Troubleshooting¶
Checks for the MLflow failures the published getting-started page names.
The UI does not load¶
Cause: first boot is still creating accounts, or port 5000 is not
forwarded. Initialization takes about two minutes. The server listens on
127.0.0.1:5000.
Fix:
sudo cloud-init status --wait
sudo systemctl status mlflow sushi-mlflow-firstboot sushi-mlflow-ready
Then forward port 5000 and open http://localhost:5000 on the workstation.
See Open the tracking UI.
The browser rejects the sign-in¶
Cause: the values in the form do not match
~/.config/sushi/mlflow/access.json on this instance.
Fix: open access.json in the ubuntu session editor and use that
account for the UI. /etc/mlflow/owner.json is a separate administrator
identity, readable through root. Do not paste either file into a ticket.
A forgotten administrator password¶
Cause: the administrator identity is generated once and then preserved.
Fix: use the authenticated rotation procedure in
/opt/sushi/docs/README.md. Do not delete /var/lib/mlflow or /etc/mlflow
to obtain a new database. If the image notes do not contain the rotation
steps, email info@softwaresushi.com and keep
the instance as it is.
sushi-mlflow read does not see the demo¶
Cause: the manifest path is from a different run, or the service was restored from only one of the SQLite databases.
Fix: pass the manifest path the demo printed under ~/mlflow-results/.
A consistent backup includes /var/lib/mlflow, /etc/mlflow, and the
customer access file. See Backups.
A model file was uploaded for scoring¶
Cause: this server stores artifacts. It does not serve uploaded models, execute pickles, run server jobs, run periodic job tasks, or run custom scorers.
Fix: leave the file as a stored artifact. Read the parameters, metrics, and artifact bytes in the tracking UI.
Last updated October 27, 2026 · Suggest an improvement