Hadoop Filesystem Implementation for GEDS

This project implements the Hadoop Fileystem API for GEDS.

Design

Apache Spark and Hadoop interact with GEDS in the following way:

Building

Make sure SBT is installed on the system. Install GEDS and then run the package script:

export GEDS_INSTALL=$HOME/geds-install # GEDS install path
sbt package

The library will be located in target/scala-2.12/.

Testing

The tests can be executed by running sbt test. The backing object store can be configured with the following environment variables:

• TEST_BUCKET: Bucket for the storage location (default: test)
• AWS_ACCESS_KEY_ID: The access key for the object store (default: minioadmin)
• AWS_SECRET_ACCESS_KEY: The secret access key for the object store (default: minioadmin)
• AWS_ENDPOINT_URL: The URL of the backing object store (http://10.100.0.15:9000)

Configuration

Place the GEDS-HDFS plugin, geds.jar into the Java class path, and libgeds_java.so into the LD_LIBRARY path. GEDS can then be configured with the following variables:

• fs.geds.impl: com.ibm.geds.hdfs.GEDSHadoopFileSystem - Required
• fs.geds.metadataserver: Ip or DNS of the metadata server - Required (can be alternatively configured by setting the GEDS_METADATASERVER environment variable)
• fs.geds.cache_block_size: GEDS cache block size in bytes. Example: 33554432 - Optional (can be alternatively configured by setting the GEDS_CACHE_BLOCK_SIZE environment variable)
• fs.geds.path: Local GEDS path for ephemeral data. A path that ends with XXXXXX will be randomized with mktempd. Default: /tmp/GEDSHadoop - Optional (can be alternatively configured by setting the GEDS_LOCAL_STORAGE_PATH environment variable).
• fs.geds.port: Local port for the GEDS service. - Optional (can be alternatively configured by setting the GEDS_PORT environment variable)
• fs.geds.http_server_port: Local port for the HTTP service. - Optional (can be alternatively configured by setting the GEDS_HTTP_SERVER_PORT environment variable)

GEDS allows mapping individual buckets to S3. For each bucket, the following configuration variables can be passed to enable a S3 mapping.

• fs.geds.BUCKET_NAME.accessKey: S3 Access key for BUCKET_NAME.
• fs.geds.BUCKET_NAME.secretKey: S3 Secret key for BUCKET_NAME.
• fs.geds.BUCKET_NAME.endpoint: S3 Endpoint to use for BUCKET_NAME.

Example Spark configuration

The bucket zrlio-tpcds is mapped to S3 to enable caching.

  --conf spark.hadoop.fs.geds.impl="com.ibm.geds.hdfs.GEDSHadoopFileSystem"
  --conf spark.hadoop.fs.geds.metadataserver="geds-service:40001"
  --conf spark.hadoop.fs.geds.blocksize=$((32*1024*1024))
  --conf spark.hadoop.fs.geds.path=/geds
  --conf spark.hadoop.fs.geds.zrlio-tpcds.accessKey="${S3A_ACCESS_KEY}"
  --conf spark.hadoop.fs.geds.zrlio-tpcds.secretKey="${S3A_SECRET_KEY}"
  --conf spark.hadoop.fs.geds.zrlio-tpcds.endpoint="${S3A_ENDPOINT}"

Implementation Notes

FileSystem construction

The Hadoop FileSystem constructor is called once per hostname in the filesystem URI.

• For s3a this means, that for every bucket there is a separate instance of the S3AFilesystem. Each bucket is accessed through s3a://bucket/key. For each bucket we can define different S3 credentials or an endpoint.
• For crail this means, that the constructor for CrailHDFS is called once, since the Crail filesystem is based on the URL to the namenode. There is typically only one Crail namenode running. Crail filesystems are typically defined with crail://namenode:port/.
• For geds we want to emulate the s3a model to keep the bucket configuration, but forward the calls to a single GEDS instance. This allows us to add additional configuration (for example S3 mapping) for each bucket. Each GEDSHadoopFileSystem instance will forward the calls to a GEDS singleton. The singleton also allows us to more efficiently cache files locally.

Folders

Hadoop implements folders on Key/Value filesystems by placing an empty key with the name _$folder$.

Example: A folder structure Nase/Baer would create the following keys on the key/value filesystem:

• Nase/_$folder$
• Nase/Baer/_$folder$

Rename

The Hadoop API specifies the following API for renaming paths (source):

  /**
   * Renames Path src to Path dst.
   * @param src path to be renamed
   * @param dst new path after rename
   * @throws IOException on failure
   * @return true if rename is successful
   */
  public abstract boolean rename(Path src, Path dst) throws IOException;

• Rename needs to return true if the command has been successful
• An IOException otherwise.

Sample implementations

Below is a list of projects that implement the Hadoop Filesystem which we can use as a reference.

• s3a:// S3AFileSystem

  Minimal config:

  --conf spark.hadoop.fs.s3a.access.key=Nase
  --conf spark.hadoop.fs.s3a.secret.key=Baer
  --conf spark.hadoop.fs.s3a.endpoint=http://endpoint/
  --conf spark.hadoop.fs.s3a.impl=org.apache.hadoop.fs.s3a.S3AFileSystem
  

• crail:// CrailHDFS

  Minimal config:

  <property>
   <name>fs.crail.impl</name>
   <value>org.apache.crail.hdfs.CrailHadoopFileSystem</value>
  </property>
  <property>
    <name>fs.defaultFS</name>
    <value>crail://localhost:9060</value>
  </property>
  <property>
    <name>fs.AbstractFileSystem.crail.impl</name>
    <value>org.apache.crail.hdfs.CrailHDFS</value>
  </property>

  Note: Crail is special since it defines fs.defaultFS and uses AbstractFileSystem as a base implementation. For our use-case we want to model the s3a-approach:

  • We don't want to override fs.defaultFS
  • GEDSHadoopFileSystem should inherit from org.apache.hadoop.fs.FileSystem directly

• https:// AbstractHttpFileSystem

  Note: This filesystem does not allow listing objects.