On 4 March 2014 07:21, Brian Tarbox <[email protected]> wrote:
> We're running a five node cluster in AWS and sometimes if I do a read
> immediately after doing a write ZK does not find the data I just wrote.
>
> I may just have a flawed model but is this a case of "writes go to the
> master, while reads can go to any node in the cluster?". We are thinking
> that perhaps our write went to the master and then our read went to a
> follower that hadn't gotten the write yet?
>
Perhaps you can call sync() to ensure reads after writes?
> Along the same lines...when does a watch fire? After the leader has the
> change or after all followers have gotten the change?
>
The pipeline is, in a poor's man diagram, as follow:
client → write request → zk_server_1
↓
↓
write request is sent to the leader
↓
↓
leader fans out a transaction proposal to all followers
↓
↓
↓ ↓ ↓ ↓
follower 1 follower 2 follower 3 follower 4
↓
↓
followers, non-synchronously, acknowledge the proposal
↓
↓
when the leader gets a sufficient number of acks (from a "majority")
it fans out COMMIT msgs to followers
↓
↓
↓ ↓ ↓ ↓
follower 1 follower 2 follower 3 follower 4
↓
↓
followers will then apply the write which triggers watches, that in
turn will then flow to listening clients...
Hth,
-rgs
