Fernando Padilla wrote:
One use case I have a question about, is using Hadoop to power a web
search or other query. So the full job should be done in under a second,
from start to finish.
I don't think you should be using hadoop to answer the results of a
user's search query.
you should be looking at things like SOLR (with the distributed patch),
or CloudDB/Mysql Clusters.
some good research has also been done on this..
see CRUSH by Sage Weil :- www.ssrc.ucsc.edu/Papers/weil-sc06.pdf
or the work on Chord# for wikipedia called 'onscale' :-
http://onscale.de/onscaledb.html
both would be better suited for OLTP type operations I would think.
You know, you have a huge datastore, and you have to run a query against
that, implemented as a MR query. Is there a way to optimize that use
case, where the code doesn't change, but maybe the input parameters of
the job? So a MR job could reuse the java code, and even the same JVM to
avoid all of the startup costs..
<digression>
I bet hadoop isn't built for that yet (and enough reasons not to support
it yet).. but maybe it's a usecase that shouldn't be totally ignored.
And if you think about it, this is similar to what HBase is doing, at
least the query execution part.. A dedicated MR daemon running ontop of
the Hadoop infrastructure, so you don't incur the cost of distributing
and starting fresh MR/JVM processes across the cluster.. maybe someone
would want to refactor this thought process a little bit..
</digression>
Matt Kent wrote:
We use Hadoop in a similar manner, to process batches of data in
real-time every few minutes. However, we do substantial amounts of
processing on that data, so we use Hadoop to distribute our computation.
Unless you have a significant amount of work to be done, I wouldn't
recommend using Hadoop because it's not worth the overhead of launching
the jobs and moving the data around.
Matt
On Tue, 2008-06-24 at 13:34 +1000, Ian Holsman (Lists) wrote:
Interesting.
we are planning on using hadoop to provide 'near' real time log
analysis. we plan on having files close every 5 minutes (1 per log
machine, so 80 files every 5 minutes) and then have a m/r to merge it
into a single file that will get processed by other jobs later on.
do you think this will namespace will explode?
I wasn't thinking of clouddb.. it might be an interesting alternative
once it is a bit more stable.
regards
Ian
Stefan Groschupf wrote:
Hadoop might be the wrong technology for you.
Map Reduce is a batch processing mechanism. Also HDFS might be critical
since to access your data you need to close the file - means you might
have many small file, a situation where hdfs is not very strong
(namespace is hold in memory).
Hbase might be an interesting tool for you, also zookeeper if you want
to do something home grown...
On Jun 23, 2008, at 11:31 PM, Vadim Zaliva wrote:
Hi!
I am considering using Hadoop for (almost) realime data processing. I
have data coming every second and I would like to use hadoop cluster
to process
it as fast as possible. I need to be able to maintain some guaranteed
max. processing time, for example under 3 minutes.
Does anybody have experience with using Hadoop in such manner? I will
appreciate if you can share your experience or give me pointers
to some articles or pages on the subject.
Vadim
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
101tec Inc.
Menlo Park, California, USA
http://www.101tec.com
