zookeeper-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Isabel Muñoz Fernández <imu...@eui.upm.es>
Subject Re: A question about sequential order consistency that zookeeper garantees
Date Tue, 11 Nov 2014 15:32:57 GMT

> On 24 Oct 2014, at 04:15, bit1129@163.com wrote:
> 
> Thanks Ritesh for the detailed explanation.
> 
> With FiFO client order and Asynchronous Linearizability as you mentioned, Can I say that
even the same client sends two request A and B asynchronously , the server can still guarantee
to process the request A ahead of B?  If this is true, the client must do something to help
achieve this Asynchronous Linearizability
> 
> Thanks.
> 
> 
> 
> bit1129@163.com
> 
> From: Ritesh Jaltare
> Date: 2014-10-24 04:54
> To: user
> CC: bit1129@163.com
> Subject: Re: RE: A question about sequential order consistency that zookeeper garantees
> Hi bit1129
> 
> Zookeeper provides 2 basic ordering guarantees:
> 
> 1) Linearizable writes: What this means is all updates that change the state of zookeeper
are serializable and respect the precedence order. 
> 
> 2) FIFO client order: All requests from a given client are executed in the order that
they were sent by the client.
> 
> Now let us take your scenario and see how these 2 ordering guarantees interact. . Suppose
Client 1 sends request-1A and  request-1B in order  and Client2 sends request-2A and request-2B
in order. These requests can reach the zookeeper server in any order. What zookeeper guarantees
is that  request-1A and request-1B would retain the same order in the processing pipeline.
That is,  request-1B would be processed by zookeeper only after request-1A. And similarly
for Client 2 , request-2B would be processed by zookeeper only after request-2A.
> 
> So the possible valid orderings are
> 
> Request-2A -> Request-1A -> Request 2B -> Request-1B
> 
> Request-1A -> Request-2A -> Request 2B -> Request-1B
> 
> Request-1A -> Request-2A -> Request 1B -> Request-2B
> 
> and so on...
> 
> In each of thee orderings, you can see that zookeeper maintains a fifo client order.
What zookeeper offers is Asynchronous Linearizability and not sequential consistency. Asynchronous
Linearizability is a stonger condition that sequential consistency.
> 
> Thanks
> Ritesh Jaltare, Pivotal
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> On Wed, Oct 22, 2014 at 10:35 PM, Rakesh R <rakeshr@huawei.com> wrote:
> 
> Yes.
> 
> -Rakesh
> 
> From: bit1129@163.com [mailto:bit1129@163.com]
> Sent: 23 October 2014 10:43
> To: Rakesh R
> Subject: Re: RE: A question about sequential order consistency that zookeeper garantees
> 
> Thanks Rakesh for the quick reply. So, I understand that the time that clients send the
request doesn't matter,but the server will "queue" the request and process the request.
> 
> 
> 
> ________________________________
> bit1129@163.com<mailto:bit1129@163.com>
> 
> From: Rakesh R<mailto:rakeshr@huawei.com>
> Date: 2014-10-23 12:33
> To: user@zookeeper.apache.org<mailto:user@zookeeper.apache.org>; bit1129@163.com<mailto:bit1129@163.com>
> Subject: RE: A question about sequential order consistency that zookeeper garantees
> Hi,
> 
> Sequential Order Guarantee:
> ---------------------------
> 
> Server view:- It will process the requests in sequential fashion.
> 
> Say, there are two clients connected with the server, client-1 and client-2.
> Assume application had created a znode /myapp and there is a logic to do distributed
locking mechanism. Now both these clients will try to acquire lock by creating '/myapp/lock'
znode. I'm calling these requests as request-1(create /myapp/lock) and request-2(create /myapp/lock).
Again assume, server receives the requests in the order of request-2 and then request-1. For
simplicity, you can imagine there is a Queue maintained at the server and all these requests
will be put into this Queue and executes in FIFO order. First, server will pick request-2
and creates /myapp/lock znode. Also, it will send watcher notifications if anyone has registered.
Then move on to the next request-2(create /myapp/lock), while executing this will throw exception
saying NodeExistsException.
> 
> ZooKeeper maintains global ordering but how? - Every write is assigned a globally unique
id, sequentially ordered identifier called a zxid, or ZooKeeper transaction id. This guarantees
a global order to all updates in a ZooKeeper ensemble.
> 
> 
> Regards,
> Rakesh
> -----Original Message-----
> From: bit1129@163.com<mailto:bit1129@163.com> [mailto:bit1129@163.com]
> Sent: 23 October 2014 08:46
> To: user
> Subject: A question about sequential order consistency that zookeeper garantees
> 
> Hi,Zookeepers,
> 
> I'd like to ask a question about sequential order consistency that zookeeper garantees.
> 
> Say, 2 zookeeper clients want to update the same znode data one after the other.Client1
sends the request 1 second before Client2.
> What would be the end result? There is no order gurantee here, dure to network delay,
it is likely that clients1 send the request before client2, but the zookeeper servers first
execute the request from Client1 before client2
> 
> If this is not what sequential order consistency means, then what exactly does sequential
order consistencymeans, what does zookeeper exactly garantee?
> 
> Thanks.
> 
> 
> bit1129@163.com<mailto:bit1129@163.com>
> 
> 
> 
> -- 
> Ritesh Jaltare, Pivotal


Mime
View raw message