HA HDFS on Apache Mesos

Starts 1 active NameNode (with JournalNode and ZKFC), 1 standby NN (+JN,ZKFC), 1 JN, and everything else is DataNodes.

Prerequisites

Install tar, unzip, wget in your build host. Set proxy for maven / gradle and wget if needed.
Install curl for all hosts in cluster.
$JAVA_HOME needs to be set on the host running your HDFS scheduler. This can be set through setting the environment variable on the host, export JAVA_HOME=/path/to/jre, or specifying the environment variable in Marathon.

NOTE: The build process current supports maven and gradle. The gradle wrapper meta-data is included in the project and is self boot-straping (meaning it isn't a prerequisite install). Maven as the build system is being deprecated.

Building HDFS-Mesos

Customize configuration in conf/*-site.xml. All configuration files updated here will be used by the scheduler and also bundled with the executors.
./bin/build-hdfs
Run ./bin/build-hdfs nocompile to skip the gradlew clean package step and just re-bundle the binaries.
To remove the project build output and downloaded binaries, run ./bin/build-hdfs clean.

NOTE: The build process builds the artifacts under the $PROJ_DIR/build directory. A number of zip and tar files are cached under the cache directory for faster subsequent builds. The tarball used for installation is hdfs-mesos-x.x.x.tgz which contains the scheduler and the executor to be distributed.

Installing HDFS-Mesos on your Cluster

Upload hdfs-mesos-*.tgz to a node in your Mesos cluster (which is built to $PROJ_DIR/build/hdfs-mesos-x.x.x.tgz).
Extract it with tar zxvf hdfs-mesos-*.tgz.
Optional: Customize any additional configurations that weren't updated at compile time in hdfs-mesos-*/etc/hadoop/*-site.xml Note that if you update hdfs-site.xml, it will be used by the scheduler and bundled with the executors. However, core-site.xml and mesos-site.xml will be used by the scheduler only.
Check that hostname on that node resolves to a non-localhost IP; update /etc/hosts if necessary.

If you have Hadoop pre-installed in your cluster

If you have Hadoop installed across your cluster, you don't need the Mesos scheduler application to distribute the binaries. You can set the mesos.hdfs.native-hadoop-binaries configuration parameter in mesos-site.xml if don't want the binaries distributed.

Mesos-DNS custom configuration

You can see the example configuration in the example-conf/dcos directory. Since Mesos-DNS provides native bindings for master detection, we can simply use those names in our mesos and hdfs configurations. The example configuration assumes your Mesos masters and your zookeeper nodes are colocated. If they aren't you'll need to specify your zookeeper nodes separately. Also, note that you are using the example in example-conf/dcos, the mesos.hdfs.native-hadoop-binaries property needs to be set to false if your HDFS binaries are not predistributed.

Starting HDFS-Mesos

cd hdfs-mesos-*
./bin/hdfs-mesos
Check the Mesos web console to wait until all tasks are RUNNING (monitor status in JN sandboxes)

Using HDFS

See some of the many HDFS tutorials out there for more details and explore the web UI at
http://<ActiveNameNode>:50070.
Note that you can access commands through hdfs://<mesos.hdfs.framework.name>/ (default: hdfs://hdfs/). Also here is a quick sanity check:

hadoop fs -ls hdfs://hdfs/ should show nothing for starters
hadoop fs -put /path/to/src_file hdfs://hdfs/
hadoop fs -ls hdfs://hdfs/ should now list src_file

Resource Reservation Instructions (Optional)

In mesos-site.xml, change mesos.hdfs.role to hdfs.
On master, add the role for HDFS, by running echo hdfs > /etc/mesos-master/role or by setting the —-role=hdfs.
Then restart the master by running sudo service mesos-master restart.
On each slave where you want to reserve resources, add specific resource reservations for the HDFS role. Here is one example:
cpus(*):8;cpus(hdfs):4;mem(*):16384;mem(hdfs):8192 > /etc/mesos-slave/resources
or by setting —-resources=cpus(*):8;cpus(hdfs):4;mem(*):16384;mem(hdfs):8192.
On each slave with the new settings, stop the mesos slave by running
sudo service mesos-slave stop.
On each slave with the new settings, remove the old slave state by running
rm -f /tmp/mesos/meta/slaves/latest.
Note: This will also remove task state, so you will want to manually kill any running tasks as a precaution.
On each slave with the new settings, start the mesos slave by running
sudo service mesos-slave start.

Applying mesos slave constraints (Optional)

In mesos-site.xml, add the configuration mesos.hdfs.constraints
Set the value of configuration as ";" separated set of key:value pairs. Key and value has to be separated by the ":". Key represents the attribute name. Value can be exact match, less than or equal to, subset or value within the range for attribute of type text, scalar, set and range, respectively. For example:

<property>
  <name>mesos.hdfs.constraints</name>
  <value>zone:west,east;cpu:4;quality:optimized-disk;id:4</value>
</property>

"zone" is type of set with members {"west","east"}.
"cpu" is type of scalar. 
"quality" is type of text. 
"id" may be type of range.

Authentication with CRAM-MD5 (Optional)

In mesos-site.xml add the "mesos.hdfs.principal" and "mesos.hdfs.secret" properties. For example:

<property>
  <name>mesos.hdfs.principal</name>
  <value>hdfs</value>
</property>

<property>
  <name>mesos.hdfs.secret</name>
  <value>%ComplexPassword%123</value>
</property>

Ensure that the Mesos master has access to the same credentials. See the Mesos configuration documentation, in particular the --credentials flag. Authentication defaults to CRAM-MD5 so setting the --authenticators flag is not necessary.

Shutdown Instructions (Optional)

In Marathon (or your other long-running process monitor) stop the hdfs scheduler application
Shutdown the hdfs framework in Mesos: curl -d "frameworkId=YOUR_FRAMEWORK_ID" -X POST http://YOUR_MESOS_URL:5050/master/shutdown
Access your zookeeper instance: /PATH/TO/zookeeper/bin/zkCli.sh
Remove hdfs-mesos framework state from zookeeper: rmr /hdfs-mesos
(Optional) Clear your data directories as specified in your mesos-site.xml. This is necessary to relaunch HDFS in the same directory.

Developer Notes

The project uses guice which is a light weight dependency injection framework. In this project it is used during application startup initialization. This is accomplished by using the @Inject annotation. Guice is aware of all concrete classes which are annotated with @Singleton, however when it comes to interfaces, guice needs to be "bound" to an implementation. This is accomplished with the HdfsSchedulerModule guice module class and is initialized in the main class with:

  // this is initializes guice with all the singletons + the passed in module
  Injector injector = Guice.createInjector(new HdfsSchedulerModule());
  
  // if this returns successfully, then the object was "wired" correctly.
  injector.getInstance(ConfigServer.class);

If you have a singleton, mark it as such. If you have an interface + implemention class then bind it in the HdfsSchedulerModule such as:

  // bind(<interface>.class).to(<impl>.class);
  bind(IPersistentStateStore.class).to(PersistentStateStore.class);

In this case, when an @Inject is encountered during the initialization of a guice initialized class, parameters of type <interface> will have an instance of the <impl> class passed.

The advantage of this technique is that the interface can easily have a mock class provided for testing. For more motivation read guice's motivation page

jan-zajic / mesos-hdfs