My homelab is a k3s cluster running on a handful of raspberry pis, using an external mysql database for k3s datastore. Most of these pis have just 1Gb of memory, and with just the basics installed they are already using 40%-50% of available memory, with the server using over 80%.
I'm always on the lookout for ways to conserve memory, and recently someone mentioned cilium being "better" than the other CNI implementations. Does anyone have experience with the resource usage of various CNI implementations compared to flannel in a k3s cluster?