Thanks, Camille. I just asked on user@zookeeper in addition to your blog because I thought someone here would know, and that you might be busy.
Flavio, just to outline the scenario, there are 3 regions involved (say 1, 2, and 3), each with A and B datacenters. A and B datacenters are connected by a fat low-latency pipe, but regions are on different continents. Each region has its own quorum, with N peers in datacenter A, N peers in datacenter B, and 1 peer P in the nearest datacenter in a different region. The main requirement is that the services provided by the system should remain available if any one datacenter gets lost or partitioned. P is the peer that shouldn't become the leader, if possible, so that the more remote P isn't on the critical path for update ops. P is expected to lag in voting, but that's OK, as normally at least N+1 from the remaining 2N nodes in the same region will vote with low latency. Anyhow, thanks for the responses! Ben On Thu, Sep 5, 2013 at 9:52 AM, Camille Fournier <[email protected]> wrote: > Hi Ben, > > Sorry I haven't replied to your post comment. This was "manual" in that we > made sure things like restart scripts would restart this instance last, and > we would monitor that server and restart it if it ever became the elected > leader. It's not ideal but it works, and that's the only way to do what I > suggested because observers are not voting members and thus can't help with > the datacenter failure issue. > > Best, > Camille > > > On Thu, Sep 5, 2013 at 3:00 AM, Flavio Junqueira <[email protected]>wrote: > >> Are you asking about observers? Observers don't vote and are not elected. >> >> -Flavio >> >> -----Original Message----- >> From: "Ben Horowitz" <[email protected]> >> Sent: 05/09/2013 06:57 >> To: "[email protected]" <[email protected]> >> Subject: Peer that can't become leader >> >> Hi all, >> >> Camille Fournier in her blog post on a global service discovery >> infrastructure using ZK [1] describes a setup in which, by design, a >> distinguished peer P never becomes leader. >> >> Is there any way of accomplishing this setup with ZK config files? Or >> would one have to watch the ensemble and restart P if ever P becomes >> leader? >> >> [1] >> http://whilefalse.blogspot.com/2012/12/building-global-highly-available.html#comment-form >> >> Thanks, >> Ben >>
