hadoop spark #21

Overview

The Apache Hadoop software library is a framework that allows for the
distributed processing of large data sets across clusters of computers
using a simple programming model.

Hadoop is designed to scale from a few servers to thousands of machines,
each offering local computation and storage. Rather than rely on hardware
to deliver high-availability, Hadoop can detect and handle failures at the
application layer. This provides a highly-available service on top of a cluster
of machines, each of which may be prone to failure.

Spark is a fast and general engine for large-scale data processing.

This bundle provides a complete deployment of Hadoop and Spark components from
Apache Bigtop that performs distributed data processing at scale. Ganglia
and rsyslog applications are also provided to monitor cluster health and syslog
activity.

Bundle Composition

The applications that comprise this bundle are spread across 9 units as
follows:

  • NameNode (HDFS)
  • ResourceManager (YARN)
    • Colocated on the NameNode unit
  • Slave (DataNode and NodeManager)
    • 3 separate units
  • Spark (Master in yarn-client mode)
  • Zookeeper
    • 3 separate units
  • Client (Hadoop endpoint)
  • Plugin (Facilitates communication with the Hadoop cluster)
    • Colocated on the Spark and Client units
  • Ganglia (Web interface for monitoring cluster metrics)
    • Colocated on the Client unit
  • Rsyslog (Aggregate cluster syslog events in a single location)
    • Colocated on the Client unit

Deploying this bundle results in a fully configured Apache Bigtop
cluster on any supported cloud, which can be scaled to meet workload
demands.

Deploying

A working Juju installation is assumed to be present. If Juju is not yet set
up, please follow the getting-started instructions prior to deploying this
bundle.

Note: This bundle requires hardware resources that may exceed limits
of Free-tier or Trial accounts on some clouds. To deploy to these
environments, modify a local copy of bundle.yaml to set
services: 'X': num_units: 1 and machines: 'X': constraints: mem=3G as
needed to satisfy account limits.

Deploy this bundle from the Juju charm store with the juju deploy command:

juju deploy hadoop-spark

Note: The above assumes Juju 2.0 or greater. If using an earlier version
of Juju, use juju-quickstart with the following syntax: juju quickstart hadoop-spark.

Alternatively, deploy a locally modified bundle.yaml with:

juju deploy /path/to/bundle.yaml

Note: The above assumes Juju 2.0 or greater. If using an earlier version
of Juju, use juju-quickstart with the following syntax: juju quickstart /path/to/bundle.yaml.

The charms in this bundle can also be built from their source layers in the
Bigtop charm repository. See the Bigtop charm README for instructions
on building and deploying these charms locally.

Network-Restricted Environments

Charms can be deployed in environments with limited network access. To deploy
in this environment, configure a Juju model with appropriate proxy and/or
mirror options. See Configuring Models for more information.

Verifying

Status

The applications that make up this bundle provide status messages to indicate
when they are ready:

juju status

This is particularly useful when combined with watch to track the on-going
progress of the deployment:

watch -n 2 juju status

The message for each unit will provide information about that unit's state.
Once they all indicate that they are ready, perform application smoke tests
to verify that the bundle is working as expected.

Smoke Test

The charms for each core component (namenode, resourcemanager, slave, spark,
and zookeeper) provide a smoke-test action that can be used to verify the
application is functioning as expected. Note that the 'slave' component runs
extensive tests provided by Apache Bigtop and may take up to 30 minutes to
complete. Run the smoke-test actions as follows:

juju run-action namenode/0 smoke-test
juju run-action resourcemanager/0 smoke-test
juju run-action slave/0 smoke-test
juju run-action spark/0 smoke-test
juju run-action zookeeper/0 smoke-test

Note: The above assumes Juju 2.0 or greater. If using an earlier version
of Juju, the syntax is juju action do <application>/0 smoke-test.

Watch the progress of the smoke test actions with:

watch -n 2 juju show-action-status

Note: The above assumes Juju 2.0 or greater. If using an earlier version
of Juju, the syntax is juju action status.

Eventually, all of the actions should settle to status: completed. If
any report status: failed, that application is not working as expected. Get
more information about a specific smoke test with:

juju show-action-output <action-id>

Note: The above assumes Juju 2.0 or greater. If using an earlier version
of Juju, the syntax is juju action fetch <action-id>.

Utilities

Applications in this bundle include command line and web utilities that
can be used to verify information about the cluster.

From the command line, show the HDFS dfsadmin report and view the current list
of YARN NodeManager units with the following:

juju run --application namenode "su hdfs -c 'hdfs dfsadmin -report'"
juju run --application resourcemanager "su yarn -c 'yarn node -list'"

Show the list of Zookeeper nodes with the following:

juju run --unit zookeeper/0 'echo "ls /" | /usr/lib/zookeeper/bin/zkCli.sh'

To access the HDFS web console, find the PUBLIC-ADDRESS of the namenode
application and expose it:

juju status namenode
juju expose namenode

The web interface will be available at the following URL:

http://NAMENODE_PUBLIC_IP:50070

Similarly, to access the Resource Manager web consoles, find the
PUBLIC-ADDRESS of the resourcemanager application and expose it:

juju status resourcemanager
juju expose resourcemanager

The YARN and Job History web interfaces will be available at the following URLs:

http://RESOURCEMANAGER_PUBLIC_IP:8088
http://RESOURCEMANAGER_PUBLIC_IP:19888

Finally, to access the Spark web console, find the PUBLIC-ADDRESS of the
spark application and expose it:

juju status spark
juju expose spark

The web interface will be available at the following URL:

http://SPARK_PUBLIC_IP:8080

Monitoring

This bundle includes Ganglia for system-level monitoring of the namenode,
resourcemanager, slave, spark, and zookeeper units. Metrics are sent to a
centralized ganglia unit for easy viewing in a browser. To view the ganglia web
interface, find the PUBLIC-ADDRESS of the Ganglia application and expose it:

juju status ganglia
juju expose ganglia

The web interface will be available at:

http://GANGLIA_PUBLIC_IP/ganglia

Logging

This bundle includes rsyslog to collect syslog data from the namenode,
resourcemanager, slave, spark, and zookeeper units. These logs are sent to a
centralized rsyslog unit for easy syslog analysis. One method of viewing this
log data is to simply cat syslog from the rsyslog unit:

juju run --unit rsyslog/0 'sudo cat /var/log/syslog'

Logs may also be forwarded to an external rsyslog processing service. See
the Forwarding logs to a system outside of the Juju environment section of
the rsyslog README for more information.

