=================================
=				=
= GORACI README			=
= @author Keith Turner		=
=================================

GoraCI has now been integrated into [Apache Gora][A] master branch where
current development is ongoing. If you would like to learn more about GoraCI
please see the [Gora documentation][B].

[A] http://gora.apache.org
[B] http://gora.apache.org/current/index.html#goraci-integration-testsing-suite

BACKGROUND
------------

Apache Accumulo [0] has a simple test suite that verifies that data is not lost
at scale.  This test suite is called continuous ingest [5].  This test runs
many ingest clients that continually create linked lists containing 25 million
nodes. At some point the clients are stopped and a map reduce job is run to
ensure no linked list has a hole. A hole indicates data was lost.    

The nodes in the linked list are random.  This causes each linked list to
spread across the table.  Therefore if one part of a table loses data, then it
will be detected by references in another part of the table.

This project is a version of the test suite written using Apache Gora [1].
Goraci has been tested against Accumulo and HBase.  

THE ANATOMY OF GORACI TESTS
----------------------------

Below is rough sketch of how data is written.  For specific details look at the
Generator code (src/main/java/goraci/Generator.java)

  1 Write out 1 million nodes 
  2 Flush the client 
  3 Write out 1 million that reference previous million 
  4 If this is the 25th set of 1 million nodes, then update 1st set of million
    to point to last 
  5 goto 1

The key is that nodes only reference flushed nodes.  Therefore a node should
never reference a missing node, even if the ingest client is killed at any
point in time.

When running this test suite w/ Accumulo there is a script running in parallel
called the Aggitator that randomly and continuously kills server processes.  
The outcome was that many data loss bugs were found in Accumulo by doing this. 
This test suite can also help find bugs that impact uptime and stability when 
run for days or weeks.  

This test suite consists the following 
- a few Java programs 
- a little helper script to run the java programs
- a maven script to build it.  

When generating data, its best to have each map task generate a multiple of 25
million.  The reason for this is that circular linked list are generated every
25M.  Not generating a multiple in 25M will result in some nodes in the linked
list not having references.  The loss of an unreferenced node can not be
detected.

BUILDING GORACI
---------------

This code currently depends on an unreleased version of Gora.  To build Gora
0.2 run the following commands.

  svn export http://svn.apache.org/repos/asf/gora/trunk gora
  cd gora
  mvn install -DskipTests

After this you can build goraci.

  git clone git://github.com/keith-turner/goraci.git
  cd goraci
  mvn compile

The maven pom file has some profiles that attempt to make it easier to run
goraci against different gora backends by copying the jars you need into lib.
Before packaging its important to edit gora.properties and set it correctly
for your datastore.  To run against accumulo do the following.

  vim src/main/resources/gora.properties (set Accumulo properties)
  mvn package -Paccumulo-1.4

To run against hbase, do the following.

  vim src/main/resources/gora.properties (set HBase properties)
  mvn package -Phbase-0.92

To run against cassandra, do the following.

  vim src/main/resources/gora.properties (set Cassandra properties)
  mvn package -Pcassandra-1.1.2

For other datastores mentioned in gora.properties, you will need to copy the
appropriate deps into lib.  Feel free to update the pom with other profiles and
send me pull request.

JAVA CLASS DESCRIPTION
-----------------

Below is a description of the Java programs

  * goraci.Generator - A map only job that generates data.  As stated previously, 
                       its best to generate data in multiples of 25M.
  * goraci.Verify    - A map reduce job that looks for holes.  Look at the
                       counts after running.  REFERENCED and UNREFERENCED are 
                       ok, any UNDEFINED counts are bad. Do not run at the 
                       same time as the Generator.
  * goraci.Walker    - A standalong program that start following a linked list 
                       and emits timing info.  
  * goraci.Print     - A standalone program that prints nodes in the linked list
  * goraci.Delete    - A standalone program that deletes a single node
  * goraci.Loop      - Runs generation and verify in a loop

goraci.sh is a helper script that you can use to run the above programs.  It
assumes all needed jars are in the lib dir.  It does not need the package name.
You can just run "./goraci.sh Generator", below is an example.

  $ ./goraci.sh Generator
  Usage : Generator <num mappers> <num nodes>

For Gora to work, it needs a gora.properties file on the classpath and a
mapping file on the classpath, the contents of both are datastore specific,
more details can be found here [2]. You can edit the ones in src/main/resources
and build the goraci-${version}-SNAPSHOT.jar with those. Alternatively remove
those and put them on the classpath through some other means.

GORA AND HADOOP
-----------------

Gora uses Avro which uses a Json library that Hadoop has an old version of.
The two libraries  jackson-core and jackson-mapper need to be updated in
<HADOOP_HOME>/lib and <HADOOP_HOME>/share/hadoop/lib/.  I updated these to
jackson-core-asl-1.4.2.jar and jackson-mapper-asl-1.4.2.jar.  For details see
HADOOP-6945 [3]. 

