Today, if a node transfers its leases away, any other node that is neither the old nor new holder of the lease but has a cached record of that lease, will first attempt to contact the old leaseholder before it "discovers" that the lease is somewhere else.

If the old leaseholder is down at that time (as would be expected if it was e.g. stopped or restarted), a brief period of unavailability ensures until the requesting node looks up an update LH record.

We want to avoid this unavailability by preemptively noticing all nodes in the cluster of LH changes, so they can invalidate their caches.

@andreimatei suggests to do this via gossip.

Jira issue: CRDB-4143