Benchmarking

The resourcemanager charm in this bundle provide several benchmarks to gauge
the performance of the Hadoop cluster. Each benchmark is an action that can be
run with juju run-action:

$ juju actions resourcemanager
ACTION      DESCRIPTION
mrbench     Mapreduce benchmark for small jobs
nnbench     Load test the NameNode hardware and configuration
smoke-test  Run an Apache Bigtop smoke test.
teragen     Generate data with teragen
terasort    Runs teragen to generate sample data, and then runs terasort to sort that data
testdfsio   DFS IO Testing

$ juju run-action resourcemanager/0 nnbench
Action queued with id: 55887b40-116c-4020-8b35-1e28a54cc622

$ juju show-action-output 55887b40-116c-4020-8b35-1e28a54cc622
results:
  meta:
    composite:
      direction: asc
      units: secs
      value: "128"
    start: 2016-02-04T14:55:39Z
    stop: 2016-02-04T14:57:47Z
  results:
    raw: '{"BAD_ID": "0", "FILE: Number of read operations": "0", "Reduce input groups":
      "8", "Reduce input records": "95", "Map output bytes": "1823", "Map input records":
      "12", "Combine input records": "0", "HDFS: Number of bytes read": "18635", "FILE:
      Number of bytes written": "32999982", "HDFS: Number of write operations": "330",
      "Combine output records": "0", "Total committed heap usage (bytes)": "3144749056",
      "Bytes Written": "164", "WRONG_LENGTH": "0", "Failed Shuffles": "0", "FILE:
      Number of bytes read": "27879457", "WRONG_MAP": "0", "Spilled Records": "190",
      "Merged Map outputs": "72", "HDFS: Number of large read operations": "0", "Reduce
      shuffle bytes": "2445", "FILE: Number of large read operations": "0", "Map output
      materialized bytes": "2445", "IO_ERROR": "0", "CONNECTION": "0", "HDFS: Number
      of read operations": "567", "Map output records": "95", "Reduce output records":
      "8", "WRONG_REDUCE": "0", "HDFS: Number of bytes written": "27412", "GC time
      elapsed (ms)": "603", "Input split bytes": "1610", "Shuffled Maps ": "72", "FILE:
      Number of write operations": "0", "Bytes Read": "1490"}'
status: completed
timing:
  completed: 2016-02-04 14:57:48 +0000 UTC
  enqueued: 2016-02-04 14:55:14 +0000 UTC
  started: 2016-02-04 14:55:27 +0000 UTC

The spark charm in this bundle also provides several benchmarks to gauge
the performance of the Spark cluster. Each benchmark is an action that can be
run with juju run-action:

$ juju actions spark | grep Bench
connectedcomponent                Run the Spark Bench ConnectedComponent benchmark.
decisiontree                      Run the Spark Bench DecisionTree benchmark.
kmeans                            Run the Spark Bench KMeans benchmark.
linearregression                  Run the Spark Bench LinearRegression benchmark.
logisticregression                Run the Spark Bench LogisticRegression benchmark.
matrixfactorization               Run the Spark Bench MatrixFactorization benchmark.
pagerank                          Run the Spark Bench PageRank benchmark.
pca                               Run the Spark Bench PCA benchmark.
pregeloperation                   Run the Spark Bench PregelOperation benchmark.
shortestpaths                     Run the Spark Bench ShortestPaths benchmark.
sql                               Run the Spark Bench SQL benchmark.
stronglyconnectedcomponent        Run the Spark Bench StronglyConnectedComponent benchmark.
svdplusplus                       Run the Spark Bench SVDPlusPlus benchmark.
svm                               Run the Spark Bench SVM benchmark.

$ juju run-action spark/0 svdplusplus
Action queued with id: 339cec1f-e903-4ee7-85ca-876fb0c3d28e

$ juju show-action-output 339cec1f-e903-4ee7-85ca-876fb0c3d28e
results:
  meta:
    composite:
      direction: asc
      units: secs
      value: "200.754000"
    raw: |
      SVDPlusPlus,2016-11-02-03:08:26,200.754000,85.974071,.428255,0,SVDPlusPlus-MLlibConfig,,,,,10,,,50000,4.0,1.3,
    start: 2016-11-02T03:08:26Z
    stop: 2016-11-02T03:11:47Z
  results:
    duration:
      direction: asc
      units: secs
      value: "200.754000"
    throughput:
      direction: desc
      units: MB/sec
      value: ".428255"
status: completed
timing:
  completed: 2016-11-02 03:11:48 +0000 UTC
  enqueued: 2016-11-02 03:08:21 +0000 UTC
  started: 2016-11-02 03:08:26 +0000 UTC

Scaling

By default, three Hadoop slave and three zookeeper units are deployed. Scaling
these applications is as simple as adding more units. To add one unit:

juju add-unit slave
juju add-unit zookeeper

Multiple units may be added at once. For example, add four more slave units:

juju add-unit -n4 slave

Contact Information

Resources