Justyna Betkier

The initial discussion about this proposal was on a doc: http://docs/document/d/1bcct-luMPP51YAeUhVuFV7MIXud5wqHsVBDah9WuKeo?tab=t.0 (there are still unresolved conversations there)

Summary of open topics: - controller vs library used by autoscalers for translating buffers to pod spec - k8s quotas limitations (lacking in the proposal, will research that and fill...

Update: 1. added k8s resource quota handling to proposal 2. after some discussions I am keeping the proposal to have a separate controller. When implementing we will reconsider embedding pod...

@ellistarn as FYI After an offline review with @towca - Changed name Buffer to NodeAutoscalingBuffer - Changed the Type field to ProvisioningStrategy and added information that if not specified the...

@ellistarn responding 1. It feels to me that NodeBuffer could incorrectly suggest that the buffer is empty nodes and this is not the intention as it is just spare capacity...

@ellistarn * 1 reservations are considered to be added on the scheduler level (some docs: [older one](https://bit.ly/k8s-reservations), newer approach that was rejected, but may come back in a similar form:...

@towca @ellistarn Yesterday I met with @liggitt to discuss embedding the pod spec. Following his advice I decided to get rid of embedding and the the buffer status we will...

@towca For the format of the ReplicaBufferCapacity - based on discussion with @liggitt it seems that int or string field is not the easiest to maintain over time so I...

For the controller discussion there is one more argument. For buffers for objects implementing scale subresource if we go with a library and no pod template in the status we...

@ellistarn @jonathan-innis @jackfrancis @towca @x13n Hi! Today we discussed with our product team and based on these discussions would like to put up for comments changed design: https://docs.google.com/document/d/1wrPLdWqPRJzRTDPMT7WRYDYYx7L65LNd99qBD6y6U38/edit?tab=t.0 Changes: -...