Hi Daidong, In addition to Flavio's response, I'll try to address some of your specific questions.
> In my opinion, an atomic broadcast protocol must guarantee all the non- > faulty servers have the same status eventually. So in the 2PC protocol, > the coordinator must block until "all" the servers reply "ok". Designed this way, the protocol wouldn't be able to tolerate any failures - the leader could block waiting for a response from a server that had crashed. The idea is to receive enough "ok" messages to guarantee that even if a minority of servers crash, the information is still not lost. That's why the leader waits for a majority of acks. Messages are still sent to all followers, so they will eventually get them (or if they disconnect they will later reconnect and synch with the leader automatically). Regarding your second question - formally, sequential consistency guarantees that operations of each client take effect in the order they were submitted by the client - so a client's read is guaranteed to see its own last complete write. In the example you mention, the client first executes a create() and then getChildren(). If clients C1 and C2 both submit a create() concurrently, one of these requests will reach the leader and will be scheduled by the leader before the other one, suppose the create() request of C1. Then, when C2 is notified about the completion of its own create, FIFO ensures that it also finds out about any operation that completed before that create() (these messages were sent by the leader earlier). So when C2 finally runs getChildren(), its local state will already have every operation that was scheduled by the leader before its own create() completed. In general, ZAB implements state-machine replication by executing consensus on each operation. To understand the general idea, I recommend reading Lamport's "Paxos made simple" paper I sent earlier - it has a constructive explanation of this (although the algorithm is somewhat different from ZAB). Alex > -----Original Message----- > From: daidong [mailto:] > Sent: Wednesday, April 20, 2011 11:31 PM > To: zookeeper-u...@hadoop.apache.org > Subject: Re: RE: Problems about Zab protocol > > Hi, Alex > > Thanks for your reply. :) > > I knew ZAB has two modes, but things i do not quit understand focus on > the broadcast mode. In the ZAB paper, authors said ZAB is a simple > version of two phases commit protocol because we don't have abort > actions in followers. I do not quit understand this. > > In my opinion, an atomic broadcast protocol must guarantee all the non- > faulty servers have the same status eventually. So in the 2PC protocol, > the coordinator must block until "all" the servers reply "ok". If there > is not any abort too, consider the situation that we have a very slow > follower F who processes messages slower than other followers. > According TCP and FIFO channel, We can say all the messages will be > processed orderly in F, however, the messages will assemble if > coordinator continues to broadcasting. What happens if the receive > buffer in F is overflow? > > Is there any mechanism i have not noticed to avoid this situation in > ZAB? > > About my second questions, I read the consistency guarantees section, > thanks for your tips. I still have a question, if zookeeper do not make > sure that all the clients will see the latest value, how the lock > mechanism works? i checked the recipe example code in Zookeeper 3.3.3, > when a client try to get the write lock, it does not sync() before call > getChildren(). If other client has created a ephemeral node with the > lowest number suffix, this client does not get this information as > getChildren() do not sync with leader. Is there any possibility that > two clients will think they both got the lock? > > Thanks for any words. :) > -- > daidong > Sent with Sparrow > On 2011年4月21日星期四 at 上午2:30, Alexander Shraer [via zookeeper- > user] wrote: > > Hi, > > > > Regarding your first question - ZAB has two parts - the broadcast > protocol you mention, > > which is executed by a leader, and the leader election protocol, > which recovers from a leader failure. > > This is similar to the way other state-machine replication algorithms > work, where you have > > a fast normal mode and a slower recovery mode (you don't need to > execute both all the time - only when the leader fails). > > See Paxos state-machine replication for example (section 3): > http://research.microsoft.com/en- > us/um/people/lamport/pubs/pubs.html#paxos-simple > > > > Regarding your second question - Zookeeper basically guarantees so > called "sequential consistency" semantics. > > This guarantees that the real execution looks to clients like some > sequential execution in which > > the operations of every client appear in the order they were > submitted. It does not guarantee that a read of one client > > returns the latest value written by another client. This allows reads > to be executed locally. If you need to return the latest > > state, you can use the sync() call which flushes the pending updates > between the leader and a follower. > > See also the "consistency guarantees" section here: > > > http://hadoop.apache.org/zookeeper/docs/r3.3.1/zookeeperProgrammers.htm > l > > > > Alex > > > > > -----Original Message----- > > > From: daidong [mailto:[hidden email]] > > > Sent: Wednesday, April 20, 2011 2:38 AM > > > To: [hidden email] > > > Subject: Problems about Zab protocol > > > > > > Hi, everyone. > > > > > > Recently, i read the paper "a simple total ordered broadcast > protocol" > > > and > > > there are some problems i can not figure out. Hope anyone can help > > > me... :P > > > > > > The paper describes the Zab protocol as a 2 phase commit protocol > when > > > system is under broadcast mode. However some paper(Skeen 82, "A > Quorum > > > Based > > > Commit Protocol") has mentioned if we want to extend an 2PC to > adapt a > > > quorum based commit protocol we must introduce a three phase commit > > > protocol(In fact, i haven't quit understood this, :( ). However > > > according > > > Zab paper, this still can be done. Why and how to do this? > > > > > > Secondly, even Zookeeper can guarantee that status in different > > > followers > > > are consistent. However, this consistency only works among a quorum > of > > > followers that has acked the COMMIT. As the client can connect to > any > > > followers when perform reading action, so what happens if the > client > > > happens > > > to connect with the follower that has not acked the COMMIT? I can > not > > > find > > > the information in this paper... > > > > > > If i ask some naive question, Hope anybody can tell me where i can > find > > > the > > > answer or some suggestions, thanks :) > > > > > > > > > -- > > > View this message in context: http://zookeeper- > > > user.578899.n2.nabble.com/Problems-about-Zab-protocol- > > > tp6290102p6290102.html > > > Sent from the zookeeper-user mailing list archive at Nabble.com. > > > > > > If you reply to this email, your message will be added to the > discussion below: http://zookeeper-user.578899.n2.nabble.com/Problems- > about-Zab-protocol-tp6290102p6291775.html > > To unsubscribe from Problems about Zab protocol, click here. > > > > > > > > > > -- > View this message in context: http://zookeeper- > user.578899.n2.nabble.com/Problems-about-Zab-protocol- > tp6290102p6293369.html > Sent from the zookeeper-user mailing list archive at Nabble.com.
