For months now we've been finding that we've been experiencing frequent data truncation issues when reading from S3 using the s3n:// protocol. I finally was able to gather some debugging output on the issue in a job I ran last night, and so can finally file a bug report.
The job I ran last night was on a 16-node cluster (all of them AWS EC2 cc2.8xlarge machines, running Ubuntu 13.04 and Cloudera CDH4.3.0). The job was a Hadoop streaming job, which reads through a large number (i.e., ~55,000) of files on S3, each of them approximately 300K bytes in size.
All of the files contain 46 columns of data in each record. But I added in an extra check in my mapper code to count and verify the number of columns in every record - throwing an error and crashing the map task if the column count is wrong.
If you look in the attached task logs, you'll see 2 attempts on the same task. The first one fails due to data truncated (i.e., my job intentionally fails the map task due to the current record failing the column count check). The task then gets retried on a different machine and runs to a succesful completion.
You can see further evidence of the truncation further down in the task logs, where it displays the count of the records read: the failed task says 32953 records read, while the successful task says 63133.
Any idea what the problem might be here and/or how to work around it? This issue is a very common occurrence on our clusters. E.g., in the job I ran last night before I had gone to bed I had already encountered 8 such failuers, and the job was only 10% complete. (~25,000 out of ~250,000 tasks.)
I realize that it's common for I/O errors to occur - possibly even frequently - in a large Hadoop job. But I would think that if an I/O failure (like a truncated read) did occur, that something in the underlying infrastructure code (i.e., either in NativeS3FileSystem or in jets3t) should detect the error and throw an IOException accordingly. It shouldn't be up to the calling code to detect such failures, IMO.