Christian, I am really not an expert on Hive, but from what I know Hive translates the SQL into 1 or more Map/Reduce jobs to execute the query. It does optimizations to try and reduce the number of jobs that it launches and to try and speed things up. I also know that Pig has looked into supporting some alternative paradigms similar to spark to speed up processing of small jobs, but that is still in the discussion stages and I have no idea where that will end up. I can only assume that Hive is also looking into similar things.
-- Bobby On 8/19/11 1:44 PM, "Christian Schäfer" <[email protected]> wrote: Hi Bobby, thanks for the information provided :) I'm glad there are some possibilities to use hadoop+hbase....was a bit afraid a had to discard that mighty tool (in my project) As I'm still at the beginning of learning hadoop I just got one basic question: Is every query i send via hive to hbase in the background realized as a map/reduce-job or does it work in another (more efficient) kind? (I know RTFM would be an appropriate answer...but it still searched...and did not find the "answer" yet. the mesos and storm stuff looks interesting..will take it into account for my evaluation if possible. somehow I think pig + hive + cloudera tools will be implemented later because of proven tech, high level, tooling and possibility of getting support. But I will check out the spark and storm as they seem to have some interesting concepts :) regards Christian ________________________________ Von: Robert Evans <[email protected]> An: "[email protected]" <[email protected]> Gesendet: Freitag, den 19. August 2011, 17:35:08 Uhr Betreff: Re: Hadoop Case Studies with interactive applications...an antagonism? Christian, Hadoop is best for batch processing because it is optimized for that use case. It is not that it cannot handle small jobs. Those jobs tend to be some what slower then other systems and also not as consistent in their processing time as some use cases really need. You can get around this some what by over provisioning your grid. If you want to do monitoring of sensor data Hadoop should be able to handle it, so long as your SLAs are not extremely tight. This is especially true as the size of your data grows. You might want to look at HBase. It can be very fast and interactive, and because it stores the data in HDFS you can process it with Map/Reduce if you need to. There are a number of interactive/fast processing solutions on top of HDFS too that are either available now or should be soon once MRV2 stabilizes some more. Look at Spark which is part of the mesos project at Berkley (www.mesosproject.org). Another thing to look at is Hive or Pig if you want to be able to query the data with a higher level language. Another solution that looks very interesting once it is released as open source is storm http://engineering.twitter.com/2011/08/storm-is-coming-more-details-and-plans.html It looks like it could be modified a bit to run under YARN (MRV2) and then you can store your modules state in HBase. That would compliment Hadoop's MapReduce processing very nicely and do a lot of what you are looking at doing in real time. --Bobby On 8/19/11 8:06 AM, "Christian Schäfer" <[email protected]> wrote: Hi Hadoopians, I'm a noob in hadoop (what a rhyme) ....and got some questions relating to the white papers posted on cloudera.com as follows: in IQT QUARTERLY: HADOOP: Scalable, Flexible Data Storage and Analysis - By Mike Olson I got an antagonism when comparing case studies and the following pros&cons of hadoop. pros: hadoop(M/R) mostly used in batch operation (running mins or hours to complete) cons: hadoop(M/R) not usable for interactive applications and the case study: OpenPDC where it is used for monitoring and to be able to react quickly: "Close monitoring and rapid response to changes in the state of the grid allow utilities to minimize or prevent blackouts," another case study from "Ten Common Hadoopable Problems - Real-World Hadoop Use Cases": "Fast detection allows the bank to protect itself from considerable losses." If there is a better non-commercial place to ask this questions please let me know. Background: I'm intending to set up a system for another domain where lots of sensordata need to be stored and queried to implement monitoring an detect problem situations kind regards Christian
