Also, as you pointed out, ZAB requires this FIFO property of the
point-to-point links. Paxos copes with more adversarial networks which allow
reordering and missed messages. It's easy to alter Paxos so as not to
'publish' the results of consensus rounds where there are gaps in the
previous commit history. (You may be interested in the 'Fast Paxos' paper by
Lamport which talks about making the protocol 2-message optimal in all cases
when order is not important, i.e. the messages commute). You can express the
ordering dependency between messages by supplying a proposal number with
each that is monotonically increasing in causal order.
ZAB takes care of all of this for you by using TCP sequence numbers and
getting the deep pipelining available by knowing that there are no updates
being voted on depend on updates that have not yet arrived, the 'cost' is
relying on a stronger network model than Paxos presupposes.
2010/1/20 Benjamin Reed <br...@yahoo-inc.com>
> hi Qing,
> i'm glad you like the page and Zab.
> yes, we are very familiar with Paxos. that page is meant to show a weakness
> of Paxos and a design point for Zab. it is not to say Paxos is not useful.
> Paxos is used in the real world in production systems. sometimes there are
> not order dependencies between messages, so Paxos is fine.
> in cases where order is important, multiple messages are batched into a
> single operation and only one operation is outstanding at a time. (i believe
> that this is what Chubby does, for example.) this is the solution you allude
> to: wait for 27 to commit before 28 is issued.
> for ZooKeeper we do have order dependencies and we wanted to have multiple
> operations in progress at various stages of the pipeline to allow us to
> lower latencies as well as increase our bandwidth utilization, which led us
> to Zab.
> Qing Yan wrote:
>> Anyone familer with Paxos protocol here?
>> I was doing some comparision of ZAB vs Paxos... first of all, ZAB's
>> based protocol is really cool!
>> http://wiki.apache.org/hadoop/ZooKeeper/PaxosRun mentioned the
>> inconsistency case for Paxos("the state change B depends upon A, but A was
>> not committed").
>> In the "Paxos made simple" paper, author suggests fill the GAP (lost
>> machine changes) with "NO OP" opeartion.
>> Now I have some serious doubts how could Paxos be any useful in the real
>> world. yeah you do reach the consesus - albeit the content
>> is inconsistent/corrupted !?
>> E.g. on the wiki page, why the Paxos state machine allow fire off 27,28
>> concurrently where there is actually depedency? Shouldn't you wait
>> 27 to be committed before start 28?
>> Did I miss something?
>> Thanks for the enlight!