You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
To optimize costs further, one could make user-placeholder pods come and pre-warm some seats at the start of the day, and then leave at the end of the day. That way, one could avoid making users needing to wait for scale-up events during the day, but scale down properly during the night.
For something like this, I imagine a CRD + Controller binary pod running inspecting the CRD. The CRD would then specify at what kind of time one needs certain user placeholders somehow, perhaps using a cron-syntax or similar.
As far as I understand, the current implementation of "permanent" placeholders is mainly useful if the number of users varies slowly over a node startup time (i.e. number of new users arriving within 5 minutes is always small compared to the total number of users on the system). Otherwise, you either need many placeholders (expensive) or your users will encounter the node startup delay.
A possible alternative to the proposal of scheduled pre-warming would be the concept of "one-time" placeholders that simply do not reschedule after they've been taken over. An admin would add those one-time placeholders to the system a couple of minutes before an increase in the number of users is expected.
I don't know whether that would be easier or harder from a kubernetes perspective; either approach would be a solution to the use case.
To optimize costs further, one could make user-placeholder pods come and pre-warm some seats at the start of the day, and then leave at the end of the day. That way, one could avoid making users needing to wait for scale-up events during the day, but scale down properly during the night.
For something like this, I imagine a CRD + Controller binary pod running inspecting the CRD. The CRD would then specify at what kind of time one needs certain user placeholders somehow, perhaps using a cron-syntax or similar.
An example controller using a CRD can be found here: https://github.com/kubernetes/sample-controller
The text was updated successfully, but these errors were encountered: