Is there any way you can find a third rack/server room/power supply nearby just for the 1 extra zookeeper node? You don’t have to put any kafka brokers there, just a single zookeeper. It’s less likely to have a 3-way split brain because of a network partition. It’s so much cleaner with 3 availability zones because everything would be automatic failover. This is how most people run when deployed in Amazon.
Baring that I would say the next best thing would be 3 zookeepers in one zone and 2 zookeepers in the other zone so it will auto-failover if the 2 zk zone fails. If the 3 zk zone fails you can setup a well tested set of manual steps to carefully configure a 3rd zookeeper clone (which matches the id of one of the failed nodes) and still get your system back up and running. If this is not something you have done before I suggest getting a few days of expert consulting to have someone help you set it up, test it, and document the proper failover and recovery procedures. -hans > On Mar 6, 2017, at 10:44 AM, Le Cyberian <lecyber...@gmail.com> wrote: > > Thanks Han and Alexander for taking time out and your responses. > > I now understand the risks and the possible outcome of having the desired > setup. > > What would be better in your opinion to have failover (active-active) > between both of these server rooms to avoid switching to the clone / 3rd > zookeeper. > > I mean even if there are 5 nodes having 3 in one server room and 2 in other > still there would be problem related to zookeeper majority leader election > if the server room goes down that has 3 nodes. > > is there some way to achieve this ? > > Thanks again! > > Lee > > On Mon, Mar 6, 2017 at 4:16 PM, Alexander Binzberger < > alexander.binzber...@wingcon.com> wrote: > >> I agree on this is one cluster but having one additional ZK node per site >> does not help. (as far as I understand ZK) >> >> A 3 out of 6 is also not a majority. So I think you mean 3/5 with a cloned >> 3rd one. This would mean manually switching the cloned one for majority >> which can cause issues again. >> 1. You actually build a master/slave ZK with manually switch over. >> 2. While switching the clone from room to room you would have downtime. >> 3. If you switch on both ZK node clones at the same time (by mistake) you >> screwed. >> 4. If you "switch" clones instead of moving it will all data on disk you >> generate a split brain from which you have to recover first. >> >> So if you loose the connection between the rooms / the rooms get separated >> / you loose one room: >> * You (might) need manual interaction >> * loose automatic fail-over between the rooms >> * might face complete outage if your "master" room with the active 3rd >> node is hit. >> Actually this is the same scenario with 2/3 nodes spread over two >> locations. >> >> What you need is a third cross connected location for real fault tolerance >> and distribute your 3 or 5 ZK nodes over those. >> Or live with a possible outage in such a scenario. >> >> Additional Hints: >> * You can run any number of Kafka brokers on a ZK cluster. In your case >> this could be 4 Kafka brokers on 3 ZK nodes. >> * You should set topic replication to 2 (can be done at any time) and some >> other producer/broker settings to ensure your messages will not get lost in >> switch over cases. >> * ZK service does not react nicely on disk full. >> >> >> >> Am 06.03.2017 um 15:10 schrieb Hans Jespersen: >> >>> In that case it’s really one cluster. Make sure to set different rack ids >>> for each server room so kafka will ensure that the replicas always span >>> both floors and you don’t loose availability of data if a server room goes >>> down. >>> You will have to configure one addition zookeeper node in each site which >>> you will only ever startup if a site goes down because otherwise 2 of 4 >>> zookeeper nodes is not a quorum.Again you would be better with 3 nodes >>> because then you would only have to do this in the site that has the single >>> active node. >>> >>> -hans >>> >>> >>> On Mar 6, 2017, at 5:57 AM, Le Cyberian <lecyber...@gmail.com> wrote: >>>> >>>> Hi Hans, >>>> >>>> Thank you for your reply. >>>> >>>> Its basically two different server rooms on different floors and they are >>>> connected with fiber connectivity so its almost like a local connection >>>> between them no network latencies / lag. >>>> >>>> If i do a Mirror Maker / Replicator then i will not be able to use them >>>> at >>>> the same time for writes./ producers. because the consumers / producers >>>> will request from all of them >>>> >>>> BR, >>>> >>>> Lee >>>> >>>> On Mon, Mar 6, 2017 at 2:50 PM, Hans Jespersen <h...