[HADOOP-5861] s3n files are not getting split by default

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.19.1
• Fix Version/s: 0.21.0
• Component/s: fs/s3

Environment

ec2

• Hadoop Flags: Incompatible change, Reviewed
• Release Note: Files stored on the native S3 filesystem (s3n:// URIs) now report a block size determined by the fs.s3n.block.size property (default 64MB).
• Assignee: Thomas White
• Reporter: Joydeep Sen Sarma
• Created: 17/May/09 04:33
• Updated: 24/Aug/10 20:37
• Resolved: 03/Jun/09 11:34

Description

running with stock ec2 scripts against hadoop-19 - i tried to run a job against a directory with 4 text files - each about 2G in size. These were not split (only 4 mappers were run).

The reason seems to have two parts - primarily that S3N files report a block size of 5G. This causes FileInputFormat.getSplits to fall back on goal size (which is totalsize/conf.get("mapred.map.tasks")).Goal Size in this case was 4G - hence the files were not split. This is not an issue with other file systems since the block size reported is much smaller and the splits get based on block size (not goal size).

can we make the S3N files report a more reasonable block size?