File Systems

Apache Flink uses file systems to consume and persistently store data, both for the results of applications and for fault tolerance and recovery. These are some of most of the popular file systems, including local, hadoop-compatible, Amazon S3, Aliyun OSS and Azure Blob Storage.

The file system used for a particular file is determined by its URI scheme. For example, file:///home/user/text.txt refers to a file in the local file system, while hdfs://namenode:50010/data/user/text.txt is a file in a specific HDFS cluster.

File system instances are instantiated once per process and then cached/pooled, to avoid configuration overhead per stream creation and to enforce certain constraints, such as connection/stream limits.

Local File System

Flink has built-in support for the file system of the local machine, including any NFS or SAN drives mounted into that local file system. It can be used by default without additional configuration. Local files are referenced with the file:// URI scheme.

Pluggable File Systems

The Apache Flink project supports the following file systems:

  • Amazon S3 object storage is supported by two alternative implementations: flink-s3-fs-presto and flink-s3-fs-hadoop. Both implementations are self-contained with no dependency footprint.

  • Aliyun Object Storage Service is supported by flink-oss-fs-hadoop and registered under the oss:// URI scheme. The implementation is based on the Hadoop Project but is self-contained with no dependency footprint.

  • Azure Data Lake Store Gen2 is supported by flink-azure-fs-hadoop and registered under the abfs(s):// URI schemes. The implementation is based on the Hadoop Project but is self-contained with no dependency footprint.

  • Azure Blob Storage is supported by flink-azure-fs-hadoop and registered under the wasb(s):// URI schemes. The implementation is based on the Hadoop Project but is self-contained with no dependency footprint.

  • Google Cloud Storage is supported by gcs-connector and registered under the gs:// URI scheme. The implementation is based on the Hadoop Project but is self-contained with no dependency footprint.

You can and should use any of them as plugins.

To use a pluggable file systems, copy the corresponding JAR file from the opt directory to a directory under plugins directory of your Flink distribution before starting Flink, e.g.

  1. mkdir ./plugins/s3-fs-hadoop
  2. cp ./opt/flink-s3-fs-hadoop-1.20.0.jar ./plugins/s3-fs-hadoop/

Attention The plugin mechanism for file systems was introduced in Flink version 1.9 to support dedicated Java class loaders per plugin and to move away from the class shading mechanism. You can still use the provided file systems (or your own implementations) via the old mechanism by copying the corresponding JAR file into lib directory. However, since 1.10, s3 plugins must be loaded through the plugin mechanism; the old way no longer works as these plugins are not shaded anymore (or more specifically the classes are not relocated since 1.10).

It’s encouraged to use the plugins-based loading mechanism for file systems that support it. Loading file systems components from the lib directory will not supported in future Flink versions.

Adding a new pluggable File System implementation

File systems are represented via the org.apache.flink.core.fs.FileSystem class, which captures the ways to access and modify files and objects in that file system.

To add a new file system:

  • Add the File System implementation, which is a subclass of org.apache.flink.core.fs.FileSystem.
  • Add a factory that instantiates that file system and declares the scheme under which the FileSystem is registered. This must be a subclass of org.apache.flink.core.fs.FileSystemFactory.
  • Add a service entry. Create a file META-INF/services/org.apache.flink.core.fs.FileSystemFactory which contains the class name of your file system factory class (see the Java Service Loader docs for more details).

During plugins discovery, the file system factory class will be loaded by a dedicated Java class loader to avoid class conflicts with other plugins and Flink components. The same class loader should be used during file system instantiation and the file system operation calls.

In practice, it means you should avoid using Thread.currentThread().getContextClassLoader() class loader in your implementation.

Hadoop File System (HDFS) and its other implementations

For all schemes where Flink cannot find a directly supported file system, it falls back to Hadoop. All Hadoop file systems are automatically available when flink-runtime and the Hadoop libraries are on the classpath.

This way, Flink seamlessly supports all of Hadoop file systems implementing the org.apache.hadoop.fs.FileSystem interface, and all Hadoop-compatible file systems (HCFS).

  • HDFS (tested)
  • Alluxio (tested, see configuration specifics below)
  • XtreemFS (tested)
  • FTP via Hftp (not tested)
  • HAR (not tested)

The Hadoop configuration has to have an entry for the required file system implementation in the core-site.xml file. See example for Alluxio.

We recommend using Flink’s built-in file systems unless required otherwise. Using a Hadoop File System directly may be required, for example, when using that file system for YARN’s resource storage, via the fs.defaultFS configuration property in Hadoop’s core-site.xml.

Alluxio

For Alluxio support add the following entry into the core-site.xml file:

  1. <property>
  2. <name>fs.alluxio.impl</name>
  3. <value>alluxio.hadoop.FileSystem</value>
  4. </property>