@confluent.io> >>>> wrote: >>>> >>>> What do you mean when you say you have "2 sites not datacenters"? You >>>>> should be very careful configuring a stretch cluster across multiple >>>>> sites. >>>>> What is the RTT between the two sites? Why do you think that MIrror >>>>> Maker >>>>> (or Confluent Replicator) would not work between the sites and yet you >>>>> think a stretch cluster will work? That seems wrong. >>>>> >>>>> -hans >>>>> >>>>> /** >>>>> * Hans Jespersen, Principal Systems Engineer, Confluent Inc. >>>>> * h...@confluent.io (650)924-2670 >>>>> */ >>>>> >>>>> On Mon, Mar 6, 2017 at 5:37 AM, Le Cyberian <lecyber...@gmail.com> >>>>> wrote: >>>>> >>>>> Hi Guys, >>>>>> >>>>>> Thank you very much for you reply. >>>>>> >>>>>> The scenario which i have to implement is that i have 2 sites not >>>>>> datacenters so mirror maker would not work here. >>>>>> >>>>>> There will be 4 nodes in total, like 2 in Site A and 2 in Site B. The >>>>>> >>>>> idea >>>>> >>>>>> is to have Active-Active setup along with fault tolerance so that if >>>>>> one >>>>>> >>>>> of >>>>> >>>>>> the site goes on the operations are normal. >>>>>> >>>>>> In this case if i go ahead with 4 node-cluster of both zookeeper and >>>>>> >>>>> kafka >>>>> >>>>>> it will give failover tolerance for 1 node only. >>>>>> >>>>>> What do you suggest to do in this case ? because to divide between 2 >>>>>> >>>>> sites >>>>> >>>>>> it needs to be even number if that makes sense ? Also if possible some >>>>>> >>>>> help >>>>> >>>>>> regarding partitions for topic and replication factor. >>>>>> >>>>>> I already have Kafka running with quiet few topics having replication >>>>>> factor 1 along with 1 default partition, is there a way to repartition >>>>>> / >>>>>> increase partition of existing topics when i migrate to above setup ? I >>>>>> think we can increase replication factor by Kafka rebalance tool. >>>>>> >>>>>> Thanks alot for your help and time looking into this. >>>>>> >>>>>> BR, >>>>>> >>>>>> Le >>>>>> >>>>>> On Mon, Mar 6, 2017 at 12:20 PM, Hans Jespersen <h...@confluent.io> >>>>>> >>>>> wrote: >>>>> >>>>>> Jens, >>>>>>> >>>>>>> I think you are correct that a 4 node zookeeper ensemble can be made >>>>>>> to >>>>>>> work but it will be slightly less resilient than a 3 node ensemble >>>>>>> >>>>>> because >>>>>> >>>>>>> it can only tolerate 1 failure (same as a 3 node ensemble) and the >>>>>>> likelihood of node failures is higher because there is 1 more node >>>>>>> that >>>>>>> could fail. >>>>>>> So it SHOULD be an odd number of zookeeper nodes (not MUST). >>>>>>> >>>>>>> -hans >>>>>>> >>>>>>> >>>>>>> On Mar 6, 2017, at 12:20 AM, Jens Rantil <jens.ran...@tink.se> >>>>>>>> >>>>>>> wrote: >>>>> >>>>>> Hi Hans, >>>>>>>> >>>>>>>> On Mon, Mar 6, 2017 at 12:10 AM, Hans Jespersen <h...@confluent.io> >>>>>>>>> >>>>>>>> wrote: >>>>>>> >>>>>>>> A 4 node zookeeper ensemble will not even work. It MUST be an odd >>>>>>>>> >>>>>>>> number >>>>>> >>>>>>> of zookeeper nodes to start. >>>>>>>>> >>>>>>>> >>>>>>>> Are you sure about that? If Zookeer doesn't run with four nodes, that >>>>>>>> >>>>>>> means >>>>>>> >>>>>>>> a running ensemble of three can't be live-migrated to other nodes >>>>>>>> >>>>>>> (because >>>>>>> >>>>>>>> that's done by increasing the ensemble and then reducing it in the >>>>>>>> >>>>>>> case >>>>> >>>>>> of >>>>>>> >>>>>>>> 3-node ensembles). IIRC, you can run four Zookeeper nodes, but that >>>>>>>> >>>>>>> means >>>>>> >>>>>>> quorum will be three nodes, so there's no added benefit in terms of >>>>>>>> availability since you can only loose one node just like with a three >>>>>>>> >>>>>>> node >>>>>>> >>>>>>>> cluster. >>>>>>>> >>>>>>>> Cheers, >>>>>>>> Jens >>>>>>>> >>>>>>>> >>>>>>>> -- >>>>>>>> Jens Rantil >>>>>>>> Backend engineer >>>>>>>> Tink AB >>>>>>>> >>>>>>>> Email: jens.ran...@tink.se >>>>>>>> Phone: +46 708 84 18 32 >>>>>>>> Web: www.tink.se >>>>>>>> >>>>>>>> Facebook <https://www.facebook.com/#!/tink.se> Linkedin >>>>>>>> <http://www.linkedin.com/company/2735919?trk=vsrp_ >>>>>>>> >>>>>>> companies_res_photo&trkInfo=VSRPsearchId%3A1057023381369207406670% >>>>>>> 2CVSRPtargetId%3A2735919%2CVSRPcmpt%3Aprimary> >>>>>>> >>>>>>>> Twitter <https://twitter.com/tink> >>>>>>>> >>>>>>> >>> >>
