|Subject:||retreat from zookeeper|
|From:||Thomas Koch (tho...@koch.ro)|
|Date:||May 19, 2011 2:21:08 am|
as you may have noticed, I haven't been active in the ZooKeeper project anymore for a couple of months. I'm a full time student again since march so that any further activity in Hadoop/ZooKeeper would need to be auto-motivated.
Since I don't want to just fade away and I'll still give a talk about ZooKeeper on the BerlinBuzzWords conf (Berlin, june 6/7), I listed the reasons why I wouldn't like to work on the current ZooKeeper code base anymore.
I plan the following structure for my talk:
1) theoretical model / protocol of ZooKeeper 2) practical applications, projects using ZooKeeper 3) shortcomings of the current ZooKeeper code base
A tentative brain dump of part three is listed below. I appreciate any comments that could help me to give a balanced presentation of the ZooKeeper project.
If I'd need a ZooKeeper implementation right now I'd probably do a minimal- feature rewrite in Scala + Akka. I do appreciate ZooKeeper as an invaluable proof-of-concept implementation and pioneer. But as in american history there should come others after the pioneers that don't look like Clint Eastwood anymore and build more tidy things.
* The code is tightly coupled * most so called "Unit-Tests" are actualy integration tests. They run the whole application and test one specific functionality.
* no uniform configuration: command line parameters, system properties, configuration file (java properties) * configuration properties copied to static class members
* feature bloat on fragile foundation: e.g. chroot + automatic resubscribtion does not work
* implementation unlike specification: allowed characters in path
* still on ant instead of maven (depends how you see ant vs. maven)
* circular object dependencies (e.g. ZooKeeper <-> ClientCnxn)
* methods with +100 lines of code and nested conditions depth well over 5
* general attitude against refactoring, no knowledge or appreciation of "effective java" (Josh Bloch) or "clean code" (Robert C. Martin)
* magic numbers instead of enum
* still bound to inline copy of jute (HadoopIO, avro predecessor) * even hand coded (de)serialization in leader election
* no client-only jar. Every client gets the full server code.
* unhandy API triggered (at least) two client API wrappers: zkClient, cages
* insane amounts of code duplication
* horrible, fragile thread programming: plenty of "XYZ extends Threads" instead of - implements runnable - or better: executor framework - or much better: actors (see Akka) -> leads to fear of refactoring, because nobody understands all synchronization needs.
Thomas Koch, http://www.koch.ro