wrt defining interface stability we should adopt something like hadoop is now doing:


Mahadev, do you know if this is documented somewhere? "final" documentation, rather than the running commentary thats on this jira? We could adopt something similar/same. Can you create a jira for that?


On 08/11/2010 08:23 AM, Thomas Koch wrote:
Hallo Mahadev,

thank you for your nice answer. Yes, we'll of cause preserve compatibility.
Otherwise there is no chance to get accepted.

I assume the following things must keep their interfaces:
ZooKeeper (It'll call the new interface in the background), ASyncCallback,
We may want to change: ClientCnxn (faktor out some things, remove dep on

I think other classes should not be involved at all in our issues. My collegue
Patrick was so kind to fill the jira issues.

Best regards,


Mahadev Konar:
Also, I am assuming you have backwards compatability in mind when you
suggest these changes right?

The interfaces of zookeeper client should not be changing as part of this,
though the recursive delete hasn't been introduced yet (its only available
in 3.4, so we can move it out into a helper class).


On 8/11/10 7:40 AM, "Mahadev Konar"<maha...@yahoo-inc.com>  wrote:

HI Thomas,
   I read through the list of issues you posted, most of them seem
reasonable to fix. The one's you have mentioned below might take quite a
bit of time to fix and again a lot of testing! (just a warning :)). It
would be great if you'd want to clean this up for 3.4. Please go ahead and
file a jira. These improvements would be good to have in the zookeeper
java client.

For deleteRecursive, I definitely agree that it should be a helper class. I
don't believe it should be in the direct zookeeper api!


On 8/11/10 2:45 AM, "Thomas Koch"<tho...@koch.ro>  wrote:


I started yesterday to work on my idea of an alternative ZooKeeper client
interface.[1] Instead of methods on a ZooKeeper class, a user should
instantiate an Operation (Create, Delete, ...) and forward it to an
Executor which handles session loss errors and alikes.

By doing that, I got shocked by the sheer number of WTF issues I found. I'm
sorry for ranting now, but it gets quicker to the poing.

- Hostlist as string

The hostlist is parsed in the ctor of ClientCnxn. This violates the rule of
not doing (too much) work in a ctor. Instead the ClientCnxn should receive
an object of class "HostSet". HostSet could then be instantiated e.g. with
a comma separated string.

- cyclic dependency ClientCnxn, ZooKeeper

ZooKeeper instantiates ClientCnxn in its ctor with this and therefor builds
a cyclic dependency graph between both objects. This means, you can't have
the one without the other. So why did you bother do make them to separate
classes in the first place?
ClientCnxn accesses ZooKeeper.state. State should rather be a property of
ClientCnxn. And ClientCnxn accesses zooKeeper.get???Watches() in its method
primeConnection(). I've not yet checked, how this dependency should be
resolved better.

- Chroot is an attribute of ClientCnxn

I'd like to have one process that uses ZooKeeper for different things
(managing a list of work, locking some unrelated locks elsewhere). So I've
components that do this work inside the same process. These components
should get the same zookeeper-client reference chroot'ed for their needs.
So it'd be much better, if the ClientCnxn would not care about the chroot.

- deleteRecursive does not belong to the other methods

DeleteRecursive has been committed to trunk already as a method to the
zookeeper class. So in the API it has the same level as the atomic
operations create, delete, getData, setData, etc. The user must get the
false impression, that deleteRecursive is also an atomic operation.
It would be better to have deleteRecursive in some helper class but not
that deep in zookeeper's core code. Maybe I'd like to have another policy
on how to react if deleteRecursive fails in the middle of its work?

- massive code duplication in zookeeper class

Each operation calls validatePath, handles the chroot, calls ClientCnxn and
checks the return header for error. I'd like to address this with the
operation classes:
Each operation should receive a prechecked Path object. Calling ClientCnxn
and error checking is not (or only partly) the concern of the operation
but of an "executor" like class.

- stat is returned by parameter

Since one can return only one value in java it's the only choice to do so.
Still it feels more like C then like Java. However with operator classes
one could simply get the result values with getter functions after the

- stat calls static method on org.apache.zookeeper.server.DataTree

It's a huge jump from client code to the internal server class DataTree.
Shouldn't there rather be some class related to the protobuffer stat class
that knows how to copy a stat?

- Session class?

Maybe it'd make sense to combine hostlist, sessionId, sessionPassword and
sessionTimeout in a Session class so that the ctor of ClientCnxn won't get
too long?

I may have missed some items. :-)

Once again, please excuse my harsh tone. May I put the above issues in jira
and would you accept (backwards compatible) patches for it for 3.4.0?

Zookeeper is a fascinating project. Cudos to the devs. I've only looked in
the client side code, which is what most users of zookeeper will ever see
if they see any zookeeper internal code at all. So it may make sense to
make this piece of the project as nice and clean as possible.

[1] http://mail-archives.apache.org/mod_mbox/hadoop-zookeeper-

Best regards,

Thomas Koch, http://www.koch.ro

Reply via email to