The EMRFS S3-optimized committer and multipart uploads - Amazon EMR
Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

The EMRFS S3-optimized committer and multipart uploads

To use the EMRFS S3-optimized committer, you must enable multipart uploads for Amazon EMR . Multipart uploads are enabled by default. You can re-enable it if required. For more information, see Configure multipart upload for Amazon S3 in the Amazon EMR Management Guide.

The EMRFS S3-optimized committer uses the transaction-like characteristics of multipart uploads to ensure files written by task attempts only appear in the job's output location upon task commit. By using multipart uploads in this way, the committer improves task commit performance over the default FileOutputCommitter algorithm version 2. When using the EMRFS S3-optimized committer, there are some key differences from traditional multipart upload behavior to consider:

  • Multipart uploads are always performed regardless of the file size. This differs from the default behavior of EMRFS, where the fs.s3n.multipart.uploads.split.size property controls the file size at which multipart uploads are triggered.

  • Multipart uploads are left in an incomplete state for a longer period of time until the task commits or aborts. This differs from the default behavior of EMRFS where a multipart upload completes when a task finishes writing a given file.

Because of these differences, if a Spark Executor JVM crashes or is killed while tasks are running and writing data to Amazon S3, incomplete multipart uploads are more likely to be left behind. For this reason, when you use the EMRFS S3-optimized committer, be sure to follow the best practices for managing failed multipart uploads. For more information, see Best practices for working with Amazon S3 buckets in the Amazon EMR Management Guide.