Kubecon SLC 2024

kubecon post thoughts.

1. PodDisription Conidtions and Eviction Backpressure on replicasets-deployments and stateful sets. Started an issue 128815 to get a condition on for blocked evictions. The former should help k8s-pdb-autoscaler along but the latter should make it not necessary. The idea of back pressure is replicasets controllers would monitor their pods and update their own status if they’re seeing seeing eviction blocked on pod status. Deployment controller could then scale up if its the most recent replicaset or if there is a newere replicaset that can’t make progress scale that down then scalue up the older defacto replicaset to unblock evictions.

2. Antonio’s/Network-Sig’s kube-network-policies may be able to be trivailly adapted to windows with this tailscale windows filter library. Theorically you could use ebpf to send packets into user space as well.

3. Cilum Scaling. Several good presentations thoug many stared with cluster mesh. Not really a desire for us to got etcd but there are limitations
   1. CBOR Serializer may help some
   2. Maybe using apiserserv-mesh in built etcd as a proxy as that is what kvstoremesh is doing.
   3. Ryan was excited about replacing watch with a messaging system. I am not sure that solves our problems but loves the thinking
   4. Should definitely tighten up our identity garbage collection but also why isnt the identity the hash.
   5. Sig API maching thought streaming lists were going to help a lot. Of course growing
   6. Talks: Hemeth Clustermesh scalability, etcd vs api server, Ryan Drews clustermesh scalabiluy of cluster mesh , Plea for help testing

4. Api server Server roll back is being worked on! Need to catch up

5. Api Server backed by Postgresql was a little dissapointing. Need to measure watch performnce and write a follow up using azure postgres on cosmosdb