What is the impact of PARALLEL - all replicas at the same time ? Will it make repair faster,? Do we expect more CPU , Load and memory usage in case if we use Parallel , compare to other settings ?
On 21 May 2018 at 22:55, Alexander Dejanovski <a...@thelastpickle.com> wrote: > You won't be able to have less segments than vnodes, so just use 256 > segments per node, use parallel as repair parallelism, and set intensity to > 1. > > You apparently have more than 3TB per node, and that kind of density is > always challenging when it comes to run "fast" repairs. > > Cheers, > > Le mar. 22 mai 2018 à 07:28, Surbhi Gupta <surbhi.gupt...@gmail.com> a > écrit : > >> We are on Dse 4.8.15 and it is cassandra 2.1. >> What are the best configuration to use for reaper for 144 nodes with 256 >> vnodes and it shows around 532TB data when we start opscenter repairs. >> >> We need to finish repair soon. >> >> On Mon, May 21, 2018 at 10:53 AM Alexander Dejanovski < >> a...@thelastpickle.com> wrote: >> >>> Hi Subri, >>> >>> Reaper might indeed be your best chance to reduce the overhead of vnodes >>> there. >>> The latest betas include a new feature that will group vnodes sharing >>> the same replicas in the same segment. This will allow to have less >>> segments than vnodes, and is available with Cassandra 2.2 and onwards (the >>> improvement is especially beneficial with Cassandra 3.0+ as such token >>> ranges will be repaired in a single session). >>> >>> We have a gitter that you can join if you want to ask questions. >>> >>> Cheers, >>> >>> Le lun. 21 mai 2018 à 15:29, Surbhi Gupta <surbhi.gupt...@gmail.com> a >>> écrit : >>> >>>> Thanks Abdul >>>> >>>> On Mon, May 21, 2018 at 6:28 AM Abdul Patel <abd786...@gmail.com> >>>> wrote: >>>> >>>>> We have a paramater in reaper yaml file called >>>>> repairManagerSchrdulingIntervalSeconds default is 10 seconds , i >>>>> tested with 8,6,5 seconds and found 5 seconds optimal for my environment >>>>> ..you go down further but it will have cascading effects in cpu and memory >>>>> consumption. >>>>> So test well. >>>>> >>>>> >>>>> On Monday, May 21, 2018, Surbhi Gupta <surbhi.gupt...@gmail.com> >>>>> wrote: >>>>> >>>>>> Thanks a lot for your inputs, >>>>>> Abdul, how did u tune reaper? >>>>>> >>>>>> On Sun, May 20, 2018 at 10:10 AM Jonathan Haddad <j...@jonhaddad.com> >>>>>> wrote: >>>>>> >>>>>>> FWIW the largest deployment I know about is a single reaper instance >>>>>>> managing 50 clusters and over 2000 nodes. >>>>>>> >>>>>>> There might be bigger, but I either don’t know about it or can’t >>>>>>> remember. >>>>>>> >>>>>>> On Sun, May 20, 2018 at 10:04 AM Abdul Patel <abd786...@gmail.com> >>>>>>> wrote: >>>>>>> >>>>>>>> Hi, >>>>>>>> >>>>>>>> I recently tested reaper and it actually helped us alot. Even with >>>>>>>> our small footprint 18 node reaper takes close to 6 hrs.<intially took >>>>>>>> 13 >>>>>>>> hrs ,i was able to tune it 50%>. But it really depends on number >>>>>>>> nodes. For >>>>>>>> example if you have 4 nodes then it runs on 4*256<vnodes> =1024 >>>>>>>> segements , >>>>>>>> so for your env. Ut will be 256*144 close to 36k segements. >>>>>>>> Better test on poc box how much time it takes and then proceed >>>>>>>> further ..i have tested so far in 1 dc only , we can actually have >>>>>>>> seperate >>>>>>>> reaper instance handling seperate dc but havent tested it yet. >>>>>>>> >>>>>>>> >>>>>>>> On Sunday, May 20, 2018, Surbhi Gupta <surbhi.gupt...@gmail.com> >>>>>>>> wrote: >>>>>>>> >>>>>>>>> Hi, >>>>>>>>> >>>>>>>>> We have a cluster with 144 nodes( 3 datacenter) with 256 Vnodes . >>>>>>>>> When we tried to start repairs from opscenter then it showed >>>>>>>>> 1.9Million ranges to repair . >>>>>>>>> And even after doing compaction and strekamthroughput to 0 , >>>>>>>>> opscenter is not able to help us much to finish repair in 9 days >>>>>>>>> timeframe . >>>>>>>>> >>>>>>>>> What is your thought on Reaper ? >>>>>>>>> Do you think , Reaper might be able to help us in this scenario ? >>>>>>>>> >>>>>>>>> Thanks >>>>>>>>> Surbhi >>>>>>>>> >>>>>>>>> >>>>>>>>> -- >>>>>>> Jon Haddad >>>>>>> http://www.rustyrazorblade.com >>>>>>> twitter: rustyrazorblade >>>>>>> >>>>>>> >>>>>>> >>>>>> >>>>>> >>>> >>>> -- >>> ----------------- >>> Alexander Dejanovski >>> France >>> @alexanderdeja >>> >>> Consultant >>> Apache Cassandra Consulting >>> http://www.thelastpickle.com >>> >>> >>> -- > ----------------- > Alexander Dejanovski > France > @alexanderdeja > > Consultant > Apache Cassandra Consulting > http://www.thelastpickle.com >
