Hi Neha, Camille, I wanted to share something we did as part of the dynamic membership change feature development (ZK-107) that seems very related to the discussion here and might solve the problem.
When the membership changes, similarly to what you wrote below, clients sometimes need to move. Obviously this is the case if the server they are connected to is no longer in the cluster. The idea is that they should move in a way that minimizes unnecessary client migration (such as the one that would happen if you have every client re-shuffle the new server list) and yet leaves the system in a balance state (the number of clients connected to each server is the same in expectation). The idea was to come up with a set of probabilistic rules that each client applies locally to see whether and where it should migrate. The resulting rules as well as the evaluation in Zookeeper are in the document attached to the jira (https://issues.apache.org/jira/browse/ZOOKEEPER-1355) I'll provide a patch soon. (this is part of a larger paper about reconfiguration that is under submission). In terms of implementation changes, I added a command "zk.updateServerList(hostlist);" and an implantation of the probabilistic rules in StaticHostProvider. I wanted to get your feedback and hopefully this may solve the problem you're discussing here. Thanks a lot, Alex > -----Original Message----- > From: c...@renttherunway.com [mailto:c...@renttherunway.com] On Behalf Of > Camille Fournier > Sent: Monday, January 09, 2012 12:47 PM > To: dev@zookeeper.apache.org > Subject: Re: Performing no downtime hardware changes to a live > zookeeper cluster > > Sounds fine with me, probably should make it a flaggable option. > > C > > > On Mon, Jan 9, 2012 at 3:33 PM, Neha Narkhede > <neha.narkh...@gmail.com>wrote: > > > >> If you just have machine names in a list that you pass in, then > yes, we > > could re-resolve on every reconnect and you could just re-alias that > name > > to a new IP. But you'll have to put in logic that will do that but > not > > break people using DNS RR. > > > > Having a list of machine names that can be changed to point to new > IPs > > seems reasonable too. To be able to do the upgrade without having to > > restart all clients, besides turning off DNS caching in the JVM, we > > still have to solve the problem of zookeeper client caching the IPs > in > > code. Having 2 levels of DNS caching, one in the JVM and one in code > > (which cannot be turned off) doesn't look like a good idea. Unless > I'm > > missing the purpose of such IP caching in zookeeper ? > > > > >> I realize that moving machines is difficult when you have lots of > > clients. > > I'm a bit surprised your admins can't maintain machine IP addresses > on a > > machine move given a cluster of that complexity, though > > > > Its not like it can't be done, it definitely has quite some > > operational overhead. We are trying to brainstorm various approaches > > and come up with one that will involve the least overhead on such > > upgrades going forward. > > > > Having said that, seems like re-resolving host names in reconnect > > doesn't look like a bad idea, provided it doesn't break the DNS RR > use > > case. If that sounds good, can I go ahead a file a JIRA for this ? > > > > Thanks, > > Neha > > > > On Mon, Jan 9, 2012 at 11:04 AM, Camille Fournier > <cami...@apache.org> > > wrote: > > > We don't shuffle IPs after the initial resolution of IP addresses. > > > > > > In DNS RR, you resolve to a list of IPs, shuffle these, and then we > round > > > robin through them trying to connect. If you re-resolve on every > > > round-robin, you have to put in logic to know which ones have > changed and > > > somehow maintain that shuffle order or you aren't doing a fair back > end > > > round robin, which people using the ZK client against DNS RR are > relying > > on > > > today. > > > > > > If you just have machine names in a list that you pass in, then > yes, we > > > could re-resolve on every reconnect and you could just re-alias > that name > > > to a new IP. But you'll have to put in logic that will do that but > not > > > break people using DNS RR. > > > > > > I realize that moving machines is difficult when you have lots of > > clients. > > > I'm a bit surprised your admins can't maintain machine IP addresses > on a > > > machine move given a cluster of that complexity, though. I also > think > > that > > > if we're going to be putting special cases like this in we might > just > > want > > > to go all the way to a pluggable reconnection scheme, but maybe > that is > > too > > > aggressive. > > > > > > C > > > > > > On Mon, Jan 9, 2012 at 1:51 PM, Neha Narkhede > <neha.narkh...@gmail.com > > >wrote: > > > > > >> Maybe I didn't express myself clearly. When I said DNS RR, I meant > its > > >> simplest implementation which resolves a hostname to multiple IPs. > > >> > > >> Whatever method you use to map host names to IPs, the problem is > that > > >> the zookeeper client code will always cache the IPs. So to be able > to > > >> swap out a machine, all clients would have to be restarted, which > if > > >> you have 100s of clients, is a major pain. If you want to move the > > >> entire cluster to new machines, this becomes even harder. > > >> > > >> I don't see why re-resolving host names to IPs in the reconnect > logic > > >> is a problem for zookeeper, since you shuffle the list of IPs > anyways. > > >> > > >> Thanks, > > >> Neha > > >> > > >> > > >> On Mon, Jan 9, 2012 at 10:31 AM, Camille Fournier > <cami...@apache.org> > > >> wrote: > > >> > You can't sensibly round robin within the client code if you > > re-resolve > > >> on > > >> > every reconnect, if you're using dns rr. If that's your goal > you'd > > want a > > >> > list of dns alias names and re-resolve each hostname when you > hit it > > on > > >> > reconnect. But that will break people using dns rr. > > >> > You can look into writing a pluggable reconnect logic into the > zk > > client, > > >> > that's what would be required to do this but at the end of the > day > > you'll > > >> > have to give your users special clients to make that work. > > >> > > > >> > C > > >> > On Jan 9, 2012 1:16 PM, "Neha Narkhede" > <neha.narkh...@gmail.com> > > >> wrote: > > >> > > > >> >> I was reading through the client code and saw that zookeeper > client > > >> >> caches the server IPs during startup and maintains it for the > rest of > > >> >> its lifetime. If we go with the DNS RR approach or a load > balancer > > >> >> approach, and later swap out a server with a new one ( with a > new IP > > >> >> ), all clients would have to be restarted to be able to > "forget" the > > >> >> old IP and see the new one. That doesn't look like a clean > approach > > to > > >> >> such upgrades. One way of getting around this problem, is > adding the > > >> >> resolution of host names to IPs in the "reconnect" logic in > addition > > >> >> to the constructor. So when such upgrades happen and the client > > >> >> reconnects, it will see the new list of IPs, and wouldn't > require to > > >> >> be restarted. > > >> >> > > >> >> Does this approach sound good or am I missing something here ? > > >> >> > > >> >> Thanks, > > >> >> Neha > > >> >> > > >> >> On Wed, Dec 21, 2011 at 7:21 PM, Camille Fournier < > > cami...@apache.org> > > >> >> wrote: > > >> >> > DNS RR is good. I had good experiences using that for my > client > > >> >> > configs for exactly the reasons you are listing. > > >> >> > > > >> >> > On Wed, Dec 21, 2011 at 8:43 PM, Neha Narkhede < > > >> neha.narkh...@gmail.com> > > >> >> wrote: > > >> >> >> Thanks for the responses! > > >> >> >> > > >> >> >>>> How are your clients configured to find the zks now? > > >> >> >> > > >> >> >> Our clients currently use the list of hostnames and ports > that > > >> >> >> comprise the zookeeper cluster. For example, > > >> >> >> zoo1:port1,zoo2:port2,zoo3:port3 > > >> >> >> > > >> >> >>>> > - switch DNS, > > >> >> >>> - wait for caches to die, > > >> >> >> > > >> >> >> This is something we thought about however, if I understand > it > > >> >> >> correctly, doesn't JVM cache DNS entries forever until it is > > >> restarted > > >> >> >> ? We haven't specifically turned DNS caching