Installing Hadoop 2 on a Mac

I've had a lot of trouble getting Hadoop 2 and yarn 2 running on my MAC.  There are some tutorials out there but they are often for
beta and alpha versions of the hadoop 2.0 family.  These are the steps I used to get Hadoop 2.2.0 working on my MAC running OSX 10.9

Note:  watch for version differences in this blog.  It was written for Hadoop 2.2.0, we are currently on 2.6.2 so that will need to be changed throughout.

Get hadoop from http://www.apache.org/dyn/closer.cgi/hadoop/common/

make sure JAVA_HOME is set (if you have Java 6 on your machine):
export JAVA_HOME=`/usr/libexec/java_home -v1.6`
(Note your Java version should be 1.7 or 1.8)

point HADOOP_INSTALL to the hadoop installation directory
export HADOOP_INSTALL=/Applications/hadoop-2.2.0

And set the path
export PATH=$PATH:$HADOOP_INSTALL/bin:$HADOOP_INSTALL/sbin

You can test hadoop is found with
hadoop -version

make sure ssh is set up on your machine:
system preferences -> sharing -> remote login is ticked

try:
ssh @localhost

where is the name you used to logon.

in $HADOOP_INSTALL/etc these are the conf files I changed.

core-site.xml

 <configuration>  
 <property>  
   <name>fs.default.name</name>  
   <value>hdfs://localhost:9000</value>  
  </property>  
 </configuration>  

hdfs-site.xml

 <configuration>  
 <property>  
   <name>dfs.replication</name>  
   <value>1</value>  
  </property>  
  <property>  
   <name>dfs.namenode.name.dir</name>  
   <value>file:/Users/Administrator/hadoop/namenode</value>  
  </property>  
  <property>  
   <name>dfs.datanode.data.dir</name>  
   <value>file:/Users/Administrator/hadoop/datanode</value>  
  </property>  
 </configuration>  

Make the directories for the namenode and datanode data (note the file above and the mkdir below will need to reflect where you  want to store the files, I've stored mine in the home directory of the Administrator user on my Mac).

mkdir -p /Users/Administrator/hadoop/namenode
mkdir -p /Users/Administrator/hadoop/datanode

hadoop namenode -format

yarn-site.xml

 <configuration>  
 <!-- Site specific YARN configuration properties -->  
 <property>  
 <name>yarn.resourcemanager.address</name>  
 <value>localhost:8032</value>  
 </property>  
 <property>  
 <name>yarn.nodemanager-aux-services</name>  
 <value>madpreduce.shuffle</value>  
 </property>  
 </configuration>  

start-dfs.sh
start-yarn.sh
jps

should give
9430 ResourceManager
9325 SecondaryNameNode
9513 NodeManager
9225 DataNode
9916 Jps
9140 NameNode

if not check log files.  If data node is not started and  you get incompatible id's error, stop everything delete datanode directory and recreate
datanode directory

try  a ls
hadoop fs -ls

if you get

ls: `.': No such file or directory

then there is no home directory in the hadoop file system.  So

hadoop fs -mkdir /user
hadoop fs -mkdir /user/<username>
where is the name you are logged onto the machine with.

now change to $HADOOP_INSTALL directory and upload a file

hadoop fs -put LICENSE.txt


finally try a mapreduce job:

cd share/hadoop/mapreduce
hadoop jar ./hadoop-mapreduce-examples-2.2.0.jar wordcount LICENSE.txt out

Mapping CQL's sets and maps to column families

In this post we are going to explore how CQL implements sets and maps in Cassandra’s column store.

(in a bizarre twist of fate, John Berryman. created this post http://www.planetcassandra.org/blog/post/understanding-how-cql3-maps-to-cassandras-internal-data--structure yesterday on the same subject, I swear I hadn't seen it when I started working on this post, yesterday as well !  It's just how it goes sometimes,  Johns post is great it has to be said !. )

In CQL version 3 wide tables have been supported through the use of sets, maps and lists.  These features have been supported since Cassandra 1.2 (http://www.datastax.com/dev/blog/cql3_collections) and should now be the de facto way of creating “wide tables”  the canonical example of sets is the use of multiple email addresses for a user .  In the relational world you might create a email address table with a foreign key pointing to the user id for each address.   This is going to cause a join just for any request that needs details of the user and their valid addresses. 

