[
https://issues.apache.org/jira/browse/HADOOP-16378?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
K S updated HADOOP-16378:
-------------------------
Affects Version/s: 3.3.0
Environment: Ubuntu 18.04, Hadoop 2.7.3 (Though this problem exists
on later versions of Hadoop as well), Java 8 ( + Java 11)
Description:
Bug occurs when Hadoop creates temporary ".nfs*" files as part of file moves
and accesses. If this file is deleted very quickly after being created, a
RuntimeException is thrown. The root cause is in the loadPermissionInfo method
in org.apache.hadoop.fs.RawLocalFileSystem. To get the permission info, it
first does
{code:java}
ls -ld{code}
and then attempts to get permissions info about each file. If a file
disappears between these two steps, an exception is thrown.
*Reproduction Steps:*
An isolated way to reproduce the bug is to run FileInputFormat.listStatus over
and over on the same dir that we’re creating those temp files in. On Ubuntu or
any other Linux-based system, this should fail intermittently
*Fix:*
One way in which we managed to fix this was to ignore the exception being
thrown in loadPemissionInfo() if the exit code is 1 or 2. Alternatively, it's
possible that turning "useDeprecatedFileStatus" off in RawLocalFileSystem would
fix this issue, though we never tested this, and the flag was implemented to
fix HADOOP-9652
was:
Bug occurs when Hadoop creates temporary ".nfs*" files as part of file moves
and accesses. If this file is deleted very quickly after being created, a
RuntimeException is thrown. The root cause is in the loadPermissionInfo method
in org.apache.hadoop.fs.RawLocalFileSystem. To get the permission info, it
first does
{code:java}
ls -ld{code}
and then attempts to get permissions info about each file. If a file
disappears between these two steps, an exception is thrown.
*Reproduction Steps:*
An isolated way to reproduce the bug is to run FileInputFormat.listStatus over
and over on the same dir that we’re creating those temp files in. On Ubuntu or
any other Linux-based system, this should fail intermittently. On MacOS (due to
differences in how `ls` returns status codes) this should not fail.
*Fix:*
One way in which we managed to fix this was to ignore the exception being
thrown in loadPemissionInfo() if the exit code is 1 or 2.
> RawLocalFileStatus throws exception if a file is created and deleted quickly
> ----------------------------------------------------------------------------
>
> Key: HADOOP-16378
> URL: https://issues.apache.org/jira/browse/HADOOP-16378
> Project: Hadoop Common
> Issue Type: Bug
> Components: fs
> Affects Versions: 3.3.0
> Environment: Ubuntu 18.04, Hadoop 2.7.3 (Though this problem exists
> on later versions of Hadoop as well), Java 8 ( + Java 11)
> Reporter: K S
> Priority: Major
>
> Bug occurs when Hadoop creates temporary ".nfs*" files as part of file moves
> and accesses. If this file is deleted very quickly after being created, a
> RuntimeException is thrown. The root cause is in the loadPermissionInfo
> method in org.apache.hadoop.fs.RawLocalFileSystem. To get the permission
> info, it first does
>
> {code:java}
> ls -ld{code}
> and then attempts to get permissions info about each file. If a file
> disappears between these two steps, an exception is thrown.
> *Reproduction Steps:*
> An isolated way to reproduce the bug is to run FileInputFormat.listStatus
> over and over on the same dir that we’re creating those temp files in. On
> Ubuntu or any other Linux-based system, this should fail intermittently
> *Fix:*
> One way in which we managed to fix this was to ignore the exception being
> thrown in loadPemissionInfo() if the exit code is 1 or 2. Alternatively, it's
> possible that turning "useDeprecatedFileStatus" off in RawLocalFileSystem
> would fix this issue, though we never tested this, and the flag was
> implemented to fix HADOOP-9652
>
>
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
