hadoop-common-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Pieter Reuse (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (HADOOP-11262) Enable YARN to use S3A
Date Fri, 03 Jul 2015 08:16:04 GMT

     [ https://issues.apache.org/jira/browse/HADOOP-11262?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Pieter Reuse updated HADOOP-11262:
----------------------------------
    Attachment: HADOOP-11262-6.patch

Patch 6:

As requested, expanded patch 5 with tests by extending the following tests, overriding them
with S3A specifics:
* _TestFileContext.java_
* _FileContextCreateMkdirBaseTest.java_
* _FileContextMainOperationsBaseTest.java_
* _FCStatisticsBaseTest.java_
* _FileContextURIBase.java_
* _FileContextUtilBase.java_

In doing so, fixed following bugs in _FileContextMainOperationsBaseTest.java_:
* _line 1169_: creating a symlink on an FS that doesn't support this throws an _UnsupportedOperationException_,
not an _IOException_ (see _FileContext.java:1441_).
* _lines 1252 and 1313_: the contract of _read()_ is not to read the whole file - that's the
contract of _readFully()_. For this reason tests assuming that the whole file has been read
should use _readFully()_ instead of _read()_.

And added an enhancement for object storage systems in the same file:
* _line 1238_: an object storage system throws an _IOException_ as a file does not exist *before*
the file is closed (nor does it have a checksum at that moment). This object-storage issue
is resolved by changing the order of _fc.setVerifyChecksum(true, path)_ and _out.write(data,
0, data.length)_, while this does not impact the behaviour on hdfs or other file systems.

Discovered and patched the following related bugs in S3A:
* Bugfix in _S3AFileSystem.java_: ports on s3 should be ignored, which corresponds with a
value of -1 (instead of the default 0 in FileSystem).
* Another bugfix is in _S3AFileStatus.java_: _getModificationTime()_ is overwritten for directories.
It returns _System.currentTimeMillis()_ because an ObjectStore does not keep track of modification-times
of directories. Because some parts of the Hadoop ecosystem use modification time to ignore
or delete "old" directories (e.g. the YarnHistorySever), returning 0 for directories is not
the best option here.

Added _TestS3AMiniYarnCluster.java_, which runs a simple _WordCount_-MapReduce job on a _YarnMiniCluster_
using S3A as filesystem.

> Enable YARN to use S3A 
> -----------------------
>
>                 Key: HADOOP-11262
>                 URL: https://issues.apache.org/jira/browse/HADOOP-11262
>             Project: Hadoop Common
>          Issue Type: Sub-task
>          Components: fs/s3
>            Reporter: Thomas Demoor
>            Assignee: Pieter Reuse
>              Labels: amazon, s3
>         Attachments: HADOOP-11262-2.patch, HADOOP-11262-3.patch, HADOOP-11262-4.patch,
HADOOP-11262-5.patch, HADOOP-11262-6.patch, HADOOP-11262.patch
>
>
> Uses DelegateToFileSystem to expose S3A as an AbstractFileSystem.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message