mlflow/mlflow · error · MlflowException

Databricks spark job only supports 'python' command in the e

Error message

Databricks spark job only supports 'python' command in the entry point configuration.

What it means

When a project's databricks_spark_job spec does not define python_file/parameters, MLflow computes the entry point's command and requires it to be a plain `python ...` invocation, since Databricks Spark jobs can only execute Python scripts this way. A non-python command raises this MlflowException.

Source

Thrown at mlflow/projects/databricks.py:300

        _logger.info(
            "=== Running databricks spark job of project %s on Databricks ===", project_uri
        )

        if project_spec.databricks_spark_job_spec.python_file is not None:
            if entry_point != "main" or parameters:
                _logger.warning(
                    "You configured Databricks spark job python_file and parameters within the "
                    "MLProject file's databricks_spark_job section. '--entry-point' "
                    "and '--param-list' arguments specified in the 'mlflow run' command are "
                    "ignored."
                )
            job_code_file = project_spec.databricks_spark_job_spec.python_file
            job_parameters = project_spec.databricks_spark_job_spec.parameters
        else:
            command = project_spec.get_entry_point(entry_point).compute_command(parameters, None)
            command_splits = command.split(" ")
            if command_splits[0] != "python":
                raise MlflowException(
                    "Databricks spark job only supports 'python' command in the entry point "
                    "configuration."
                )
            job_code_file = command_splits[1]
            job_parameters = command_splits[2:]

        tmp_dir = Path(get_or_create_tmp_dir())
        origin_job_code = (Path(work_dir) / job_code_file).read_text()
        job_code_filename = f"{uuid.uuid4().hex}.py"
        new_job_code_file = tmp_dir / job_code_filename

        project_dir, extracting_tar_command = _get_project_dir_and_extracting_tar_command(
            dbfs_fuse_uri
        )

        env_vars_str = json.dumps(env_vars)
        new_job_code_file.write_text(
            f"""

View on GitHub (pinned to 6a27f2decc)

Solutions

1. Rewrite the entry point command so the first token is `python` and the second token is the script file
2. Move setup steps out of the command into the Python script itself
3. Use the databricks_spark_job spec with an explicit python_file instead of relying on command computation

Example fix

// before
class MLproject
entry_points:
  main:
    command: "bash run.sh"
// after
entry_points:
  main:
    command: "python train.py --alpha {alpha}"

cmd = project.get_entry_point('main').command
assert cmd.split(' ')[0] == 'python', 'Databricks spark job entry points must start with python'

null

from mlflow.exceptions import MlflowException
try:
    run_databricks_spark_job(...)
except MlflowException as e:
    if "only supports 'python' command" in str(e):
        rewrite_entry_point_to_python()

When it happens

Trigger: Defining an MLproject entry point whose command starts with something other than `python` (e.g. `bash script.sh`, `spark-submit`, `python3 -m ...` wrappers are still fine only if the first token is `python`) and running it via `backend='databricks'` with the databricks_spark_job spec.

Common situations: MLproject written for local execution uses shell commands; entry point wraps python in a shell script; multi-token commands where the first token is not literally 'python'.

Related errors

• Running docker-based projects on Databricks is not yet suppo
• The entry point '{entry_point}' is not defined in the Databr
• Backend spec must be provided when launching MLflow project
• INVALID_PARAMETER_VALUE
• When running on Databricks, the MLflow tracking URI must be

AI-assisted analysis of mlflow/mlflow@6a27f2decc (2026-08-29). Data as JSON: /api/errors/fc5e27642cfb894d. Report an issue: GitHub.