Launch And Operate
The normal runtime lifecycle is:
Choose The Right Runtime
| Runtime family | Use it for | Connect through |
|---|---|---|
| Compute | Raw GPU node or slice capacity | SSH, browser terminal, metrics, workload detail |
| Scheduler app | Slurm, Kubernetes-style clusters, distributed jobs | Scheduler UI, terminal, route, job tools |
| Compose app | Container stacks and services | App route, declared ports, logs, health |
| Notebook or IDE | Interactive development | Jupyter, VS Code, browser route, terminal |
| Inference endpoint | Model serving and API-compatible endpoints | Endpoint route, model health, API clients |
See Runtime Families for a deeper comparison.
Launch Checklist
Before submitting a launch:
- Confirm tenant and project context.
- Confirm runtime family, SKU, region, and image.
- Confirm required keys, service accounts, or storage references.
- Review hourly burn, quota, entitlement, and budget posture.
- Submit once; retry only if the product tells you the request is safe to retry.
Runtime States
| State | Meaning | User action |
|---|---|---|
| Provisioning | AI Cloud accepted the request and is preparing the runtime. | Wait, then open the workload detail page. |
| Active | The runtime is ready for supported connect actions. | Connect and work. |
| Needs action | A user, tenant admin, or operator action is required. | Follow the visible action and owner. |
| Releasing | Cleanup is running. | Wait for release completion. |
| Released | Runtime usage has ended. | No further runtime action needed. |
| Failed | Provisioning or runtime setup did not complete. | Use the correlation ID for support. |
Connect
Connect actions depend on the runtime. A workload may expose one or more of:
- SSH;
- browser terminal;
- notebook or IDE route;
- scheduler UI;
- application route;
- inference API endpoint;
- metrics or health view.
If a connection action fails, return to the workload detail page and refresh the runtime state. If it still fails, capture the correlation ID and route/action name for support.
Release And Cleanup
Release unused runtimes promptly. A release should make billing stop and move the workload to a released or cleanup-complete state. If release needs action, use the visible retry or support path rather than abandoning the runtime.
Cost And Storage Habits
- Check hourly burn before starting long-running work.
- Keep project context correct before launching data-heavy jobs.
- Attach only the buckets or storage paths required by the workload.
- Release test runtimes when validation is complete.
See Billing And Storage for spend, quota, and storage review guidance.
Blocked States
Continue to Support And Troubleshooting for access denied, capacity unavailable, route failure, billing restriction, quota block, release failure, and MFA recovery paths.