[lethain]

Regarding the hard-coded IP piece, I've run into a similar scenario wherein the early versions of the distributed counters patch for Cassandra relied on vector-clocks which used IPs as part of those clocks. Certainly not a robust design in the case of machine failure, but I've seen smart people make awkward decisions, and if someone had gone into production with that patch and lost a machine, it would have been remarkably annoying to recover (it would have required rewriting all the SSTables to a new IP, which isn't impossible or anything, but kind of a pita).