I think I'm with Bill on this.
There are useful Chukwa deployments without HBase, and I don't want to
scare anybody away by saying "a full deployment has HBase."
I would say "you can use Chukwa on either HDFS or HBase. HBase has
lower latency, and requires Hadoop 0.20. Using HDFS+Mapred works with
On Mon, Oct 11, 2010 at 4:31 PM, Eric Yang <[EMAIL PROTECTED]> wrote:
> For streaming data to Hbase 0.20.6, Hadoop 0.20 is required. If you are
> using the sequence file and mapreduce, then Hadoop 0.18+ should work fine.
> I would like to move forward with map reduce on table instead of map reduce
> on file to reduce data processing latency. Hence, I am making Hbase a
> required component for trunk for a full end to end deployment. The setup
> may sound difficult to do, but it is trivial to setup. In the long run,
> demux does not need to be a map reduce job, and demux will only get
> optimized for incremental updates, when developers start to adopt the table
> approach. Considering this migration as planning for a better future. :)
> For now, the existing Chukwa 0.4 approach is still works in trunk but I am
> not enhancing the batch model.
> On 10/11/10 1:11 PM, "Bill Graham" <[EMAIL PROTECTED]> wrote:
>> Hi Eric,
>> A read this over and have a few comments:
>> - Prerequisites say that Hadoop 0.20+ is required, which I think isn't
>> entirely true. Agents, collectors and the data processor processes all
>> can run with 0.18.3. I though HICC was the only thing that required
>> 0.20, no? Either way, we should clarify which components require what
>> - You talk about a minimal Chukwa install including HBase, which also
>> seems misleading. Will HBase be required to use Chukwa going forward?
>> If not, no need to make the barrier for entry sound higher than it
>> needs to be IMO.
>> On Sat, Oct 9, 2010 at 6:47 PM, Eric Yang <[EMAIL PROTECTED]> wrote:
>>> I have written some instructions on how to deploy Chukwa 0.5
>>> pre-release system with HBase. Instructions are available here:
>>> Suggestions and feedback are welcome.
Ari Rabkin [EMAIL PROTECTED]
UC Berkeley Computer Science Department