Spark is a special snowflake and its WebUI follow-on links are pretty heavily based on having the cluster IPs routable from the browser. Attached is a sample screenshot of what I mean. As part of the drive in #16517, I have the first page working just fine as part of kubectl proxy, but, of course, the links to the workers and application itself don't work, because those are cluster IPs.

If we had a way to get through the service proxy by endpoint spec and directly proxy rather than just proxying an individual service, it would be then be straightforward to write a separate pod that was just a URL rewriter for nginx for all "cluster-like" URLs (handwaving a bit, but assuming you passed the CIDR in, it seems straightforward).

cc @brendandburns