The title really states it all.
My current use case: I want to run a DHT node that remembers as many good nodes as possible.
Allowing the k-buckets to be (infinitely) large results in good nodes not being forgotten once they are stored in the routing table.