• Type: New Feature
    • Status: Resolved
    • Priority: Major
    • Resolution: Duplicate
    • Affects Version/s: None
    • Fix Version/s: None
    • Component/s: fs/s3
    • Labels:


      Object stores do not have an efficient rename operation, which is used by the Hadoop FileOutputCommitter to atomically promote the "winning" attempt out of the multiple (speculative) attempts to the final path. These slow job commits are one of the main friction points when using object stores in Hadoop.There have been quite some attempts at resolving this: HADOOP-9565, Apache Spark DirectOutputCommitters, ... but they have proven not to be robust in face of adversity (network partitions, ...).

      The current ticket proposes to do the atomic commit by using the S3 Multipart API, which allows multiple concurrent uploads on the same objectname, each in its own "temporary space, identified by the UploadId which is returned as a response to InitiateMultipartUpload. Every attempt writes directly to the final outputPath. Data is uploaded using Put Part and as a response an ETag for the part is returned and stored. The CompleteMultipartUpload is postponed. Instead, we persist the UploadId (using a _temporary subdir or elsewhere) and the ETags. When a certain "job" wins CompleteMultipartUpload is called for each of its files using the proper list of Part ETags.

      Completing a MultipartUpload is a metadata only operation (internally in S3) and is thus orders of magnitude faster than the rename-based approach which moves all the data.

      Required work:

      • Expose the multipart initiate and complete calls in S3AOutputStream to S3AFilesystem
      • Use these multipart calls in a custom committer as described above. I propose to build on the S3ACommitter Steve Loughran is doing for HADOOP-13786


          Issue Links



              • Assignee:
                Thomas Demoor Thomas Demoor
                Thomas Demoor Thomas Demoor
              • Votes:
                0 Vote for this issue
                8 Start watching this issue


                • Created: