| I believe the goal of the examples was never to be a complete solutions to barriers or queues, but just to give a quick bootstrap to beginners. It is true, though, that the documentation page does not make that claim, and can be misleading.
I see two possible action points out of this discussion: 1- State clearly in the beginning that the example discussed is not correct under the assumption that a process may finish the computation before another has started, and the example is there for illustration purposes; 2- Have another example following the current one that discusses the problem and shows how to fix it. This is an interesting option that illustrates how one could reason about a solution when developing with zookeeper.
If you are interested in helping us fix it, Semih, then you could perhaps create a jira and assign yourself to fix it. I can help you out.
-Flavio
On Mar 7, 2011, at 11:23 AM, Semih Salihoglu wrote: Hi Mahadev, Sorry for the late response. I agree, actually in this other documentation http://hadoop.apache.org/zookeeper/docs/r3.0.0/recipes.html, where there is only the pseudo-code, I think this situation is avoided. Here there is another znode /ready that all nodes have a watch on. And after each node writes their own ephemeral child, they don't wait. They read how many of has been written and the last one writes the /ready znode and everyone wakes up. The only race condition in this one is that there can be two nodes trying to write /ready and only one of them will succeed but this is ok. Thank you again, semih On Sat, Mar 5, 2011 at 6:41 PM, Mahadev Konar < [email protected]> wrote: Semih,
You pointed it out right. It is possible ot enter into a situation
like that. The recipe does have a bug. It can be fixed with the last
client creating a special znode and every node in the list watching
for that (so itll be an indication for entering the barrier). no?
thanks
mahadev
On Sat, Mar 5, 2011 at 5:06 PM, Semih Salihoglu <[email protected]>
wrote:
Hi All,
I am new to this group and to ZooKeeper. I was readin the Barrier
tutorial
in one of the ZooKeeper documentations.
http://hadoop.apache.org/zookeeper/docs/current/zookeeperTutorial.html .
A
barrier primitive is exactly how I want to use ZooKeeper. I have a
question
about this example. It's not really a ZooKeeper question, it's more a
question about the Barrier primitive I think. Here it is: In the enter
method of this Barrier implementation below
boolean enter() throws KeeperException, InterruptedException{
zk.create(root + "/" + name, new byte[0], Ids.OPEN_ACL_UNSAFE,
CreateMode.EPHEMERAL_SEQUENTIAL);
while (true) {
synchronized (mutex) {
List<String> list = zk.getChildren(root, true);
if (list.size() < size) {
mutex.wait();
} else {
return true;
}
}
}
}
could there be a race condition? Let's say there are two
machines/nodes: node1 and node2 that will use this code to synchronize
over ZK. Let's say the following steps take place:
1. node1 calls the zk.create method and then reads the number of
children, and sees that it's 1 and starts waiting.
2. node2 calls the zk.create method (doesn't call the
zk.getChildren method yet, let's say it's very slow)
3. node1 is notified that the number of children on the znode
changed, it checks that the size is 2 so it leaves the barrier, it
does its work and then leaves the barrier, deleting its node.
4. node2 calls zk.getChildren and because node1 has already left,
it sees that the number of children is equal to 1. Since node1 will
never enter the barrier again, it will keep waiting.
Could this scenario happen? If not, what is preventing this? I haven't
copied the code piece that enters barrier-does work-leaves barrier.
But in the link I pasted above, it's the barrierTest(String args[])
method.
Thank you very much in advance,
semih
flaviojunqueira research scientist [email protected]direct +34 93-183-8828 avinguda diagonal 177, 8th floor, barcelona, 08018, esphone (408) 349 3300 fax (408) 349 3301
 |
