What is a typical hub start-up time?

User server startup time ranges from a few seconds to 10 minutes, depending on two main factors:

Factor 1: Resources available on a node

When a server starts, it requests resources from a node.

If a node already exists with sufficient resources:

• Server creation takes only a few seconds

If a suitable node doesn’t exist or lacks resources:

• The autoscaler creates a new node, which takes 2 to 10+ minutes

• Node creation time depends on the cloud provider and machine type (larger machines like GPUs often take longer)

• If startup exceeds 10 minutes, you’ll see a TimeoutError and need to restart your server to use the newly created node

• This timeout is uncommon but can occur with certain machine types

Factor 2: Environment image availability

When a user server starts, JupyterHub then finds and loads the environment image, which contains all the tools available to the user.

If the image is already on the node:

• Server startup is much faster since no download is needed

If the environment image isn’t on the node:

• The image must be pulled from the container registry

• This takes a few minutes depending on image size and network conditions

• Image pulling always happens if a new node was just created

Important

Startup time can be reduced through strategies like maintaining minimum node counts or pre-pulling images, but these increase costs. Learn more about balancing responsiveness and cost in How to balance cloud responsiveness and cost.