Have you seen RAMP transactions? I think that's a much better fit for C* than fully linearizable operations cross-partition.
https://issues.apache.org/jira/browse/CASSANDRA-7056 On Fri, Aug 7, 2015 at 7:56 AM, Marek Lewandowski < marekmlewandow...@gmail.com> wrote: > actually I have been also thinking about doing something like redundant > execution of transaction. So you have this *single active thing* that > executes transaction, but you can also have redundancy of form of other > _followers_ that try to execute same transactions (like a dry-run) and upon > detection of failure of *single active thing* one of them could pick > transaction execution and finish it. Still it's a little bit vague and > needs a lot more details, but now system could recover from failure of this > _single active thing_. What do you think? > > 2015-08-07 14:48 GMT+02:00 Robert Stupp <sn...@snazy.de>: > > > > > > On 07 Aug 2015, at 14:35, Marek Lewandowski < > marekmlewandow...@gmail.com> > > wrote: > > > > > > In both of my ideas there > > > is some central piece. > > > > > > That’s the point - a single thing. A single thing IS a > > single-point-of-failure. > > Sorry to reply that drastically: that’s an absolute no-go in C*. Every > > node must be equal - no special “this” or special “that”. > > > > — > > Robert Stupp > > @snazy > > > > > > > -- > Marek Lewandowski > -- Jonathan Ellis Project Chair, Apache Cassandra co-founder, http://www.datastax.com @spyced