GORACI AND HBASE
-----------------

To improve performance running read jobs such as the Verify step, enable
scanner caching on the command line.  For example:

    $ ./gorachi.sh Verify -Dhbase.client.scanner.caching=1000 \
         -Dmapred.map.tasks.speculative.execution=false verify_dir 1000

Dependent on how you have your hadoop and hbase deployed, you may need to
change the gorachi.sh script around some.  Here is one suggestion that may help
in the case where your hadoop and hbase configuration are other than under the
hadoop and hbase home directories.

  diff --git a/goraci.sh b/goraci.sh
  index db1562a..31c3c94 100755
  --- a/goraci.sh
  +++ b/goraci.sh
  @@ -95,6 +95,4 @@ done
   #run it
   export HADOOP_CLASSPATH="$CLASSPATH"
   LIBJARS=`echo $HADOOP_CLASSPATH | tr : ,`
  -hadoop jar "$GORACI_HOME/lib/goraci-0.0.1-SNAPSHOT.jar" $CLASS -libjars "$LIBJARS" "$@"
  -
  -
  +CLASSPATH="${HBASE_CONF_DIR}" hadoop --config "${HADOOP_CONF_DIR} jar "$GORACI_HOME/lib/goraci-0.0.1-SNAPSHOT.jar" $CLASS -files "${HBASE_CONF_DIR}/hbase-site.xml" -libjars "$LIBJARS" "$@"

You will need to define HBASE_CONF_DIR and HADOOP_CONF_DIR before you run your
goraci jobs.  For example:

  $ export HADOOP_CONF_DIR=/home/you/hadoop-conf
  $ export HBASE_CONF_DIR=/home/you/hbase-conf
  $ PATH=/home/you/hadoop-1.0.2/bin:$PATH ./goraci.sh Generator 1000 1000000

CONCURRENCY
------------

Its possible to run verification at the same time as generation.  To do this
supply the -c option to Generator and Verify.  This will cause Genertor to
create a secondary table which holds information about what verification can
safely verify.  Running Verify with the -c option will make it run slower
because more information must be brought back to the client side for filtering
purposes.  The Loop program also supports the -c option, which will cause it to
run verification concurrently with generation.

If verification is run at the same time as generation without the -c option,
then it will inevitably fail.  This is because verification mappers read
different parts of the table at different times and giving an inconsistent view
of the table.  So one mapper may read a part of a table before a node is
written, when the node is later referenced it will appear to be missing.  The
-c option basically filters out newer information using data written to the
secondary table.

CONCLUSIONS
------------

This test suite does not do everything that the Accumulo test suite does,
mainly it does not collect statistics and generate reports.  The reports
are useful for assesing performance.

Below shows running a test of the test.  Ingest one linked list, deleted a node
in it, ensure the verifaction map reduce job notices that the node is missing.
Not all output is shown, just the important parts.

  $ ./goraci.sh Generator  1 25000000
  $ ./goraci.sh Print -s 2000000000000000 -l 1
  2000001f65dbd238:30350f9ae6f6e8f7:000004265852:ef09f9dd-75b1-4c16-9f14-0fa84f3029b6
  $ ./goraci.sh Print -s 30350f9ae6f6e8f7 -l 1
  30350f9ae6f6e8f7:4867fe03de6ea6c8:000003265852:ef09f9dd-75b1-4c16-9f14-0fa84f3029b6
  $ ./goraci.sh Delete 30350f9ae6f6e8f7
  Delete returned true
  $ ./goraci.sh Verify gci_verify_1 2 
  11/12/20 17:12:31 INFO mapred.JobClient:   goraci.Verify$Counts
  11/12/20 17:12:31 INFO mapred.JobClient:     UNDEFINED=1
  11/12/20 17:12:31 INFO mapred.JobClient:     REFERENCED=24999998
  11/12/20 17:12:31 INFO mapred.JobClient:     UNREFERENCED=1
  $ hadoop fs -cat gci_verify_1/part\*
  30350f9ae6f6e8f7	2000001f65dbd238

The map reduce job found the one undefined node and gave the node that
referenced it.

Below are some timing statistics for running goraci on a 10 node cluster. 

  Store           | Task                   | Time    | Undef  | Unref | Ref        
  ----------------+------------------------+---------+--------+-------+------------
  accumulo-1.4.0  | Generator 10 100000000 | 40m 16s |    N/A |   N/A |        N/A     
  accumulo-1.4.0  | Verify /tmp/goraci1 40 |  6m  7s |      0 |     0 | 1000000000  
  hbase-0.92.1    | Generator 10 100000000 |  2h 44m |    N/A |   N/A |        N/A     
  hbase-0.92.1    | Verify /tmp/goraci2 40 |  6m 34s |      0 |     0 | 1000000000

Hbase and Accumulo are configured differently out-of-the-box.  We used the Accumulo 
3G, native configuration examples in the conf/examples directory.

To provide a comparable memory footprint, we increased the HBase jvm to "-Xmx4000m", 
and turned on compression for the ci table:

create 'ci', {NAME=>'meta', COMPRESSION=>'GZ'}

We also turned down the replication of write-ahead logs to be comparable to Accumulo:

  <property>
    <name>hbase.regionserver.hlog.replication</name>
    <value>2</value>
  </property>

For the accumulo run, we set the split threshold to 512M:

 shell> config -t ci -s table.split.threshold=512M

This was done so that Accumulo would end up with 64 tablets, which is the
number of regions hbase had.   The number of tablets/regions determines how
much parallelism there is in the map phase of the verify step.

Sometimes when this test suite is run against HBase data is lost.  This issue
is being tracked under HBASE-5754 [4].

[0] http://accumulo.apache.org
[1] http://gora.apache.org
[2] http://gora.apache.org/docs/current/gora-conf.html
[3] https://issues.apache.org/jira/browse/HADOOP-6945
[4] https://issues.apache.org/jira/browse/HBASE-5754
[5] http://svn.apache.org/viewvc/accumulo/tags/1.4.0/test/system/continuous/ScaleTest.odp?view=co