What causes "Not implemented yet" in mlflow/mlflow?

Delete the files directly on the FTP server using an FTP client (ftplib/paramiko), resolving the path from the artifact URI.

mlflow/mlflow · error · MlflowException

Not implemented yet

Error message

Not implemented yet

What it means

FtpArtifactRepository.delete_artifacts is a stub: MLflow's FTP artifact store can list, upload, and download artifacts but has not implemented deletion. Calling it raises MlflowException('Not implemented yet') unconditionally.

Source

Thrown at mlflow/store/artifact/ftp_artifact_repo.py:133

                file_path = file_name if path is None else posixpath.join(path, file_name)
                full_file_path = posixpath.join(list_dir, file_name)
                if self._is_dir(ftp, full_file_path):
                    infos.append(FileInfo(file_path, True, None))
                else:
                    size = self._size(ftp, full_file_path)
                    infos.append(FileInfo(file_path, False, size))
        return infos

    def _download_file(self, remote_file_path, local_path):
        remote_full_path = (
            posixpath.join(self.path, remote_file_path) if remote_file_path else self.path
        )
        with self.get_ftp_client() as ftp:
            with open(local_path, "wb") as f:
                ftp.retrbinary("RETR " + remote_full_path, f.write)

    def delete_artifacts(self, artifact_path=None):
        raise MlflowException("Not implemented yet")

View on GitHub (pinned to 6a27f2decc)

Solutions

Delete the files directly on the FTP server using an FTP client (ftplib/paramiko), resolving the path from the artifact URI.
Switch the experiment's artifact location to a backend that supports deletion (local, S3, GCS, Azure).
Implement a custom ArtifactRepository subclass overriding delete_artifacts and register it via mlflow.register_artifact_repository.
Wrap the call and treat NotImplementedError-style cleanup as manual follow-up.

Example fix

// before
client.delete_artifacts(run_id, 'model')  # raises on ftp:// repo
// after
import ftplib
with ftplib.FTP(host, user, pwd) as ftp:
    for f in ftp.nlst(f'{base}/model'):
        ftp.delete(f)

Defensive patterns

Strategy: fallback

Validate before calling

from mlflow.store.artifact.ftp_artifact_repo import FtpArtifactRepository
repo = get_artifact_repository(client.get_run(run_id).info.artifact_uri)
if isinstance(repo, FtpArtifactRepository):
    # deletion unsupported: plan manual/FTP-client cleanup

Try / catch

from mlflow.exceptions import MlflowException
try:
    client.delete_artifacts(run_id, path)
except MlflowException as e:
    if 'Not implemented yet' in str(e):
        ftp_delete_fallback(run_id, path)  # ftplib-based cleanup
    else:
        raise

Prevention

Check the artifact store backend before designing deletion workflows
Prefer object-store artifact locations (S3/GCS/Azure) for experiments that need lifecycle cleanup
Track FTP-backed runs in a cleanup queue instead of calling delete_artifacts

When it happens

Trigger: Calling MlflowClient().delete_artifacts(run_id, path) (or repo.delete_artifacts()) when the run's artifact URI scheme is ftp:// or sftp://, i.e. the artifact repository resolves to FtpArtifactRepository.

Common situations: Cleanup scripts that delete old run artifacts working fine on local/S3 repos but failing on FTP-backed experiments; generic artifact-retention tooling that assumes delete is supported everywhere.

Related errors

AI-assisted analysis of mlflow/mlflow@6a27f2decc (2026-08-29). Data as JSON: /api/errors/ea6a8cc8a51ab101. Report an issue: GitHub.