Suppose we create a simple keyspace in the usual fashion:

create keyspace Keyspace3 WITH replication = {'class':'SimpleStrategy', 'replication_factor':1};

In a  Cassandra (from 1.2) you would create a table like this:

CREATE TABLE Users (   
    id uuid Primary Key,
    name text,    
    email_addresses set) ;

(This is similar to  Sylvain Lebresne’s example here http://www.datastax.com/dev/blog/cql3_collections)

We can insert data into the table (a user with 2 email addresses like this):

insert into users(id,name,email_addresses) values (88b8fd18-b1ed-4e96-bf79-4280797cba81,'tim',{'tim@example.org','timothy@example.org'});

This user has a UUID, a name and two email addresses.   You can of course get the email addresses with a select command:

select email_addresses from Users;

which will return the addresses as a set:

email_addresses
-------------------------------------------------------
            {'tim@example.org', 'timothy@example.org'}

However, how is this implemented in the column store ?  If you had used a thrift based interface (such as Hector) you may have created the column family and had the following structure:

Id: 88b8fd18-b1ed-4e96-bf79-4280797cba81 (Key)
    name: tim
  email_address: 'tim@example.org'
  email_address: 'timothy@example.org'

but how is it implemented in CQL3 ?  If you fire up Cassandra-cli you can use the list command to see what is stored in the column family:

LifeintheAirAge:bin Administrator$ ./cassandra-cli
Connected to: "Test Cluster" on 127.0.0.1/9160
Welcome to Cassandra CLI version 2.0.0

Please consider using the more convenient cqlsh instead of CLI
CQL3 is fully backwards compatible with Thrift data; see http://www.datastax.com/dev/blog/thrift-to-cql3

Type 'help;' or '?' for help.
Type 'quit;' or 'exit;' to quit.

[default@unknown] use keyspace3;
Authenticated to keyspace: keyspace3
[default@keyspace3] list users;
Using default limit of 100
Using default cell limit of 100
-------------------
RowKey: 88b8fd18-b1ed-4e96-bf79-4280797cba81
=> (name=, value=, timestamp=1381497810072000)
=> (name=email_addresses:74696d406578616d706c652e6f7267, value=, timestamp=1381497810072000)
=> (name=email_addresses:74696d6f746879406578616d706c652e6f7267, value=, timestamp=1381497810072000)
=> (name=name, value=74696d, timestamp=1381497810072000)

So we can see the rowkey as expected and the name of the user as a name value pair (the value is in ASCII in hex in this case 746976d is tim).

But for the email_addresses the values are not set.  The values of the email addresses is encoded into the name along with the “column” schema name.  name=email_addresses:74696d406578616d706c652e6f7267  is the column name email_addresses followed by tim@example.org in ASCII hex) .  Why do this ? Why not have the name as email_addresses and the value as the the hex email address ?  One reason perhaps is because this allows us to implement maps ina similar way with out needing special cases.   Suppose we alter table to include a map, we want to store details about our user, but we don’t yet know which details the user will provide (A contrived example I’ll grant you).  You can alter the table as follows:

alter table users add details map;

and insert some details as follows:

update users set details= {'tel' : '555 232341', 'twitter' : '@andycobley'} where id =88b8fd18-b1ed-4e96-bf79-4280797cba81;

What does our column now look like? Using the list command we get :

RowKey: 88b8fd18-b1ed-4e96-bf79-4280797cba81
=> (name=, value=, timestamp=1381498805511000)
=> (name=details:74656c, value=3031333832333435303738, timestamp=1381498805511000)
=> (name=details:74776974746572, value=40616e6479636f626c6579, timestamp=1381498805511000)

You can see the map key is stored with the column name  in the name part of the column family name value  pair. So  name=details:74656c contains ‘tel’ as a ASCII hex value.  The map value is simply stored in the value part of the column family name value pair.

So, we’ve seen how CQL3’s  maps and sets map on to the column family name value pairs by storing the CQL table’s column name in the name part of the column family name value pair.  It’s quite simple and elegant really.

(as ever I’m more than happy to receive corrections or further explanations !)

