualenv import _get_or_create_virtualenv from mlflow.version import VERSION _logger = logging.getLogger(__name__) _STDIN_SERVER_SCRIPT = Path(__file__).parent.joinpath("stdin_server.py") # Flavors that require Java to be installed in the environment JAVA_FLAVORS = {"johnsnowlabs", "h2o", "spark"} # Some flavor requires additional packages to be installed in the environment FLAVOR_SPECIFIC_APT_PACKAGES = { "lightgbm": ["libgomp1"], "paddle": ["libgomp1"], } # Directory to store loaded model inside the Docker context directory _MODEL_DIR_NAME = "model_dir" LOCAL_ENV_MANAGER_ERROR_MESSAGE = "We cannot use 'LOCAL' environment manager " "for your model configuration. Please specify a virtualenv or conda environment " "manager instead with `--env-manager` argument." def _set_mlflow_config_env(command_env, model_config): if model_config: command_env[scoring_server.SERVING_MODEL_CONFIG] = json.dumps(model_config) return command_env class PyFuncBackend(FlavorBackend): """ Flavor backend implementation for the generic python models. """ def __init__( self, config, env_manager, workers=1, install_mlflow=False, create_env_root_dir=False, env_root_dir=None, **kwargs, ): """ Args: config: Configuration for the backend. env_manager: Environment manager to use for preparing the environment. If None, MLflow will automatically pick the env manager based on the model's flavor configuration for generate_dockerfile. It can't be None for other methods. workers: Number of workers to use for serving the model. Defaults to 1. install_mlflow: Whether to install MLflow in the environment. Defaults to False. create_env_root_dir: Whether to create the environment root directory if it doesn't exist. Defaults to False. env_root_dir: Root path for conda env. If None, use Conda's default environments directory. Note if this is set, conda package cache path becomes "{env_root_dir}/conda_cache_pkgs" instead of the global package cache path, and pip package cache path becomes "{env_root_dir}/pip_cache_pkgs" instead of the global package cache path. kwargs: Additional keyword arguments to pass to the parent class. """ super().__init__(config=config, **kwargs) self._nworkers = workers or 1 if env_manager == em.CONDA and ENV not in config: warnings.warn( "Conda environment is not specified in config `env`. Using local environment." ) env_manager = em.LOCAL self._env_manager = env_manager self._install_mlflow = install_mlflow self._env_id = os.environ.get("MLFLOW_HOME", VERSION) if install_mlflow else None self._create_env_root_dir = create_env_root_dir self._env_root_dir = env_root_dir self._environment = None def prepare_env( self, model_uri, capture_output=False, pip_requirements_override=None, extra_envs=None ): if self._environment is not None: return self._environment @cache_return_value_per_process def _get_or_create_env_root_dir(should_use_nfs): if should_use_nfs: root_tmp_dir = get_or_create_nfs_tmp_dir() else: root_tmp_dir = get_or_create_tmp_dir() envs_root_dir = os.path.join(root_tmp_dir, "envs") os.makedirs(envs_root_dir, exist_ok=True) return envs_root_dir local_path = _download_artifact_from_uri(model_uri) if self._create_env_root_dir: if self._env_root_dir is not None: raise Exception("env_root_dir can not be set when create_env_root_dir=True") nfs_root_dir = get_nfs_cache_root_dir() env_root_dir = _get_or_create_env_root_dir(nfs_root_dir is not None) else: env_root_dir = self._env_root_dir if self._env_manager in {em.VIRTUALENV, em.UV}: activate_cmd = _get_or_create_virtualenv( local_path, self._env_id, env_root_dir=env_root_dir, capture_output=capture_output, pip_requirements_override=pip_requirements_override, env_manager=self._env_manager, extra_envs=extra_envs, ) self._environment = Environment(activate_cmd, extra_env=extra_envs) elif self._env_manager == em.CONDA: conda_env_path = os.path.join(local_path, _extract_conda_env(self._config[ENV])) self._environment = get_or_create_conda_env( conda_env_path, env_id=self._env_id, capture_output=capture_output, env_root_dir=env_root_dir, pip_requirements_override=pip_requirements_override, extra_envs=extra_envs, ) elif self._env_manager == em.LOCAL: raise Exception("Prepare env should not be called with local env manager!") else: raise Exception(f"Unexpected env manager value '{self._env_manager}'") if self._install_mlflow: self._environment.execute(_get_pip_install_mlflow()) else: self._environment.execute('python -c ""') return self._environment def predict( self, model_uri, input_path, output_path, content_type, pip_requirements_override=None, extra_envs=None, ): """ Generate predictions using generic python model saved with MLflow. The expected format of the input JSON is the MLflow scoring format. Return the prediction results as a JSON. """ local_path = _download_artifact_from_uri(model_uri) # NB: Absolute windows paths do not work with mlflow apis, use file uri to ensure # platform compatibility. local_uri = path_to_local_file_uri(local_path) if self._env_manager != em.LOCAL: predict_cmd = [ "python", _mlflow_pyfunc_backend_predict.