off on our > clients. > > So > > >> >> >> this solution would require us to restart the clients to see > the > > new > > >> >> >> list of zookeeper hosts. > > >> >> >> > > >> >> >> Another thought is to use DNS RR and have the client zk url > have > > one > > >> >> >> name that resolves to and returns a list of IPs to the > zookeeper > > >> >> >> client. This has the advantage of being able to perform > hardware > > >> >> >> migration without changing the client connection url, in the > > future. > > >> >> >> Do people have thoughts about using a DNS RR ? > > >> >> >> > > >> >> >> Thanks, > > >> >> >> Neha > > >> >> >> > > >> >> >> On Tue, Dec 20, 2011 at 1:06 PM, Ted Dunning < > > ted.dunn...@gmail.com> > > >> >> wrote: > > >> >> >>> In particular, aren't you using DNS names? If you are, > then you > > can > > >> >> >>> > > >> >> >>> - expand the quorum with the new hardware on new IP > addresses, > > >> >> >>> - switch DNS, > > >> >> >>> - wait for caches to die, > > >> >> >>> - restart applications without reconfig or otherwise force > new > > >> >> connections, > > >> >> >>> - decrease quorum size again > > >> >> >>> > > >> >> >>> On Tue, Dec 20, 2011 at 12:26 PM, Camille Fournier < > > >> cami...@apache.org > > >> >> >wrote: > > >> >> >>> > > >> >> >>>> How are your clients configured to find the zks now? How > many > > >> clients > > >> >> do > > >> >> >>>> you have? > > >> >> >>>> > > >> >> >>>> From my phone > > >> >> >>>> On Dec 20, 2011 3:14 PM, "Neha Narkhede" < > > neha.narkh...@gmail.com> > > >> >> wrote: > > >> >> >>>> > > >> >> >>>> > Hi, > > >> >> >>>> > > > >> >> >>>> > As part of upgrading to Zookeeper 3.3.4, we also have to > > migrate > > >> our > > >> >> >>>> > zookeeper cluster to new hardware. I'm trying to figure > out > > the > > >> best > > >> >> >>>> > strategy to achieve that with no downtime. > > >> >> >>>> > Here are some possible solutions I see at the moment, I > could > > >> have > > >> >> >>>> > missed a few though - > > >> >> >>>> > > > >> >> >>>> > 1. Swap each machine out with a new machine, but with > the same > > >> >> host/IP. > > >> >> >>>> > > > >> >> >>>> > Pros: No client side config needs to be changed. > > >> >> >>>> > Cons: Relatively tedious task for Operations > > >> >> >>>> > > > >> >> >>>> > 2. Add new machines, with different host/IPs to the > existing > > >> >> cluster, > > >> >> >>>> > and remove the older machines, taking care to maintain > the > > >> quorum at > > >> >> >>>> > all times > > >> >> >>>> > > > >> >> >>>> > Pros: Easier for Operations > > >> >> >>>> > Cons: Client side configs need to be changed and clients > need > > to > > >> be > > >> >> >>>> > restarted/bounced. Another problem is having a large > quorum > > for > > >> >> >>>> > sometime (potentially 9 nodes). > > >> >> >>>> > > > >> >> >>>> > 3. Hide the new cluster behind either a Hardware load > balancer > > >> or a > > >> >> >>>> > DNS server resolving to all host ips. > > >> >> >>>> > > > >> >> >>>> > Pros: Makes it easier to move hardware around in the > future > > >> >> >>>> > Cons: Possible timeout issues with load balancers > messing with > > >> >> >>>> > zookeeper functionality or performance > > >> >> >>>> > > > >> >> >>>> > Read this and found it helpful - > > >> >> >>>> > > > >> >> >>>> > > > >> >> >>>> > > >> >> > > >> > > > http://apache.markmail.org/message/44tbj53q2jufplru?q=load+balancer+lis > t:org%2Eapache%2Ehadoop%2Ezookeeper-user&page=1 > > >> >> >>>> > But would like to hear from the authors and the users > who > > might > > >> have > > >> >> >>>> > tried this in a real production setup. > > >> >> >>>> > > > >> >> >>>> > I'm very interested in finding a long term solution for > > masking > > >> the > > >> >> >>>> > zookeeper host names. Any inputs here are appreciated ! > > >> >> >>>> > > > >> >> >>>> > In addition to this, it will also be great to know what > people > > >> think > > >> >> >>>> > about options 1 and 2, as a solution for hardware > changes in > > >> >> >>>> > Zookeeper. > > >> >> >>>> > > > >> >> >>>> > Thanks, > > >> >> >>>> > Neha > > >> >> >>>> > > > >> >> >>>> > > >> >> > > >> > >
