Hi Justin, If you have 6 node cluster with RF = 3 nodes in each rack. So nodes in rac1 will be primary ower of different token ranges and their replica will be rac2 and rac3. If one of the node in rac1 goes down then their replicas present in rac2 and rac3 will be serving the request. However note that hints for the down node will be stored on cordinator. If you are using token aware policy at client side that the replica on rac2/3 will be coordinator and will store hint for the down node in rac1.
regards, Laxmikant On Wed, Mar 13, 2019 at 1:44 AM Justin Sanciangco <jsancian...@blizzard.com.invalid> wrote: > Maybe this was a specific issue to my topology in the past where I had 9 > nodes with a 3 rack implementation. Each rack contained a unique replica > set so when a node went down it put very high load on the nodes in the same > rack. How does the data get distributed in this case where there are only 2 > nodes in each of the 3 racks? > > > > - Justin Sanciangco > > > > > > *From: *Alexander Dejanovski <a...@thelastpickle.com> > *Reply-To: *"user@cassandra.apache.org" <user@cassandra.apache.org> > *Date: *Tuesday, March 12, 2019 at 10:56 AM > *To: *user <user@cassandra.apache.org> > *Subject: *Re: Changing existing Cassandra cluster from single rack > configuration to multi racks configuration > > > > Hi Justin, > > > > I'm not sure I follow your reasoning. In a 6 node cluster with 3 racks (2 > nodes per rack) and RF 3, if a node goes down you'll still have one node in > each of the other racks to serve the requests. Nodes within the same racks > aren't replicas for the same tokens (as long as the number of racks is > greater or equal to the RF). > > > > Regarding the other question with the decommission/rebootstrap procedure, > unbalances are indeed to be expected, and I'd favor the DC switch > technique, but it may not be an option. > > > > Cheers, > > > > Le mar. 12 mars 2019 à 18:28, Justin Sanciangco > <jsancian...@blizzard.com.invalid> a écrit : > > I would recommend that you do not go into a 3 rack single dc > implementation with only 6 nodes. If a node goes down in this situation, > the node that is paired with the node that is downed will have to service > all of the load instead of being evenly distributed throughout the cluster. > While its conceptually nice to have 3 rack implementation, it does have > some negative implications when not at a proper node count. > > > > What features are you trying to make use of with going with multirack? > > > > - Justin Sanciangco > > > > > > *From: *Laxmikant Upadhyay <laxmikant....@gmail.com> > *Reply-To: *"user@cassandra.apache.org" <user@cassandra.apache.org> > *Date: *Monday, March 11, 2019 at 10:52 PM > *To: *"user@cassandra.apache.org" <user@cassandra.apache.org> > *Subject: *Re: Changing existing Cassandra cluster from single rack > configuration to multi racks configuration > > > > Hi Alex, > > > > Regarding your below point the admin need to take care of temporary uneven > distribution of data util the entire process is done: > > > > "If you can't, then I guess you can for each node (one at a time), > decommission it, wipe it clean and re-bootstrap it after setting the > appropriate rack." > > > > I believe while doing so in the existing single rack cluster, the first > new node joined with different rack (rac2) will get 100% loaded in terms so > disk usage will be proportionally very high in comparison to other nodes in > rac1. > > So until both racks have equal number of nodes and we run nodetool cleaup, > the data will not be equally distributed. > > > > > > On Wed, Mar 6, 2019 at 5:50 PM Alexander Dejanovski < > a...@thelastpickle.com> wrote: > > Hi Manish, > > > > the best way, if you have the opportunity to easily add new > hardware/instances, is to create a new DC with racks and switch traffic to > the new DC when it's ready (then remove the old one). My co-worker Alain > just wrote a very handy blog post on that technique : > http://thelastpickle.com/blog/2019/02/26/data-center-switch.html > > > > If you can't, then I guess you can for each node (one at a time), > decommission it, wipe it clean and re-bootstrap it after setting the > appropriate rack. > > Also, take into account that your keyspaces must use the > NetworkTopologyStrategy so that racks can be taken into account. Change the > strategy prior to adding the new nodes if you're currently using > SimpleStrategy. > > > > You cannot (and shouldn't) try to change the rack on an existing node (the > GossipingPropertyFileSnitch won't allow it). > > > > Cheers, > > > > On Wed, Mar 6, 2019 at 12:15 PM manish khandelwal < > manishkhandelwa...@gmail.com> wrote: > > We have a 6 node Cassandra cluster in which all the nodes are in same > rack in a dc. We want to take advantage of "multi rack" cluster (example: > parallel upgrade on all the nodes in same rack without downtime). I would > like to know what is the recommended process to change an existing cluster > with single racks configuration to multi rack configuration. > > > > I want to introduce 3 racks with 2 nodes in each rack. > > > > Regards > > Manish > > > > -- > > ----------------- > > Alexander Dejanovski > > France > > @alexanderdeja > > > > Consultant > > Apache Cassandra Consulting > > http://www.thelastpickle.com > > > > > -- > > > > regards, > > Laxmikant Upadhyay > > > > -- regards, Laxmikant Upadhyay
