Good introduction. Spark is really a project to watch in the data analysis field on distributed architecture. We had performed several benchmarks and Spark keeps its promisses. 2.5x faster comparing to Pig for the same algorithm on the same cluster.
For iterative algorithm with the in-memory possibilities, performances are really good comparing to Hadoop.
The project is still young with several bugs but the documentation is really good and the code is well commented and robust.
As part of our work we have done extensive comparisons of Spark on various workloads, clusters and cluster sizes comparing with Hadoop Map Reduce, Naiad and several other frameworks. We've found Spark to be temperamental, hard to configure, and with wildly varying performance, suited only to a small set of computations for which in-memory state reuse is beneficial (mostly it isn't).
MSFT killed Naiad's predecessor Dryad in favor of Hadoop some time ago, because Hadoop was becoming popular. The primary author linked in the page works at "Microsoft Silicon Valley"which was just shut down and in fact now lists himself on LinkedIn as "Researcher At Large, previously at Microsoft"
So, how do we know Naiad has much future? . Technologically, it may be better/more reliable/faster, but if it's a niche product that gets desupported just because it never took off... it doesn't really matter.
Spark on the other hand has a great deal of momentum and in my experience, momentum and adoption trump technical elegance in the short run...
(don't get me wrong: I thought Dryad was awesome. Google's Flume is very similar in some ways. MapReduce's days are numbered except for a small number of problems which can't be easily ported).
All good points. I can't say what the future of Naiad is. What they have done to Microsoft Research Silicon Valley is disgusting (I worked there too for a short time).
In our experiences the performance claims with Spark have been more hype than substance. Naiad on the other hand has been hard to find a corner case for.
Naiad is open source licensed under an Apache License so one can only hope...
If you are interested in this, you might be interested in my (warning: shameless plug) 5 part blog series located at http://blog.caseystella.com/pyspark-openpayments-analysis.ht.... I'm using the python bindings for Spark to illustrate doing data analysis on healthcare financial data on Hadoop.
I've just start using this at work. It's far easier to jump into than MapReduce; orders a magnitude easier. Hopefully I'll be able to contribute back to the project at some point.
Spark is nice, but its memory model almost requires a full cluster overhaul. We looked at it at my last project. Our cluster nodes only had 64 GB of RAM. That was carved into 4 GB workers. In order to use Spark we'd have to halve our number of workers because of the memory requirements.
Neat project. Has its place. Requires a different cluster configuration which might limit its utility.
13 comments
[ 2.6 ms ] story [ 39.7 ms ] threadFor iterative algorithm with the in-memory possibilities, performances are really good comparing to Hadoop.
The project is still young with several bugs but the documentation is really good and the code is well commented and robust.
In nearly every test Naiad has beaten Spark.
More info on Naiad: http://research.microsoft.com/en-us/projects/naiad/
So, how do we know Naiad has much future? . Technologically, it may be better/more reliable/faster, but if it's a niche product that gets desupported just because it never took off... it doesn't really matter.
Spark on the other hand has a great deal of momentum and in my experience, momentum and adoption trump technical elegance in the short run...
(don't get me wrong: I thought Dryad was awesome. Google's Flume is very similar in some ways. MapReduce's days are numbered except for a small number of problems which can't be easily ported).
In our experiences the performance claims with Spark have been more hype than substance. Naiad on the other hand has been hard to find a corner case for.
Naiad is open source licensed under an Apache License so one can only hope...
Also, it appears to be tied to Windows (it's delivered as a VS solution).
Neat project. Has its place. Requires a different cluster configuration which might limit its utility.