Measurements from a homelab, mostly about OpenShift infrastructure.
Every post here follows three rules: a number that isn’t published anywhere else, a setup you can reproduce, and a claim you can prove wrong. If it wasn’t measured on hardware I own, it isn’t here.
I’m learning most of this as I build it.
A pod on a NoOverlay network reached a host outside the cluster with its own IP, and the host reached it back — once BGP moved off the second NIC I’d copied from the EVPN cluster. Two such networks still couldn’t reach each other, though the fabric would have routed them.
A VM on OpenShift joined the subnet two plain hosts already share, moved to another node, and lost two pings. The fabric saw the same MAC in a new place, with a sequence number to prove it.
A pod reached a VM that has never heard of Kubernetes, at ttl=62 — the cluster nodes joining the EVPN fabric as VTEPs in their own right.
The fabric already carried two hosts across leaves. A pod sits on a different subnet, so it has to be routed — and routing needs a gateway that doesn’t exist yet. Adding one changed what the fabric advertises, not what it is.
I wanted to try OpenShift’s BGP EVPN support, so first I needed a fabric for it to join. One spine, two leaves, two hosts that ping each other with no L2 path between them — and the part I had wrong: FRR never creates the tunnel.