What to check first for each symptom, and where to read more.

Accounts and permissions

A button is missing

Buttons you lack the permission for are hidden or disabled. Check your role in the current tenant under the avatar menu → Account settings (계정 설정), and find the permission you need in Roles and permissions. Roles are changed by the tenant administrator → Members

I pressed production promotion but the version did not become Production

That is expected. Promotion to Production always creates an approval request, even for roles holding the permission, and takes effect only when someone with the approve permission approves it → Approvals

I forgot my password

The sign-in screen has no reset function. Ask a global administrator to reset it → Users, Command-line tools

Installation and operations

System Health never leaves "Needs attention"

Signing keys, the pepper (a secret mixed in when tokens are hashed) and the admin password left at their defaults stay CRITICAL — even when every connection is healthy → Mandatory security settings

Large dataset or image uploads fail halfway

Usually the request-size or response-time limit of a front proxy such as a CDN or reverse proxy. Use chunked CLI upload for datasets → Reverse proxy caveats, Uploading a large dataset

A deleted tenant comes back after a restart

Bootstrap tenants (tenants listed in the server settings and created automatically) are recreated every time the server starts. Empty the bootstrap settings before the first start → Installing with docker compose

Training

A training run stays queued

When the wired GPUs (shared pool or dedicated) are busy with other runs, it waits for a free slot. A tenant with no GPU wiring cannot use GPUs (CPU only). Check the progress stages in the run detail and GPU slots. A missing Kubernetes namespace for the tenant makes the preparation stages fail → Watching training progress

Training fails in preparation / my own training image fails

Training container troubleshooting

Can I continue a stopped run?

There is no "continue" feature. Upload the checkpoint as pretrained weights and start a new run from it → Stopping and continuing training

The metric chart is empty

Metrics are read from MLflow. Make sure the trainer logs them with mlflow.log_metrics(..., step=epoch)What the trainer must do

Deployment and edge

I deployed, but the model on the device did not change

The deployment record and delivery to devices are separate. Send Download model (모델 내려받기) from the device detail → Commands in Edge Fleet → Deploying to field devices, Rolling back a deployment

A device shows as unhealthy

Its heartbeats stopped. Check the agent on the device, the network and the token → Edge SDK troubleshooting

Alerts keep firing after I silenced them

Silences apply only to Alertmanager alerts. Handle edge-device and drift alerts with Acknowledge and ResolveSilencing alerts

Integration

DataOps sync does not run by itself

Only manual sync exists today. Press Sync on the dataset screen → Sync rules

Logging to MLflow fails with "active experiment ID does not match"

Inside a platform training run, do not call set_experiment() or set_tracking_uri(); the trainer must inherit the run the server created → What the trainer must do

Written for the platform as of 2026-09-21.

© Geo-MLOps