microsoft/qlib · error · ValueError

Please make sure the recorder has been created and started p

Error message

Please make sure the recorder has been created and started properly before getting artifact uri.

What it means

Raised by Recorder.local_path when self.artifact_uri is None. The artifact URI is only populated after the recorder (an MLflow run) has been created and started; calling local_path on a recorder that was never started leaves artifact_uri None, so there is nothing to resolve into a local directory path.

Source

Thrown at qlib/workflow/recorder.py:331

        return self._artifact_uri

    def get_local_dir(self):
        """
        This function will return the directory path of this recorder.
        """
        if self.artifact_uri is not None:
            if platform.system() == "Windows":
                local_dir_path = Path(self.artifact_uri.lstrip("file:").lstrip("/")).parent
            else:
                local_dir_path = Path(self.artifact_uri.lstrip("file:")).parent
            local_dir_path = str(local_dir_path.resolve())
            if os.path.isdir(local_dir_path):
                return local_dir_path
            else:
                raise RuntimeError("This recorder is not saved in the local file system.")

        else:
            raise ValueError(
                "Please make sure the recorder has been created and started properly before getting artifact uri."
            )

    def start_run(self):
        # set the tracking uri
        mlflow.set_tracking_uri(self.uri)
        # start the run
        run = mlflow.start_run(self.id, self.experiment_id, self.name)
        # save the run id and artifact_uri
        self.id = run.info.run_id
        self._artifact_uri = run.info.artifact_uri
        self.start_time = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
        self.status = Recorder.STATUS_R
        logger.info(f"Recorder {self.id} starts running under Experiment {self.experiment_id} ...")

        # NOTE: making logging async.
        # - This may cause delay when uploading results
        # - The logging time may not be accurate

View on GitHub (pinned to 79633dd950)

Solutions

Ensure the recorder is started before touching artifacts: use exp.get_recorder() only after R.start() and a successful run, or call recorder.start_run() first.
Check recorder.artifact_uri is not None before calling local_path; if None, re-create the recorder via R.get_exp().create_recorder(...) + start_run().
If the run was created through qlib.workflow.R, verify the MLflow tracking URI is reachable so start_run() actually persisted the run.
If you truly have a valid run id, reload the recorder from the tracking store (e.g. MLflow client.get_run(id).info.artifact_uri) instead of using the un-started object.

Example fix

// before
recorder = exp.get_recorder()
path = recorder.local_path  # ValueError if run never started

// after
recorder = exp.get_recorder()
if recorder.artifact_uri is None:
    recorder.start_run()
path = recorder.local_path

Defensive patterns

Strategy: validation

Validate before calling

if recorder.artifact_uri is None:
    raise RuntimeError(f"Recorder {recorder.id} not started; artifact_uri is None")
path = recorder.local_path

Try / catch

try:
    path = recorder.local_path
except ValueError as e:
    if "artifact uri" in str(e):
        recorder.start_run()
        path = recorder.local_path
    else:
        raise

Prevention

Always obtain recorders via R.get_recorder() after a completed run, not from a fresh experiment object.
Assert recorder.artifact_uri is not None before any artifact/local-path access.
Confirm the MLflow tracking server is reachable before starting workflows.

When it happens

Trigger: Calling recorder.local_path (directly or via APIs that save/read artifacts locally) on a Recorder instance obtained before start_run()/create_recorder was committed, e.g. R.get_recorder() on a fresh experiment where the run failed to start, or a Recorder constructed manually without starting.

Common situations: Experiment was created but the recorder start threw earlier (e.g. MLflow tracking server unreachable), user reuses a stale Recorder object after the run crashed, or the tracking backend is remote (the sibling RuntimeError branch also fires when the artifact dir is not local).

Related errors

AI-assisted analysis of microsoft/qlib@79633dd950 (2026-08-15). Data as JSON: /api/errors/d0e16d4a7dfa8365. Report an issue: GitHub.