__file__, "--model-uri", str(local_uri), "--content-type", shlex.quote(str(content_type)), ] if input_path: predict_cmd += ["--input-path", shlex.quote(str(input_path))] if output_path: predict_cmd += ["--output-path", shlex.quote(str(output_path))] if pip_requirements_override and self._env_manager == em.CONDA: # Conda use = instead of == for version pinning pip_requirements_override = [ pip_req.replace("==", "=") for pip_req in pip_requirements_override ] environment = self.prepare_env( local_path, pip_requirements_override=pip_requirements_override, extra_envs=extra_envs, ) try: environment.execute(" ".join(predict_cmd)) except ShellCommandException as e: raise MlflowException( f"{e}\n\nAn exception occurred while running model prediction within a " f"{self._env_manager} environment. You can find the error message " f"from the prediction subprocess by scrolling above." ) from None else: if pip_requirements_override: raise MlflowException( "`pip_requirements_override` is not supported for local env manager." "Please use conda or virtualenv instead." ) scoring_server._predict(local_uri, input_path, output_path, content_type) def serve( self, model_uri, port, host, timeout, synchronous=True, stdout=None, stderr=None, model_config=None, ): """ Serve pyfunc model locally. """ local_path = _download_artifact_from_uri(model_uri) command, command_env = scoring_server.get_cmd( local_path, port, host, timeout, self._nworkers ) _set_mlflow_config_env(command_env, model_config) if sys.platform.startswith("linux"): def setup_sigterm_on_parent_death(): """ Uses prctl to automatically send SIGTERM to the command process when its parent is dead. This handles the case when the parent is a PySpark worker process. If a user cancels the PySpark job, the worker process gets killed, regardless of PySpark daemon and worker reuse settings. We use prctl to ensure the command process receives SIGTERM after spark job cancellation. The command process itself should handle SIGTERM properly. This is a no-op on macOS because prctl is not supported. Note: When a pyspark job canceled, the UDF python process are killed by signal "SIGKILL", This case neither "atexit" nor signal handler can capture SIGKILL signal. prctl is the only way to capture SIGKILL signal. """ try: libc = ctypes.CDLL("libc.so.6") # Set the parent process death signal of the command process to SIGTERM. libc.prctl(1, signal.SIGTERM) # PR_SET_PDEATHSIG, see prctl.h except OSError as e: # TODO: find approach for supporting MacOS/Windows system which does # not support prctl. warnings.warn(f"Setup libc.prctl PR_SET_PDEATHSIG failed, error {e!r}.") else: setup_sigterm_on_parent_death = None if not is_windows(): # Add "exec" before the starting scoring server command, so that the scoring server # process replaces the bash process, otherwise the scoring server process is created # as a child process of the bash process. # Note we in `mlflow.pyfunc.spark_udf`, use prctl PR_SET_PDEATHSIG to ensure scoring # server process being killed when UDF process exit. The PR_SET_PDEATHSIG can only # send signal to the bash process, if the scoring server process is created as a # child process of the bash process, then it cannot receive the signal sent by prctl. # TODO: For Windows, there's no equivalent things of Unix shell's exec. Windows also # does not support prctl. We need to find an approach to address it. command = "exec " + command if self._env_manager != em.LOCAL: return self.prepare_env(local_path).execute( command, command_env, stdout=stdout, stderr=stderr, preexec_fn=setup_sigterm_on_parent_death, synchronous=synchronous, ) else: _logger.info("=== Running command '%s'", command) if not is_windows(): command = ["bash", "-c", command] child_proc = subprocess.Popen( command, env=command_env, preexec_fn=setup_sigterm_on_parent_death, stdout=stdout, stderr=stderr, ) if synchronous: rc = child_proc.wait() if rc != 0: raise Exception( f"Command '{command}' returned non zero return code. Return code = {rc}" ) return 0 else: return child_proc def serve_stdin( self, model_uri, stdout=None, stderr=None, model_config=None, ): local_path = _download_artifact_from_uri(model_uri) command_env = os.environ.copy() _set_mlflow_config_env(command_env, model_config) return self.prepare_env(local_path).execute( command=(f"python {quote(str(_STDIN_SERVER_SCRIPT))} --model-uri {quote(local_path)}"), command_env=command_env, stdin=subprocess.PIPE, stdout=stdout, stderr=stderr, synchronous=False, ) def can_score_model(self): if self._env_manager == em.LOCAL: # noconda => already in python and dependencies are assumed to be installed. return True conda_path = get_conda_bin_executable("conda") try: p = subprocess.Popen( [conda_path, "--version"], stdout=subprocess.PIPE, stderr=subprocess.PIPE ) _, _ = p.communicate() return p.wait() == 0 