AutoScaling

The template scales capacity along three independent axes:

• workspaces stop themselves when idle to release the pods they hold;

• Kubernetes scales component deployments up or down to adjust for traffic load;

• the cluster adds and removes nodes dynamically to fit the running pods.

Together they keep enough capacity for active work without paying for idle nodes.

Workspace idle shutdown

To avoid paying for workspaces nobody is using, the Jupyter K8s operator stops workspaces after a period of no activity. Stopped workspaces free their pods (and thus let the Cluster Autoscaler remove now-empty nodes) while retaining their persistent storage, so a user can start again where they left off. Refer to Idle Shutdown documentation for more details.

You can control the default idle shutdown configuration with these variables:

Variable	Default	Purpose
workspaces_idle_shutdown_enabled	true	Turn idle shutdown on or off.
workspaces_idle_shutdown_timeout_default	60	Minutes of idleness before shutdown, unless a user overrides it.
workspaces_idle_shutdown_timeout_min	15	Lowest timeout a user may set (admin floor).
workspaces_idle_shutdown_timeout_max	480	Highest timeout a user may set (admin ceiling, caps cost exposure).

Note

Users can set their own timeout per workspace, bounded by the min and max you set. The floor is also bound by the operator’s idle-check cadence (5 minutes by default).

Restart a stopped workspace at any time:

• from the WebUI

• with jd: jd server start --name my-workspace

• with kubectl: kubectl patch workspace <name> --type=merge -p '{"spec":{"desiredStatus":"Running"}}'

Node autoscaling

Each EKS managed node group has a size range — min_size, desired_size, and max_size — set per node group in the node_groups variable. The default preset ships:

Node group	Role	Instance type	min	desired	max
components	components	t3.medium	1	2	3
workspaces	workspaces	c5.2xlarge	2	2	5

The Cluster Autoscaler runs on the cluster and adjusts each node group within its min/max in response to demand:

• When pods can’t be scheduled for lack of capacity (Pending pods), it adds nodes, up to max_size.

• When nodes sit underused, it removes them, down to min_size.

Note

The Cluster Autoscaler’s image version tracks the cluster’s Kubernetes minor version. The template pins them together.