blakeblackshear
Readme.md lacks Coral TPU clarification for K8s use
The readme.md should state clearly that every eligible worker node has to have its own Coral TPU.
I'm still new to trying Frigate so I might be misunderstanding your issue. I have only one Coral TPU and it is attached to one node in my multi-node cluster. I've applied a node label via kubectl and corresponding nodeSelector in the Helm values and this ensures the Frigate pod is always scheduled on that node with the TPU. Are there situations when multiple Frigate pods would be running? Or might this node label and nodeSelector combination address your issue?
One of benefits of a K8s cluster is the increase in application availability due to the ability to move a pod to a different worker node if the one it was running previously becomes inavailable for some reason. Of course, labelling and using a nodeSelector works. But the frigate pod will never be able to run on a different worker node if that particular node does not have a Coral hardware device.