We recently got a big new server at work to run Hadoop and Spark (H/S) on for a proof-of-concept test of some software we're writing for the biopharmaceutical industry and I hit a few snags while trying to get H/S up and running on Windows Server 2016. I've documented here, step-by-step, how I managed to install and run this pair of Apache products directly in the Windows cmd prompt, without any need for Linux emulation.

Get the Software

The first step is to download Java, Hadoop, and Spark. Spark seems to have trouble working with newer versions of Java, so I'm sticking with Java 8 for now:

For Java, I download the "Windows x64" version of the JDK (jdk-8u191-windows-x64.exe); for Hadoop, the binary of v3.1.1 (hadoop-3.1.1.tar.gz); for Spark, v2.3.2 "Pre-built for Apache Hadoop 2.7 and later (spark-2.3.2-bin-hadoop2.7.tgz).

Move all three of these files to C:\. Download 7-Zip to extract the *gz archives. Once they're extracted (Hadoop takes a while), you should have two directories and the JDK application:

Run the Java installer but change the destination folder from the default C:\Program Files\Java\jdk.1.8.0_191 to just C:\Java. (H/S can have trouble with directories with spaces in their names.)

Another box will pop up asking for the "Destination Folder" again. This time, use C:\Java\jre1.8.0_191. Make two more directories in C:\ called C:\Hadoop and C:\Spark and copy the hadoop-3.1.1 and spark-2.3.2-bin-hadoop2.7 directories into those directories, respectively:

If you get "name too long"-type warnings, skip those files. These are only *.html files and aren't critical to running H/S.

Set Up Your Environment Variables

Next, we need to set some environment variables. Go to Control Panel > System and Security > System > Advanced System Settings > Environment Variables...:

Add new System variables (bottom box) called:

JAVA_HOME --> C:\Java
HADOOP_HOME --> C:\Hadoop\hadoop-3.1.1
SPARK_HOME --> C:\Spark\spark-2.3.2-bin-hadoop2.7

(Adjust according to the versions of Hadoop and Spark that you've downloaded.)

Then, edit the Path and add those variables with \bin appended (also \sbin for Hadoop):

If you echo %PATH% in cmd you should now see these three directories somewhere in the middle of the path, because the User Path is appended to the System Path for the %PATH variable. You should check now that java -version, hdfs -version, and spark-shell --version return version numbers, as shown below. This means that they were correctly installed and added to your %PATH%:

Configure Hadoop

Next, go to C:\Hadoop\hadoop-3.1.1\etc\hadoop and edit (or create) the file core-site.xml so it looks like the following:

core-site.xml

<?xml version="1.0" encoding="UTF-8"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>

<configuration>
  <property>
    <name>fs.defaultFS</name>
    <value>hdfs://localhost:9000</value>
  </property>
</configuration>

In the same directory, edit (or create) mapred-site.xml with the following contents:

mapred-site.xml

<?xml version="1.0" encoding="UTF-8"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>

<configuration>
  <property>
    <name>mapreduce.framework.name</name>
    <value>yarn</value>
  </property>
</configuration>

Next, edit (or create) hdfs-site.xml:

hdfs-site.xml

<?xml version="1.0" encoding="UTF-8"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>

<configuration>
  <property>
    <name>dfs.replication</name>
    <value>1</value>
  </property>
  <property>
    <name>dfs.namenode.name.dir</name>
    <value>file:///C:/Hadoop/hadoop-3.1.1/namenode</value>
  </property>
  <property>
    <name>dfs.datanode.data.dir</name>
    <value>file:///C:/Hadoop/hadoop-3.1.1/datanode</value>
  </property>
</configuration>

...yes, they should be forward slashes, even though Windows uses backslashes. This is due to the way that Hadoop interprets these file paths. Finally, edit yarn-site.xml so it reads:

yarn-site.xml

<?xml version="1.0" encoding="UTF-8"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>

<configuration>
  <property>
    <name>yarn.nodemanager.aux-services</name>
    <value>mapreduce_shuffle</value>
  </property>
  <property>
    <name>yarn.nodemanager.auxservices.mapreduce.shuffle.class</name>  
    <value>org.apache.hadoop.mapred.ShuffleHandler</value>
  </property>
</configuration>

The last thing we need to do is create the directories that we referenced in hdfs-site.xml:

Patch Hadoop

Now, you need to apply a patch created by and posted to GitHub by user s911415. (Note that this patch is specific to the version of Hadoop that you're installing. Patches for other versions can be found here and here.) We need, in particular hadoop.dll and winutils.exe from these patches.

Download the entire directory at that link as a *.zip and extract the directory bin. Replace C:\Hadoop\hadoop-3.1.1\bin with this new directory. (If you want, save the old directory by renaming it to bin.old.)

Now, if you run hdfs namenode -format in cmd, you should see:

One more thing to do: copy hadoop-yarn-server-timelineservice-3.0.3 from C:\Hadoop\hadoop-3.1.1\share\hadoop\yarn\timelineservice to C:\Hadoop\hadoop-3.1.1\share\hadoop\yarn (the parent directory).

Boot HDFS

Finally, you can boot HDFS by running start-dfs.cmd and start-yarn.cmd in cmd:

You should verify that the namenode, datanode, resourcemanager, and nodemanager are all running. You can also open localhost:8088 and localhost:9087 in your browser to monitor your shiny, new Hadoop Distributed File System:

Finally, test that you can edit the filesystem by running hadoop fs -mkdir /test, which will make a directory called test in the root directory:

Testing Hadoop and Spark

We know now how to create directories (fs -mkdir) and list their contents (fs -ls) in HDFS, what about creating and editing files? Well, files can be copied from the local file system to HDFS with fs -put. We can then read files in the spark-shell with sc.textFile(...):

Note that you read a file from HDFS on hdfs://localhost:9000/ and not just hdfs://. This is because this is the defaultFS we defined in core-site.xml.

So there you have it! Spark running on Windows, reading files stored in HDFS. This took a bit of work to get going and I owe a lot to people who previously encountered the same bugs as me, or previously wrote tutorials which I used as a framework for this walkthrough. Here are the blogs, GitHub repos, and SO posts I used to build this tutorial: