Skip to main content

Understanding Deployment Status

StatusMeaning
PendingProvisioning — Kubernetes resources created, pod starting up
PullingDownloading the model weights — this can take several minutes
RunningModel is loaded and the endpoint is live
IdleScaled to zero — no traffic for 30 min; will auto-wake on the next request
FailedDeployment could not start — check the error message

Idle status and auto-wake

Managed CaseDesk deployments (Starter, Team, Advanced) scale to zero automatically when there has been no traffic for 30 minutes. The status changes to idle.

When a new request arrives at an idle endpoint, CaseDesk wakes the deployment automatically. The request is held open while the model loads — no retry is needed. Once the deployment reaches running, the response is returned.

Cold-start times from idle:

  • Starter (Ollama): 15–30 seconds
  • Team / Advanced (vLLM): 30–90 seconds
  • Sandbox (Ollama on AWS ASG, includes GPU node provisioning): ~2 minutes

See Autoscaling for more detail.

First deploy cold-start times

  • Managed CaseDesk: 5–10 min (cluster provisioning + image pull + model download)
  • CaseDesk Sandbox: 10–18 min (GPU node provisioning + image pull + model download)

Subsequent deploys on a warm node are significantly faster as the runtime image is already cached.

Viewing logs

From the deployment detail page, click View Logs to see the pod output in real time. This is the fastest way to diagnose a failed deployment.