Estimated costs for Jupyterhub instances

[kyle-lesinger]

Context

As we strive to move more NASA agencies (and others) into the cloud, it may be useful to have some set of metrics which can be used to estimate future costs associated with running a Hub. This could be related to:

1.) hosting hack-a-thons with a sudden spike in users and resources
2.) an increase in the number of daily users as new individuals are onboarded

This task doesn't need to be done immediately, but I'd like to use this thread to begin discussion.

Proposal

No response

Updates and actions

No response

[jnywong]

Hey @kyle-lesinger thanks for starting this discussion!

The key ingredients for predicting future costs are:

1. user activity: such as compute and home directory consumed
2. core activity: such as underlying compute, storage, networking, etc., that's always "on" regardless of user activity
3. the cloud provider's pricing for services rendered

For 1. it's reasonable that we could predict an upper limit based on the number of future users, the maximum compute they can request from a user server, and home directory quotas.

For 2. this is largely stable as long as there are no major changes to the underlying infrastructure. However, changes such as this migration we completed from EFS to EBS that took place last year make a material difference, or changing the machine type of our core nodes. So to predict core costs, we could extrapolate from the previous month into the future, for example.

For 3., well this is out of our hands! But likely stable and the same approach extrapolating from the previous month is good enough.

So, i think the greatest degree of freedom here is the number of future users which could be provided as an input parameter for modelling future costs.