Hi Jean-Pierre, Thank you for your update, here are some additional details : The configuration will be distributed on a multitude of znodes, when i talk about 256MB it is because we plan to have hundreds of them. We will keep JSon data as small as possible. Regarding the connectivity we use a private network and according to the SLA a network availability of 99.99% is expected between data centers. Knowing this do you still think that we will run into SessionExpired and ConnectionLoss issues ?
Best regards, Michael 2012/7/9 Jean-Pierre Koenig <[email protected]> > [...] > But you should beware of large payload. ZK is not designed to handle > huge amount of data [....] I highly recommend not more > than 1024 KB payload. The other point you should consider here is > (network) latency. i guess your ZK clients (your 50) will see a lot of > SessionExpired or ConnectionLoss exceptions, depending on the > connectivity of your DC's among one another. > > On Mon, Jul 9, 2012 at 2:01 PM, Michael Morello > <[email protected]> wrote: > > Hi all, > > > > I work on a project and I would be happy to have your thoughts about our > > requirements and how Zookeeper meets them. > > > > The facts : > > * We need to share configuration items between 10 data centers. > > Configuration must be synchronized between data centers (actually we can > > tolerate a few seconds of inconsistency) > > * Configuration items will be serialized in JSon and together they can > fit > > into 256MB of heap > > * R/W ratio is 90% read and 10% write and client number should be low (50 > > to 100 in each data center) > > * A client running in a DC can freely communicate with a host in an > other DC > > * Latency between data center is 20 to 60 ms > > * Only 1 host (machine) per data center might be dedicated to a Zookeeper > > process : machines are big IBM AIX boxes only one is dedicated for this > > project in each DC > > * Project must survive a data center crash > > > > Since configuration items are small and they must be synchronized and we > > need a fail-over mechanism Zookeeper appears to be a good candidate, but > > i'm not sure how to deploy it mainly because we have to start only one > > Zookeeper process in each data center. > > My idea is to deploy 1 follower in only 5 DC. This way there are 5 > > followers all over the country and we can lost 2 DC). Of course all the > > clients on all the data centers must know where are the 5 zookeeper > servers. > > Do you see any downside to do this ? > > > > I know that Zookeeper has been designed to run on a LAN and on "commodity > > hardware" but regarding the R/W ratio and the latency do you think that > it > > is a good idea to deploy it this way ? >