Setting default java on a mac

If you have multiple versions of Java on your Mac, you can set the current working version from the command line as follows:

export JAVA_HOME=`/usr/libexec/java_home -v1.6`

This will set the version to 1.6.  You can of course see the current version with

java -version

You can see all java versions installed on your mac with:

/usr/libexec/java_home -V

(tested with macos 10.8.4 )

Minor bug with C*1.2.2 and jdk8 on a pi

This is just a quick note on getting Cassandra to run on a Raspberry pi with jdk8.  It looks like a little bug has crept into the starup script in 1.2.2  that will stop C* from running on the pi with jdk1.8   To fix this locate  conf/cassandra-env.sh and comment out the following lines

#if [ "$JVM_VERSION" \> "1.7" ] ; then                                                                      
#    JVM_OPTS="$JVM_OPTS -XX:+UseCondCardMark"                                                              
#fi

It looks like jdk8 on the pi doesn't support the UseCondCardMark option, but it can be safely removed from the startup script. I'm told that the option was added "for  better lock handling especially on hotspot with multicore processor" (see https://issues.apache.org/jira/browse/CASSANDRA-4366 )  I’ll report this as a bug and hopefully it can fixed soon in a more elegant manor.

Calculating Pi to 5 Digits

Yesterday I was reading 20 controversial programming opinions which has some interesting debates going on in it.   However I was drawn to the following programming question that one responder likes to ask potential new hires in job interviews:

Given that Pi can be estimated using the function 4 * (1 – 1/3 + 1/5 – 1/7 + …) with more terms giving greater accuracy, write a function that calculates Pi to an accuracy of 5 decimal places.

The first part of this is easy, depending on the language you use.  In Java it's something simple like:

float Pi=(float)1.0;
int mult=-1;
for (int dem=3;dem <1000 dem="dem+2){<br">    float Add=(float)(1.0/(float)dem);
    Pi=Pi+(float)(mult*Add);
    System.out.println("Dem "+dem+"  "+Add+ "  Pi "+4.0*Pi);
    mult=-1*mult;
}

The problem is getting the answer to 5 digits accuracy. How can we know it's accurate unless we know the value of Pi. My solution (which does have it's problems is to iterate until the calculated value of Pi doesn't change to the accuracy we require. In the examples below I've used string formatting to track the old value and the new value until they are the same. Both versions give the same answer 3.14159

In Java

import java.text.*;
public class Pi {
   public static void main(String[] args) {
      double Pi=(double)1.0;
      //http://www.javaprogrammingforums.com/java-programming-tutorials/297-how-format-double-value-2-decimal-places.html
      DecimalFormat df = new DecimalFormat("#.#####");
      String oldPi;
      String newPi;
      long dem=3;
      oldPi =df.format((double)4.0);
      newPi =df.format(Pi);
      int mult=-1;
      while (oldPi.compareTo(newPi)!=0){
        oldPi=df.format((double)4.0*Pi);
        double Add=(double)(1.0/(double)dem);
        Pi=Pi+(double)(mult*Add);
        newPi=df.format((double)4.0*Pi);
        System.out.println("Dem "+dem+"  "+Add+ "  Pi "+df.format((double)4.0*Pi)+ "  Pi "+4.0*Pi+" : "+oldPi+" : "+newPi);
      mult=-1*mult;
     dem+=2;
     }
 }
}

And in C

#include
#include
main(){
   double Pi=(double)1.0;
   char oldPi[100];
   char newPi[100];
   long dem=3;
   sprintf(oldPi,"%.5f",(double)4.0);
   sprintf(newPi,"%.5f",(double)Pi);
   int mult=-1;
   while (strcmp(oldPi,newPi)!=0){
      sprintf(oldPi,"%.5f",(double)4.0*Pi);
      double Add=(double)(1.0/(double)dem);
      Pi=Pi+(double)(mult*Add);
      sprintf(newPi,"%.5f",(double)4.0*Pi);
      printf(" %ld Pi %.5f %s %s \n",dem,4.0*Pi,oldPi,newPi);
      mult=-1*mult;
      dem+=2;
   }
}

The problem with both these versions is they don't work if you try and increase the accuracy.  If you want 6 decimal places then the value doesn't settle down, in fact it oscillates between two values and never stays the same.

So two questions:

1. What does the code look like in other languages (particularly something like Erlang)
2. How to deal with the oscillation problem ?

Cassandra on a Raspberry Pi, 5 and 6 node insert stress tests:

Here's a quick update for the performance graphs for cassandra on Raspberry pi.  Here's the results for 5 and 6 node inserts on a stress test

I'm getting to the point in this project where I can start to build pseudo data centers and start to test performance there.

Java performance on Raspbian vs Debian

Over the past few weeks I’ve been blogging about my experience of running Apache Cassandra on the Raspberry Pi.  I plan to use the Pi as an educational resource in the University I work in, hopefully giving students the chance to play with large clusters and experiment with configurations, database models and practices in a nosql environment.  Of course performance isn’t great but  for me, it’s a  cheap way of getting lots of nodes and do real network configuration problems. 

A couple of days ago a new Debian based distro for the pi was released called Raspbian “wheezy” was released and is now the official Raspberry Pi Debian distro (I believe).   This is the first OS release for the Pi to take advantage of the Pi’s floating point hardware, which is going to make the OS a lot faster for general use.  I downloaded it for testing in my rig, sadly this is a tale of woe.

Apache Cassandra is a Java application and needs a JRE in order to run.    I’ve always used a  Oracle supplied JVM “Oracle’s java SE for  embedded “

http://www.oracle.com/technetwork/java/embedded/downloads/javase/index.html


Sadly, it seems this can’t be used on Raspbian.  Trying to run it gives :

Java: error while loading shared libraries: libjli.so: cannot open shared object file: No such file or directory

It seems that this version of Java uses the “soft float ABI (armel) which is incompatible with Raspbian”: (thanks to mpthompson on the Raspberry Pi forum for the information) so it’s looking like it can’t run.  Back to openjdk ?

But wait !  Why did I not use openjdk in the first place ?

That’s simple, performance.  In my experience (and perhaps this is a configuration problem I’m not aware of)  Open JDK is a lot slower than the Oracle version.  And  I mean a lot slower!  I set up a single node Cassandra server image, one with the old Debian image and Oracle Java the other with Raspbian and Open JDk.  I then ran stress tests from a Apple Air (something I’ve done many times !) .  Here’s the results.  The second column is interval_op_rate, you want this to be as high as possible, the third column  is  avg_latency, you want this to be as low as possible.

Raspbian and OpenJDK

>Lifeintheairage:bin Administrator$ ./stress -d 192.168.1.12 -o insert -I DeflateCompressor
Unable to create stress keyspace: Keyspace names must be case-insensitively unique ("Keyspace1" conflicts with "Keyspace1")
total,interval_op_rate,interval_key_rate,avg_latency,elapsed_time
485,48,48,0.8705896907216495,10
1042,55,55,0.9123070017953321,20
1436,39,39,1.2947030456852793,30
2010,57,57,0.9009128919860627,40
2510,50,50,0.961294,51
2743,23,23,1.922206008583691,61
3306,56,56,1.0665861456483126,71
3863,55,55,0.9055601436265709,81
4118,25,25,2.0272901960784315,91
4659,54,54,0.9333364140480591,102
5031,37,37,0.916733870967742,112
5498,46,46,1.480710920770878,122

 

Debian Squeze and Java SE for embedded

>lifeintheairage:bin Administrator$ ./stress -d 192.168.1.10 -o insert -I DeflateCompressor
Unable to create stress keyspace: Keyspace names must be case-insensitively unique ("Keyspace1" conflicts with "Keyspace1")
total,interval_op_rate,interval_key_rate,avg_latency,elapsed_time
2565,256,256,0.18891695906432748,10
4604,203,203,0.2503182932810201,20
7093,248,248,0.20536078746484532,30
9289,219,219,0.23249635701275045,40
11516,222,222,0.22830354737314773,51
14107,259,259,0.19691161713624084,61
16297,219,219,0.22646849315068493,71
18092,179,179,0.29083064066852365,81
19756,166,166,0.30374939903846154,91
21689,193,193,0.2648339368856699,102
23404,171,171,0.18766355685131195,112
25395,199,199,0.3459779005524862,122
27646,225,225,0.2330964015992892,132
29684,203,203,0.24136898920510305,142


And a graph of interval_op_rate:

(red is Java SE for embedded, blue is OpenJDK)

Java SE for embedded really is a lot faster for Apache Cassandra (and I wouldn’t be surprised for other java apps such as Arduino IDE).  For now I need to stick with the Debian release, I hope it doesn’t become unsupported.  Hopefully someone can  get in touch with Oracle and encourage them to support a official port of Java SE for embedded onto the raspberry which supports the correct Raspbian libraries.