except FileNotFoundError: # Can not find conda return False def build_image( self, model_uri, image_name, install_java=False, install_mlflow=False, mlflow_home=None, base_image=None, ): with TempDir() as tmp: cwd = tmp.path() self.generate_dockerfile( model_uri=model_uri, output_dir=cwd, install_java=install_java, install_mlflow=install_mlflow, mlflow_home=mlflow_home, base_image=base_image, ) _logger.info("Building docker image with name %s", image_name) docker_utils.build_image_from_context(context_dir=cwd, image_name=image_name) def generate_dockerfile( self, model_uri, output_dir, install_java=False, install_mlflow=False, mlflow_home=None, base_image=None, ): os.makedirs(output_dir, exist_ok=True) _logger.debug("Created all folders in path", extra={"output_directory": output_dir}) if model_uri: model_cwd = os.path.join(output_dir, _MODEL_DIR_NAME) pathlib.Path(model_cwd).mkdir(parents=True, exist_ok=True) model_path = _download_artifact_from_uri(model_uri, output_path=model_cwd) base_image = base_image or self._get_base_image(model_path, install_java) env_manager = self._env_manager or em.LOCAL if base_image.startswith("python"): # we can directly use local env for python image if env_manager in [em.CONDA, em.VIRTUALENV]: # we can directly use ubuntu image for conda and virtualenv base_image = UBUNTU_BASE_IMAGE elif base_image == UBUNTU_BASE_IMAGE: env_manager = self._env_manager or em.VIRTUALENV # installing python on ubuntu image is problematic and not recommended officially # , so we recommend using conda or virtualenv instead on ubuntu image if env_manager == em.LOCAL: raise MlflowException.invalid_parameter_value(LOCAL_ENV_MANAGER_ERROR_MESSAGE) copy_src = os.path.relpath(model_path, start=output_dir) model_install_steps = self._model_installation_steps( copy_src, model_path, env_manager, install_mlflow, ) entrypoint = f"from mlflow.models import container as C; C._serve('{env_manager}')" # if no model_uri specified, user must use virtualenv or conda env based on ubuntu image else: base_image = base_image or UBUNTU_BASE_IMAGE env_manager = self._env_manager or em.VIRTUALENV if env_manager == em.LOCAL: raise MlflowException.invalid_parameter_value(LOCAL_ENV_MANAGER_ERROR_MESSAGE) model_install_steps = "" # If model_uri is not specified, dependencies are installed at runtime entrypoint = ( self._get_install_pyfunc_deps_cmd(env_manager, install_mlflow) + f" C._serve('{env_manager}')" ) dockerfile_text = docker_utils.generate_dockerfile( output_dir=output_dir, base_image=base_image, model_install_steps=model_install_steps, entrypoint=entrypoint, env_manager=env_manager, mlflow_home=mlflow_home, # always disable env creation at runtime for pyfunc disable_env_creation_at_runtime=True, ) _logger.debug("generated dockerfile at {output_dir}", extra={"dockerfile": dockerfile_text}) def _get_base_image(self, model_path: str, install_java: bool) -> str: """ Determine the base image to use for the Dockerfile. We use Python slim base image when all the following conditions are met: 1. Model URI is specified by the user 2. Model flavor does not require Java 3. Python version is specified in the model Returns: Either the Ubuntu base image or the Python slim base image. """ # Check if the model requires Java if not install_java: flavors = _get_all_flavor_configurations(model_path).keys() if java_flavors := JAVA_FLAVORS & flavors: _logger.info(f"Detected java flavors {java_flavors}, installing Java in the image") install_java = True # Use ubuntu base image if Java is required if install_java: return UBUNTU_BASE_IMAGE # Get Python version from MLmodel try: env_conf = Model.load(model_path).flavors[pyfunc.FLAVOR_NAME][pyfunc.ENV][em.VIRTUALENV] python_env_config_path = os.path.join(model_path, env_conf) python_env = _PythonEnv.from_yaml(python_env_config_path) return PYTHON_SLIM_BASE_IMAGE.format(version=python_env.python) except Exception as e: _logger.warning( f"Failed to determine Python version from {model_path}. " f"Defaulting to {UBUNTU_BASE_IMAGE}. Error: {e}" ) return UBUNTU_BASE_IMAGE def _model_installation_steps(self, copy_src, model_path, env_manager, install_mlflow): # Copy model to image if model_uri is specified steps = ( "# Copy model to image and install dependencies\n" f"COPY {copy_src} /opt/ml/model\nRUN python -c " ) steps += f'"{self._get_install_pyfunc_deps_cmd(env_manager, install_mlflow)}"' # Install flavor-specific dependencies if needed flavors = _get_all_flavor_configurations(model_path).keys() for flavor in flavors: if flavor in FLAVOR_SPECIFIC_APT_PACKAGES: packages = " ".join(FLAVOR_SPECIFIC_APT_PACKAGES[flavor]) steps += f"\nRUN apt-get install -y --no-install-recommends {packages}" return steps def _get_install_pyfunc_deps_cmd(self, env_manager: str, install_mlflow: bool): return ( "from mlflow.models import container as C; " f"C._install_pyfunc_deps('/opt/ml/model', install_mlflow={install_mlflow}, " f"env_manager='{env_manager}');" )