Are simplified hadoop interfaces the next web cash cow?(brianbreslin.com)
brianbreslin.com
Are simplified hadoop interfaces the next web cash cow?
http://brianbreslin.com/are-simplified-hadoop-interfaces-the-next-web-cash-cow/
5 comments
well my thesis was based upon the premise that those who would need a UI for hadoop scale queries wouldn't be those doing the data collection. i.e. the marketers or research scientists who want to run queries against existing data sets, or data sets in use by others in their companies.
imagine someone in say the sports division at aol wanting to know patterns for users across aol who have viewed auto related info or something.
the idea is not having to bother the sophisticated db admins, and let end users figure out the data themselves.
imagine someone in say the sports division at aol wanting to know patterns for users across aol who have viewed auto related info or something.
the idea is not having to bother the sophisticated db admins, and let end users figure out the data themselves.
Amazon's Elastic MapReduce does a decent job of this, and there's a GUI for it in the form of the Console (http://console.aws.amazon.com): http://aws.amazon.com/elasticmapreduce/
EMR merely abstracts away the problem of configuring and hosting hadoop; the actual business of crunching data is still very much in the build-it-yourself realm.
There is a huge, huge market for a value-added service on top of EMR to help people understand their data.
There is a huge, huge market for a value-added service on top of EMR to help people understand their data.
from my perspective, a single user interface paradigm is unlikely to cover the variety of use cases that the full suite of hadoop platform services (flume, sqoop, hdfs, mapreduce, hive, pig, oozie, hbase) provides.
at cloudera, we've built and released into open source a user interface construction kit and an environment to house these various "applications". that way we can incorporate any metaphor you'd like: file browser, spreadsheet, yahoo pipes, etc.
see http://www.cloudera.com/blog/2010/07/whats-new-in-cdh3b2-hue and http://www.cloudera.com/blog/2010/07/developing-applications... for details.
at cloudera, we've built and released into open source a user interface construction kit and an environment to house these various "applications". that way we can incorporate any metaphor you'd like: file browser, spreadsheet, yahoo pipes, etc.
see http://www.cloudera.com/blog/2010/07/whats-new-in-cdh3b2-hue and http://www.cloudera.com/blog/2010/07/developing-applications... for details.
Interesting idea. How do you solve the problem of marshaling tera/petabytes of customer data into a system like this?
I honestly don't know. beyond physical hardware transfers (shipping a disk) I don't think the average person has a connection that could upload terabytes of data.
you'd have to assume much of the data is already in the cloud. i'm no expert on the matter, but I'm sure someone on HN could figure out the solution. I was merely proposing the idea.
you'd have to assume much of the data is already in the cloud. i'm no expert on the matter, but I'm sure someone on HN could figure out the solution. I was merely proposing the idea.
Tanenbaums stationwagon always seems to be one step ahead of fixed line capacity.
You don't, really.
Want to move PB of data? Rent a van.
Want to move PB of data? Rent a van.
Those interested might want to check out Cascading (http://www.cascading.org/), and open-source project tackling this very problem.
Cascading is nice, don't get me wrong, but you still know you're writing MR. And you're still far more involved in the details of getting data to and from than I would like.
However, I don't think doing so would be as huge or game-changing as he argues. Any company that has somehow aggregated petabytes of data probably has some reasonably competent developers and admins. Dealing with the inevitable scale-related issues will either burn someone out or make them competent. For these developers, getting Hadoop running now is probably easier than doing whatever else they've been doing to keep aggregating these petabytes of data. I went from knowing little about Hadoop to running relatively complex queries in a matter of hours through Amazon EMR and Hive (a SQL-ish layer on top of Hadoop). Basically, I gzipped and uploaded the data to an S3 bucket, downloaded a client to easily use the EMR API, wrote a simple Hive query derived from examples, and it just worked. All this was much easier than what I've had to do to keep a MySQL cluster or web application running.