In order to build Flink, you need the source code. Either download the source of a release or clone the git repository. In addition to that, you need Maven 3 and a JDK (Java Development Kit). Flink requires at least Java 7 to build. We recommend using Java 8.
To clone from git, enter:
git clone {{ site.github_url }}
The simplest way of building Flink is by running:
cd flink mvn clean install -DskipTests
This instructs Maven (mvn) to first remove all existing builds (clean) and then create a new Flink binary (install). The -DskipTests command prevents Maven from executing the unit tests.
Read more about Apache Maven.
This section covers building Flink for a specific Hadoop version. Most users do not need to do this manually. The download page of Flink contains binary packages for common setups.
The problem is that Flink uses HDFS and YARN which are both dependencies from Apache Hadoop. There exist many different versions of Hadoop (from both the upstream project and the different Hadoop distributions). If a user is using a wrong combination of versions, exceptions like this one occur:
ERROR: Job execution failed. org.apache.flink.runtime.client.JobExecutionException: Cannot initialize task 'TextInputFormat(/my/path)': java.io.IOException: Failed on local exception: com.google.protobuf.InvalidProtocolBufferException: Protocol message contained an invalid tag (zero).; Host Details :
There are two main versions of Hadoop that we need to differentiate:
By default, Flink is using the Hadoop 2 dependencies.
To build Flink for Hadoop 1, issue the following command:
mvn clean install -DskipTests -Dhadoop.profile=1
The -Dhadoop.profile=1 flag instructs Maven to build Flink for Hadoop 1. Note that the features included in Flink change when using a different Hadoop profile. In particular the support for YARN and the build-in HBase support are not available in Hadoop 1 builds.
You can also specify a specific Hadoop version to build against:
mvn clean install -DskipTests -Dhadoop.version=2.4.1
To build Flink against a vendor specific Hadoop version, issue the following command:
mvn clean install -DskipTests -Pvendor-repos -Dhadoop.version=2.2.0-cdh5.0.0-beta-2
The -Pvendor-repos activates a Maven build profile that includes the repositories of popular Hadoop vendors such as Cloudera, Hortonworks, or MapR.
Build Flink for hadoop2 versions before 2.2.0
Maven will automatically build Flink with its YARN client. But there were some changes in Hadoop versions before the 2.2.0 Hadoop release that are not supported by Flink's YARN client. Therefore, you can disable building the YARN client with the following string: -P!include-yarn.
So if you are building Flink for Hadoop 2.0.0-alpha, use the following command:
-P!include-yarn -Dhadoop.version=2.0.0-alpha
Note: Users that purely use the Java APIs and libraries can ignore this section.
Flink has APIs, libraries, and runtime modules written in Scala. Users of the Scala API and libraries may have to match the Scala version of Flink with the Scala version of their projects (because Scala is not strictly backwards compatible).
By default, Flink is built with the Scala 2.10. To build Flink with Scala 2.11, you need to change the default Scala binary version with a build script:
# Switch Scala binary version between 2.10 and 2.11 tools/change-scala-version.sh 2.11 # Build and install locally mvn clean install -DskipTests
To build against custom Scala versions, you need to switch to the appropriate binary version and supply the language version as additional build property. For example, to buid against Scala 2.11.4, you have to execute:
# Switch Scala binary version to 2.11 tools/change-scala-version.sh 2.11 # Build with custom Scala version 2.11.4 mvn clean install -DskipTests -Dscala.version=2.11.4
Flink is developed against Scala 2.10 and tested additionally against Scala 2.11. These two versions are known to be compatible. Earlier versions (like Scala 2.9) are not compatible.
Newer versions may be compatible, depending on breaking changes in the language features used by Flink, and the availability of Flink's dependencies in those Scala versions. The dependencies written in Scala include for example Kafka, Akka, Scalatest, and scopt.
If your home directory is encrypted you might encounter a java.io.IOException: File name too long exception. Some encrypted file systems, like encfs used by Ubuntu, do not allow long filenames, which is the cause of this error.
The workaround is to add:
<args> <arg>-Xmax-classfile-name</arg> <arg>128</arg> </args>
in the compiler configuration of the pom.xml file of the module causing the error. For example, if the error appears in the flink-yarn module, the above code should be added under the <configuration> tag of scala-maven-plugin. See this issue for more information.
The builds with Maven are controlled by properties and build profiles. There are two profiles, one for hadoop1 and one for hadoop2. When the hadoop2 profile is enabled (default), the system will also build the YARN client.
To enable the hadoop1 profile, set -Dhadoop.profile=1 when building. Depending on the profile, there are two Hadoop versions, set via properties. For “hadoop1”, we use 1.2.1 by default, for “hadoop2” it is 2.2.0.
You can change these versions with the hadoop-two.version (or hadoop-one.version) property. For example -Dhadoop-two.version=2.4.0.