Ubuntu Server - Hadoop

User Stories
------------------

Marcos has been tasked with setting up a new Hadoop compute cluster to support log analysis of this companies website activity; He is able to quickly and easily deploy Apache Hadoop on his companies internal Cloud using Juju with packages that come directly from the Canonical partner archive (as mandated by his companies IT policy).

- Implemented - charms for hadoop, zookeeper, hbase and hive avaliable in the Juju charm store albiet packages come from a PPA rather than partner for this release.

Natalie wants to setup a Hadoop cluster; she's not able to use Juju and a public cloud due to the existing infrastructure policies and design in place at her organisation but she is able to deploy the required packages directly from the Canonical partner archive.

- Implemented - packages work outside of charms albiet packages come from a PPA rather than partner for this release.

------

Status 20120411

Hadoop 1.0.2 uploaded to dev archive.

Charms for hbase, hadoop and hive written for Ubuntu precise.

Status 20120603

Testing packages avaliable for hadoop, hbase, pig and hadoop-zookeeper in http://launchpad.net/~hadoop-ubuntu/+archive/testing

sudo add-apt-repository ppa:hadoop-ubuntu/testing

Juju charms avaliable for hadoop, hbase and zookeeper:

lp:~charmers/charms/precise/zookeeper/trunk
lp:~charmers/charms/precise/hadoop/trunk
lp:~charmers/charms/precise/hbase/trunk

Note that by default the zookeeper charm will use the zookeeper packages from the main archive not the PPA - see the charm config.yaml file for details on how to use the
PPA's with these charms.

Dev PPA also now hosting backports of all packages for Ubuntu 11.10.

Status 20122501

Dev packages avaliable for hadoop, hive, pig, hbase and hadoop-zookeeper in http://launchpad.net/~hadoop-ubuntu/+archive/dev

sudo add-apt-repository ppa:hadoop-ubuntu/dev

============================

Target versions of Hadoop components:

hadoop-1.0.2
hcatalog-0.2.0
pig-0.9.1
hbase-0.92.0
hive-0.8.1
zookeeper-3.4.3

Package naming:

hadoop
hcatalog
pig
hbase
hive

Packages that target partner CANNOT have the same name as packages in the main archive so zookeeper will be:

hadoop-zookeeper

This will prevent and conflicts with the zookeeper package in the main archive.

Assumptions:

Packaging will use bigtop as a base.
 - reuse of control file structure will help support bigtop packages in juju charms as and when they support precise.
Java will NOT be rebuilt during the package build process.
Native libraries will be rebuilt during the package build process.
Q: What does this mean for the hive embedded in hcatalog (via hcatalog-0.2.0/hive/external)? I'm guessing we just use the hdp-hive package we create.
Patches may be required for native build components.
Debconf configuration in packages is useful for Juju charms and should be applied where appropriate.
--- Debconf currently in use: namenode, jobtracker and hdfs_dir
Binary distributions will be used from upstream (no source build by default).
Q: Exactly what does this mean in this context? Binary distributions of ... Are these the jar files inside the hortonworks/apache source (e.g. ./hive-0.7.1+/lib/javaewah-0.3.jar?
A: So most of the apache projects ship a binary distribution tarball which includes all of the compile Java etc so we should not build from source - just re-use these jars etc...

Challenges:

hcatalog not packaged in bigtop trunk (need to check branches)
JAVA_HOME and use of openjdk-6/7? JAVA_HOME detection - bigtop use a specific package for this (bigtop-utils) which tries to guess - might be worth re-using.
java-package as recommended approach to re-packaging Oracle java for Ubuntu (generates Debian packaging from upstream binary distro).
Kerberos security - probably actually needs to be a charm configuration option to secure a cluster.
Q: For Kerberos, do we have the necessary dependencies to get this done? If so, I can do the charm work.j
Q: Re: Kerberos. We may need to really look into this as this will end up affecting not only hadoop but, all of the other parts ( hive, pig, etc. ) as well. This can really complicate the packaging/charming as well so, we may not have time to get it all done in this cycle. Just my thoughts.
MultiArch native libraries in hdp-hadoop package - should not must.
Hive contains patches for hcatalog support - we need to work these in somehow (maybe a overlay to the binary re-distribution).

hcatalog:
External dependency on mysql-connector? and maybe mysql?
No binary distribution - only source so will need to be built locally to produce release tarbal.

Work items:
Resolve conflict between archive zookeeper and partner zookeeper (1): DONE
[james-page] Package zookeeper for partner (2): DONE
[james-page] Package hadoop for partner (2): DONE
[james-page] Rebuild native component during package rebuild for hadoop (3): DONE
Package hcatalog for partner (3): POSTPONED
[negronjl] Package pig for partner (2): DONE
[james-page] Package hbase for partner (2): DONE
[negronjl] Package hive for partner (2): DONE
[james-page] Work out solution for JAVA_HOME detection/override (1): DONE
Partner Archive upload and review: POSTPONED
[james-page] Juju charm for zookeeper (will depend on the zookeeper package): DONE
Juju charm for hcatalog (will depend on the hcatalog package): POSTPONED
Support iterating packaging as charms are developed (5): DONE
Test deployment and packaging: DONE