web5design/hadoop-s3a

An AWS SDK-backed FileSystem driver for Hadoop

An AWS SDK-backed FileSystem driver for Hadoop

This is an experimental FileSystem for Hadoop that uses the AWS SDK instead of jets3t to connect. It is intended as a replacement for the s3native FileSystem. This has not been heavily tested yet. Use at your own risk.

Features:

• File upload & copy support for files >5 GB
• Significantly faster performance, especially for large files
• Parallel upload support
• Parallel copy (rename)
• AWS S3 explorer compatible empty directories using xyz/ instead of xyz_$folder$
• Ignores _$folder$ files created by s3native and other S3 browsing utilities
• Supports multiple buffer dirs to even out IO

Build Instructions

Build or download aws-java-sdk 1.6.5

copy aws-java-sdk-1.6.5.jar, httpcore-4.2.jar, httpclient-4.2.3.jar (from aws lib) and hadoop-s3a.jar to your hadoop classpath (run hadoop classpath for the appropriate directory)

Build src

$ javac -classpath `hadoop classpath` `find src/main -name '*.java'`
$ jar cvf hadoop-s3a.jar -C src/main/java .

Add the following keys to your core-site.xml file:

<property>
  <name>fs.s3a.awsAccessKeyId</name>
  <value>...</value>
</property>

<property>
  <name>fs.s3a.awsSecretAccessKey</name>
  <value>...</value>
</property>

<property>
  <name>fs.s3a.buffer.dir</name>
  <value>${hadoop.tmp.dir}/s3a</value>
</property>

<property>
  <name>fs.s3a.impl</name>
  <value>org.apache.hadoop.fs.s3a.S3AFileSystem</value>
</property>

You probably want to add this to your log4j.properties file:

log4j.logger.com.amazonaws=ERROR
log4j.logger.com.amazonaws.http.AmazonHttpClient=ERROR
log4j.logger.org.apache.hadoop.fs.s3a.S3AFileSystem=WARN

You should now be able to run commands:

$ hadoop fs -ls s3a://bucketname/foo

Tunable parameters

These may or may not improve performance. The defaults were set without much testing.

• fs.s3a.maxConnections - Controls how many parallel connections HttpClient spawns (default: 50)
• fs.s3a.secureConnections - Enables or disables SSL connections to S3 (default: true)
• fs.s3a.maxErrorRetries - How many times we should retry commands on transient errors (default: 10)
• fs.s3a.socketTimeout - Socket connect timeout (default: 5000)
• fs.s3a.maxPagingKeys - How many keys to request from S3 when doing directory listings at a time (default: 5000)
• fs.s3a.multipartSize - How big (in bytes) to split a upload or copy operation up into (default: 10485760)
• fs.s3a.minMultipartSize - Until a file is this large (in bytes), use non-parallel upload/copy (default: 20971520)
• fs.s3a.cannedACL - Set a canned ACL on newly created/copied objects (private | public-read | public-read-write | authenticated-read | log-delivery-write | bucket-owner-read | bucket-owner-full-control)
• fs.s3a.purgeExistingMultiPart - True if you want to purge existing multipart uploads that may not have been completed/aborted correctly (default: false)
• fs.s3a.purgeExistingMultiPartAge - Minimum age in seconds of multipart uploads to purge (default: 86400)
• fs.s3a.buffer.dir - Comma separated list of directories that will be used to buffer file writes out of (default: uses fs.s3.buffer.dir)

Caveats

Hadoop uses a standard output committer which uploads files as filename.COPYING before renaming them. This can cause unnecessary performance issues with S3 because it does not have a rename operation and S3 verifies uploads against an md5 that the driver sets on the upload request. While this FileSystem should be significantly faster than the built-in s3native driver because of parallel copy support, you may want to consider setting a null output committer on our jobs to further improve performance.

Because S3 requires the file length be known before a file is uploaded, all output is buffered out to a temporary file first similar to the s3native driver.

Due to the lack of native rename() for S3, renaming extremely large files or directories make take a while. Unfortunately, there is no way to notify hadoop that progress is still being made for rename operations, so your job may time out unless you increase the task timeout.

This driver will fully ignore _$folder$ files. This was necessary so that it could interoperate with repositories that have had the s3native driver used on them.

This has only been run under CDH 4, but it should work with other distributions of hadoop. Be sure to watch out for conflicting versions of httpclient.

Statistics for the filesystem may be calcualted differently than other filesystems. When uploading a file, we do not count writing the temporary file on the local filesystem towards the local filesystem's written bytes count. When renaming files, we do not count the S3->S3 copy as read or write operations. Unlike the s3native driver, we only count bytes written when we start the upload (as opposed to the write calls to the temporary local file). The driver also counts read & write ops, but they are done mostly to keep from timing out on large s3 operations.

This is currently implemented as a FileSystem and not a AbstractFileSystem.