> > auth logins for super users is 101 replicas serving the read
This only applies to the default superuser (i.e. 'cassandra'), which is one of the reasons for recommending it is only used during initial setup[1]. Reads for all other users, including superusers, are done at LOCAL_ONE [1] http://cassandra.apache.org/doc/latest/operating/security.html#authentication On Sun, Apr 2, 2017 at 7:07 AM, Jeff Jirsa <jji...@apache.org> wrote: > > You should use a network topology strategy with high RF in each DC > > > There's some debate here - some blogs/speakers will say to put a replica > on each instance, but that falls down above a few dozen instances. Imagine > if you have (for example) 200 instances per DC, auth logins for super users > is 101 replicas serving the read - that's a really slow login that's likely > to fail (think about thread pools on the coordinator doing the read > response handling, it's an ugly ugly mess). > > Normal logins do use LOCAL_ONE though so if there are lots of replicas, > auth will be faster - so use 5-10 replicas per DC, and crank up the caching > timeouts as well > >
