[
https://issues.apache.org/jira/browse/YARN-3591?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14608089#comment-14608089
]
Lavkesh Lahngir commented on YARN-3591:
---------------------------------------
Thanks [~jlowe] and [~zxu] for detailed analysis and reviews.
Honestly it has become more evolved than I thought.
Few comments:
1. I wrote a sample program to just check the penalty we will hit in terms of
time. File.exists() along with listing on the parent(initial patch) virtually
adds nothing. