Building Hadoop Notes

This documents some of the tricks needed to make building/running the current set of Hadoop projects easier

Set Ant up

# In ANT_OPTS, set up any proxy you need. For example

export ANT_OPTS=-Dhttp.proxyHost=web-proxy -Dhttp.proxyPort=8088 -Dhttps.proxyHost=web-proxy -Dhttps.proxyPort=8088

# set ANT_ARGS to -logger org.apache.tools.ant.listener.BigProjectLogger if you want the "big project logger", which is handy when you build across different modules.

Check out the Hadoop modules

Create a root directory, ~/Hadoop

Check out from SVN Avro (just for the source when debugging), and the Hadoop modules

svn co https://svn.apache.org/repos/asf/hadoop/avro/trunk avro
svn co https://svn.apache.org/repos/asf/hadoop/common/trunk hadoop-common
svn co https://svn.apache.org/repos/asf/hadoop/hdfs/trunk hadoop-hdfs
svn co https://svn.apache.org/repos/asf/hadoop/mapreduce/trunk hadoop-mapreduce

patch the build.xml files of hadoop* to make the project names all lower case otherwise they get published with the wrong name in Ivy

sharing build.properties

The build files all load in ${user.home}/build.properties , so you can set settings there for every project.

They also load in ${basedir}/build.properties

You can share a build.properties file

Make the file: touch ~/Hadoop/build.properties
Symlink it

   ln -s  ~/Hadoop/build.properties  ~/Hadoop/hadoop-common
   ln -s  ~/Hadoop/build.properties  ~/Hadoop/hadoop-hdfs
   ln -s  ~/Hadoop/build.properties  ~/Hadoop/hadoop-mapreduce

Now changes propagate around automatically.

Create a new Hadoop version

In build.properties:

version=0.21.0-alpha-15
hadoop.version=${version}
hadoop-core.version=${version}
hadoop-mr.version=0.21.0-dev

Why the different versions? I check versions in to SVN for hudson et al, bumping up my version count every release. For stuff I build locally, we use a version counter one notch higher. This ensures that local builds get the latest versions, while the other ones get releases.

Now, we are going to have fun by symlinking from to the specific artifacts we create. This stops us having to care about publishing things. Any local build is automatically picked up -until that version number is changed.

Create the core JARs

Change to hadoop-common : cd ~/Hadoop/hadoop-common
ant clean jar jar-test ivy-publish-local
If you have made any changes to the source, run ant test until the tests pass.

run ls -l build/*.jar -expect to see two Jars

 -rw-r--r-- 1 hadoop users 1059526 2009-08-18 16:23 build/hadoop-common-0.21.0-alpha-15.jar
 -rw-r--r-- 1 hadoop users  408716 2009-08-18 16:23 build/hadoop-common-test-0.21.0-alpha-15.jar

link everything up:

 ln -s ~/Hadoop/hadoop-common/build/hadoop-common-0.21.0-alpha-15.jar ../hadoop-hdfs/lib
 ln -s ~/Hadoop/hadoop-common/build/hadoop-common-test-0.21.0-alpha-15.jar ../hadoop-hdfs/lib
 ln -s ~/Hadoop/hadoop-common/build/hadoop-common-0.21.0-alpha-15.jar ../hadoop-mapreduce/lib
 ln -s ~/Hadoop/hadoop-common/build/hadoop-common-test-0.21.0-alpha-15.jar ../hadoop-mapreduce/lib

This has now propagated the JAR to the different directories, and put it in the Ivy repository for other programs to pick up.

The JAR is not copied, just linked to the hadoop-common/build/ version. This means

Whenever a change is made to the common project, and it is rebuilt, the other projects get the change immediately.
If you change the version number in common, all the links break
If you do an ant clean in common/ the other projects will not build until you have gone ant jar jar-test again.

You could change the links to point to the ivy artifacts under ~/.ivy2/local/org.apache.hadoop/. This would give access to the cached versions, and is a half-way house between full Ivy integration for Hadoop JAR retrieval, and integration of local builds via symbolic links. But it does require to work everywhere.

Build the HDFS JARS

Change to hadoop-hdfs : cd ~/Hadoop/hadoop-hdfs
ant clean jar jar-test ivy-publish-local
If you have made any changes to the source, run ant test until the tests pass. It's OK at this point for the run-test-hdfs-with-mr tests to fail.
do an ls -l build to check the JARs are there and versioned right

Symlink the hdfs JARS into mapreduce

  ln -s ~/Hadoop/hadoop-hdfs/build/hadoop-hdfs-0.21.0-alpha-15.jar ../hadoop-mapreduce/lib
  ln -s ~/Hadoop/hadoop-hdfs/build/hadoop-hdfs-test-0.21.0-alpha-15.jar ../hadoop-mapreduce/lib

Build the MapReduce JARS

Change to hadoop-mapreduce : cd ~/Hadoop/hadoop-mapreduce
ant clean jar jar-test

Symlink these jars back into hadoop-hdfs (ooh, a loop!)

  ln -s ~/Hadoop/hadoop-mapreduce/build/hadoop-mapred-0.21.0-alpha-15.jar ../hadoop-hdfs/lib
  ln -s ~/Hadoop/hadoop-mapreduce/build/hadoop-mapred-test-0.21.0-alpha-15.jar ../hadoop-hdfs/lib

Patch build.properties to pick up the new version
```
hadoop-mr.version=${version}
```
Run the mapreduce tests
Go back to hadoop-hdfs and run the tests and make sure that the tests under run-test-hdfs-with-mr pass

BuildingHadoopFromSVN