Admin functionality in 0.8 is kind of messy. What should we do?
There are a bunch of misc. command line tools under kafka.admin. AdminUtils is the closest thing we have to an admin interface, but it is weirdly implemented as all static methods. Furthermore its APIs are really wonky and tend to describe the operation they perform in zookeeper rather than the high-level thing they accomplish.
Since people are going to want programmatic access to administrative functionality it would be good to think this through a bit. The least we could do would be to refactor AdminUtils into an Admin class and think through those APIs enough to make them usable from java/scala. This would mean having a clear API for each thing that the various tools currently do (create/delete/alter/describe/list topics, migrate partitions, shutdown broker, etc).
Arguably some of these should really be RPC apis so other languages can invoke basic operations like creating and deleting topics. But this is not critical so we can probably do it later.
I think it would be good to also think about how we ended up here. I think we are not putting in place basic APIs that provide a layer of simplification for common operations. Instead we tend to just be jabbing directly at Zookeeper from different places using ad hoc methods in ZkUtils. This is really hard to understand or change. In general we seem to have trouble thinking through APIs.
One thought I had on some of these things was that it is actually kind of funny that we implement them in zookeeper. Technically we could implement them in Kafka itself which would mean less code (potentially) and also a clean api to clients.
The idea is that many of the admin apis are basically changing cluster state. Examples are adding a topic, deleting a topic, altering the log config for a topic, changing the replication factor for a topic, migrating partitions, etc.
One way to implement these would be to define an admin topic in kafka. Then standardize a JSON format or something for these command type messages. The API to make these changes would just be to publish a message to this topic. This would allow all brokers to subscribe and carry out these commands in order, and would allow any client that has support for the producer api to publish commands.
I think in addition to providing an API this is actually the functionality needed for most of these cluster changes. For example the dynamic config patch I posted is basically implementing a queue like this but in zk for subscribing to and applying config changes.
This is not fully thought through, but thought I would just toss the idea out there.
-Jay On Fri, Feb 8, 2013 at 3:12 PM, Jay Kreps <[EMAIL PROTECTED]> wrote:
The term admin functionality itself means something that only an admin should be able to do. We had been through a discussion on admin functionality earlier and reviewed the following options -
1. Expose admin commands as RPC and in addition to that, provide command line tools that talk to the controller using this RPC. This way non-java clients also have an option of executing the admin commands from code. 2. Expose admin commands as command line tools only and notify the controller using zookeeper. The argument is that since these are "admin" operations, there is no reason someone should be allowed to fire an API request and change cluster wide metadata, configuration or state. Another assumption is that since admin operations are limited, this won't create too much zookeeper state.
The reason we picked option #1 is because it seemed reasonable for admin operations to be available through command line tools and to disallow random clients from being able to change cluster state.
However, it's not the end of the world, if we think our users like option #2 now, we can think about redesigning our admin functionality. It really depends on what most people think the meaning of admin operations is or should be.
On Monday, February 11, 2013, S Ahmed wrote:
NEW: Monitor These Apps!
All projects made searchable here are trademarks of the Apache Software Foundation.
Service operated by Sematext