[
https://issues.apache.org/jira/browse/HADOOP-10610?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13998569#comment-13998569
]
Steve Loughran commented on HADOOP-10610:
-----------------------------------------
The overall concept makes sense
# {{LocalDirAllocator}} contains the logic to worry about file capacity,
writeability &c. The more disks you list, the more likely a disk is to fail,
the more you need that code. This patch currently just bails out if one dest
dir isn't there, even if others may be present.
# This would be an ideal time to move {{"fs.s3.buffer.dir"}} from an inline
string to a constant where it can be referred to in hadoop and external code
Recommended tests
# a test with > 1 directory in the args
# basic handling of "erroneous" inputs. e.g trailing commas in options
# bad directories in input paths
> Upgrade S3n s3.fs.buffer.dir to suppoer multi directories
> ---------------------------------------------------------
>
> Key: HADOOP-10610
> URL: https://issues.apache.org/jira/browse/HADOOP-10610
> Project: Hadoop Common
> Issue Type: Improvement
> Components: fs/s3
> Affects Versions: 2.4.0
> Reporter: Ted Malaska
> Assignee: Ted Malaska
> Priority: Minor
> Attachments: HDFS-6383.patch
>
>
> s3.fs.buffer.dir defines the tmp folder where files will be written to before
> getting sent to S3. Right now this is limited to a single folder which
> causes to major issues.
> 1. You need a drive with enough space to store all the tmp files at once
> 2. You are limited to the IO speeds of a single drive
> This solution will resolve both and has been tested to increase the S3 write
> speed by 2.5x with 10 mappers on hs1.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
