In the meantime, the one-at-a-time process is remarkably usable. We use it routinely for exactly the purpose proposed by Gustavo.
Scaling by magic will be even better, but in the meantime people shouldn't be afraid to modify their clusters. On Wed, Jul 1, 2009 at 10:34 AM, Gustavo Niemeyer <gust...@niemeyer.net>wrote: > ... I guess the same basic idea that everyone else has for it: bringing > additional systems up and down dynamically to control scalability and > reliability vs. cost of having many machines running at once. I > understand it is possible to restart the servers one at a time to > change the server list without a full stop, but if we end up bundling > this in some open source framework for people to use in the wild, the > least manual interaction and procedural maintenance the best. Having > ZooKeeper clients learning about the server list dynamically will help > a lot in this scenario too. > >