Yup, I am aware of the fact that 2 is a particularly bad number for cluster
size and hopefully we should fix that soon, I was just hoping that for some
reason that was why the problem is occurring - my conjecture was, for e.g.
if the two zk servers disagree about the version there is no way to decide
who is correct without a third tie-breaker server.
But, if you say that is not the case, then I need to keep looking (sigh).
I am pretty sure that only one thread is touching that znode. We put in some
trace logging to try and pinpoint the problem and noticed that every time we
get the BadVersionException the actual version on the znode is one more than
what we expected it to be based on the previous "exists()" call.
As I said, this code gets called once every 2 seconds (or thereabouts). It
seems to fail with a BadVersionException about 3 times an hour (on average).
By the way, not sure if it is relevant, but the reason we are using 2 nodes
in the cluster and the reason why their version is 3.2.2 is because they are
the ZKs that come embedded inside HBase (we're running 2 Hbase
regionservers) - I've been meaning to pull them out and run them standalone
but just haven't got around to it (yet).
On 10 October 2011 17:35, Mahadev Konar <[EMAIL PROTECTED]> wrote:
> 2 ZK servers is definitely not the right number for running a ZK
> service but its no reason to get a Badversion exception because of
> that. For more information on the size of the ZK ensemble take a look
> As for the version on the znode, can you try reading the version when
> you get a setData/BadException?
> Also, is there any chance of a delete on the znode that removes it and
> another create happens for the same path?
> I dont think we have seen this version issue in the releases, so I'd
> be inclined to say that there could be something in the code thats
> making some changes to the znode before you set the data.
> Hope that helps
> On Fri, Oct 7, 2011 at 6:47 PM, Ishaaq Chandy <[EMAIL PROTECTED]> wrote:
> > Hi all,
> > We're seeing a puzzling error. Here's the scenario:
> > 1. We have a single thread that wakes up every two seconds (give or take)
> > and does some work
> > 2. As part of that work it updates a node on ZK. When it does this it
> > gets the Stat of the existing node and uses the version retrieved from it
> > update the value.
> > 3. There are no other processes updating the node
> > The code goes something like this:
> > final Stat stat = zooKeeper.exists(path, false);
> > // do some other work here to create the path if it does not exist - this
> > code only ever gets called once
> > zooKeeper.setData(path, value, stat.getVersion());
> > What we're seeing is that every so often (once every 5 minutes or so?) is
> > that that setData() call fails with a BadVersionException. This is very
> > unexpected because, as I mentioned previously, this thread is the sole
> > updater of that node.
> > One possibility I am considering is that we are using the wrong number of
> > ZKs in our cluster - i.e 2 nodes. I am wondering if 2 is the worst number
> > nodes possible for ZK as there is no way to resolve a disagreement.
> > Another possibility is that we are using an old version of ZK (3.2.2),
> > perhaps there is a known bug with it? Though I see nothing related to
> > in the release logs for subsequent versions.
> > Thoughts/suggestions?
> > Thanks,
> > Ishaaq