Scaling nodepools

Scaling nodepools#

When we provision Kubernetes clusters, we setup two, sometimes three, nodepools:

  • core that contains ‘always-on’ services such as the hub itself;

  • notebooks where users’ notebook servers are created;

  • and optionally, dask where dask pods are created for dask-enabled hubs.

By default, we set the notebook and dask nodepools to scale to zero in order to maximise the cost-effieciency of the cluster. The drawback of this is that the first user to trigger a nodepool scale-up event is usually left waiting a long time for their server. There are specific scenarios where we would like to avoid this, such as a hub we are running for a specific event and we can expect a certain number of users and a certain time. In these scenarios, we often scale-up the nodepools manually before the event so users are not left waiting long periods for their servers. These sections document how to manually scale a nodepool for each cloud provider.

Warning

Add documentation for how we do this for:

  • GKE

  • Azure (the “terraform way”)