Hi Patrick,

On 01.10.2009, at 08:57, Patrick Hunt wrote:

I started looking a bit more closely at the source, some questions:

1) I tried generating the javadocs (see my fork of the project on github if you want my changes to build.xml for this) but it looks like there's pretty much no javadoc. Some information, particularly on semantics of user-exposed operations would be useful (esp re my earlier README comment - some high level document describing the benefits, etc... of the library)

If I'm your proto-typical "lazy developer" (which I am :-) ), I'm really expecting some helpful docs to get me bootstrapped.

2) what purpose does ZkEventThread serve?

ZkClient updates it's connection state from the ZooKeeper events. Based on these it notifies listeners, updates it's connection state or reconnects to ZooKeeper. ZkClient has its own event thread to prevent dead-locks. When a listener blocks (because it waits until ZkClient has reconnected to Zookeeper), ZkClient wouldn't be able to receive the reconnect event from ZooKeeper anymore, if we had re-used the Zookeeper event thread to notifier listeners. See the javadoc for ZkEventThread for more information.

3) there's definitely an issue in the retryUntilConnected logic that you need to address

let's say you call zkclient.create, and the connection to the server is lost while the request is in flight. At this point ConnectionLoss is thrown on the client side, however you (client) have no information on whether the server has made the change or not. The retry method's while loop will re-run the create (after reconnect), and the result seen by the caller (user code) could be either OK or may be NODEEXISTS exception, there's no way to know which.

Mahadev is working on ZOOKEEPER-22 which will address this issue, but that's a future version, not today.

Good catch. I wasn't aware that nodes could still be have been created when receiving a ConnectionLoss. But how would you deal with that? If we create a znode and get a ConnectionLoss exception, then wait until the connection is back and check if the znode is there. There is no way of knowing whether it was us who created the node or somebody else, right?
Anyway. That's definitely a design issue.

4) when I saw that you had separated zkclient and zkconnection I thought "ah, this is interesting" however when I saw the implementation I was confused:

a) what purpose does this separation serve?

It's just to have all ZooKeeper communication in one place, where the higher lever stuff is in ZkClient. That way we are able to provide an in-memory ZkConnection implementation that doesn't connect to a real ZooKeeper. This could be used for easier testing.

b) I thought it was to allow multiple zkclients to share a single connection, however looking at zkclient.close, it closes the underlying connection.

Actually each ZkClient instance maintains one ZooKeeper connection.

5) there's a lot of wrapping of exceptions, looks like this is done in order to make them unchecked. Is this wise? How much "simpler" does it really make things? Esp things like interrupted exception? As you mentioned, one of your intents is to simplify things, but perhaps too simple? Some short, clear examples of usage would be helpful here to compare/contrast, I took a very quick look at some of the tests but that didn't help much. Is there a test(s) in particular that I should look at to see how zkclient is used, and the benefits incurred?

Checked exceptions are very painful when you are assembling together a larger number of libraries (which is true for most enterprise applications). Either you wind up having a general "throws Exception" (which I don't really like, because it's too general) at most of your interfaces, or you have to wrap checked exceptions into runtime exceptions.

We didn't want a library to introduce yet another checked exception that you MUST catch or rethrow. I know that there are different opinions about that, but that's the idea behind this.

Similar situation for the InterruptedException. ZkClient also converts this to a runtime exception and makes sure that the interrupted flag doesn't get cleared. There are just too many existing libraries that have a "catch (Exception e)" somewhere that totally ignores that this would reset the interrupt flag, if e is an InterruptedException. Therefore we better avoid having all of the methods throwing that exception.

Thanks a lot for the valuable feedback,



Patrick Hunt wrote:
Hi Stefan, two suggestions off the bat:
1) fill in something in the README, doesn't have to be final or polished, but give some insight into the what/why/how/where/goals/ etc... to get things moving quickly for reviewers & new users. 2) you should really discuss on the dev list. It's up to you to include user, but apache discourages use of user for development discussion (plus you'll pickup more developer insight there)
Stefan Groschupf wrote:
Hi Zookeeper developer,
it would be great if you guys could give us some feedback about our project zkclient.
The main idea is making the life of lazy developers that only want minimal zk functionality much easier.

We have a functionality like zkclient mock making testing easy and fast without running a real zkserver, simple call back interfaces for the different event types, reconnecting handling in case of timeout etc.

We feel we come closer to a release so it would be great if some experts could have a look and give us some feedback.

Hadoop training and consulting

Reply via email to