Skip to main content
A deployment’s status describes its lifecycle and readiness to serve requests. Check the status in your model’s deployment list, then use logs and metrics to investigate errors.

​
Model statuses

These labels appear in the dashboard for model deployments:
  • Active: The deployment is ready to serve requests.
  • Scaled to zero: The deployment has no running replicas. An inference request or a wake request starts it again. Later scaling follows its autoscaling settings.
  • Waking up: The deployment is starting replicas after scaling to zero. Requests can wait while the model loads.
  • Inactive: The deployment is deactivated and doesn’t serve requests. Activate it before sending traffic.
  • Building, Deploying, or Loading model: The deployment is preparing to serve requests. Check its logs for progress or startup errors.
  • Updating: A deployment update is in progress.
  • Deactivating: The deployment is shutting down.
  • Build stopped: The build was stopped before it finished.
  • Unhealthy: The deployment isn’t healthy and can still consume resources. Check the serving logs and resource metrics.
  • Build failed: The image build failed. Check the build logs for the failing command or dependency.
  • Failed: The deployment couldn’t start or is no longer available. Check deployment logs for the cause.
The Management API returns machine-readable status values rather than dashboard labels. For example, Waking up is WAKING_UP, and a failed deployment can return DEPLOY_FAILED or FAILED. See the status field in Get a deployment for the full set of API values.

​
Fix unhealthy deployments

  1. Open the deployment’s logs. For a failed build, inspect the build logs; for startup or serving failures, inspect the deployment logs.
  2. Check resource metrics for memory pressure or other resource limits around the failure. For a custom server, verify its health-check routes.
  3. Fix the error reported in the logs, such as a missing dependency, a model-loading failure, or an unavailable external service. Deploy the corrected code or configuration when needed.
  4. Confirm that the deployment becomes Active, then send a test request and check its response and logs.
If you need to stop the deployment while investigating, deactivate it. Deactivation makes it unavailable for inference; reactivate it when you’re ready to resume serving.