Stop, start, and restart an endpoint

Stop an endpoint during quiet periods, start it when you need it, or restart it to make it pull its model or code again. These actions are on the endpoint’s row in Model Endpoints and on its details page.

Before you begin

• Permission to update model endpoints in the project. A Member needs a policy that grants update permission for the Inference service. The portal shows only the actions your permissions allow.
• An endpoint in the state each action needs. Stop Service needs an hourly endpoint in any state other than Stopped. Start Service needs a Stopped endpoint. Restart Service needs a Running endpoint.

Stop an endpoint

1. Select Stop Service. The dialog asks “Are you sure you want to stop inference service <name>?”
2. Select Stop.

The endpoint goes to Terminating and then Stopped when every worker is gone. An endpoint already at zero workers goes straight to Stopped. Stopping removes the workers and ends compute billing. The endpoint, its address, and its settings stay, and Start Service brings it back.

Start an endpoint

1. Select Start Service. The dialog asks “Are you sure you want to start inference service <name>?”
2. Select Start.

The endpoint comes back with the worker counts it was stopped with. It is Waiting until its workers appear, or Idle when its minimum is 0 on a scale-to-zero policy. A start checks free capacity and credit again, and a refusal reads “SKU Inventory temporarily unavailable. Please try again.” or “Insufficient Credits!”. Starting begins a new billed period.

Restart an endpoint

1. Wait until the endpoint is Running, then select Restart Service under Actions.
2. Confirm the dialog “Are you sure you want to restart inference service <name>?”

TIR confirms the restart, and the state updates a few seconds later. A restart is the way to load new files that you uploaded to the endpoint’s Model Repository. The workflow of a Custom Container endpoint linked to GitHub restarts it when someone merges a change. See Deploy a custom container.

Restart one worker

On the Workers tab, select Restart on a worker’s row and confirm. TIR replaces that one worker with a new one, while the other workers keep serving. Use this when one worker misbehaves and the rest are fine.

Check the state

The state in the list and on the details page shows where the endpoint is. Only hourly endpoints have Stop Service; the Model Endpoints FAQ explains why. If an endpoint stays in Terminating, or the action is refused, see Troubleshooting Model Endpoints.