Monday, September 22, 2014

Read an article on wired

It seems that there is no more traction like in my past days

Saturday, September 20, 2014

Hands on big data - Crash Course on Spark - Start 6 nodes cluster - lesson 9

One easy way to start a cluster is to leverage those created by Amplab

After creating the keywords for AWS, I created the cluster but had to add -w 600 for timeout

Deploy (about 30mins including all the data copy)


Run the interactive scala shell which will connect to the master

Run commands

Instances on AWS

Friday, September 19, 2014

Hands on big data - Crash Course on Spark - PageRank - lesson 8

Let's compute PageRank. Below you will find the definition.

Assuming to have a pairRDD of (url, neighbors) and (url, rank) where the rank is initialized as a vector of 1s, or uniformly random. Then for a fixed number of iterations (or until the rank is not changing significantly between two consecutive iterations) 

we join links and ranks forming (url, (links, rank)) for assigning to a flatMap based on the dest. Then we reduceByKey on the destination using + as reduction. The result is computed as 0.15 + 0.85 * computed rank 

The problem is that each join needs a full shuffle over the network. So a way to reduced the overhead in computation is partition with an HashPartitioner in this case on 8 partitions

 And avoid the shuffle

However, one can directly use the SparkPageRank code distributed with Spark

Here I take a toy dataset

And compute the PageRank

Other dataset are here

Thursday, September 18, 2014

Wednesday, September 17, 2014

Hands on big data - Crash Course on Spark Cache & Master - lesson 6

Two very important aspects for Spark are the use of caching in memory for avoiding re-computations and the possibility to connect a master from the spark-shell

Caching is pretty simple. Just use .persist() at the end of the desired computation. There are also option to persist an computation on disk or to replicate among multiple nodes. There is also the option to persist objects in in-Memory shared file-systems such as Tachyon

Connect a master
There are two ways for connecting a master

val conf = new SparkConf().setAppName(appName).setMaster(master)
new SparkContext(conf)
This will connect the local master with 8 cores

$ ./bin/spark-shell --master local[8]

Tuesday, September 16, 2014

Hands on big data - Crash Course on Spark - Word count - lesson 5

Let's do  simple exercise of word count. Spark will automatically parallelize the code

1. Load the bible
2. create a flatMap where every line is put in correspondence with words
3. Map the words to tuples (words, counter where counter is simple starting with 1
4. Reduce by keys, where the reduce operation is just +

Pretty simple, isn't it?

Monday, September 15, 2014

Hands on big data - Crash Course on Spark - Optimizing Transformations - lesson 4

1. Take a list of names
2. Create buckets by initials
3. Group by keys
4. Map each each initial into the set of names, get the size
5. Collect the results into the master as an array

This code is not super-optimized. For instance, we might want to force the number of partitions - the default is 64Mb chunks

Then we can avoid to count duplicates

 or better count the unique names with a reduceByKey where the reduce function is the sum

These optimizations reduce the information sent over the wire.

Found this example very instructive [Deep Dive : Spark internals], however you should add some string sanity check

and make sure that you don't run in some heap problems if you fix the partitions

You will find more transformatios here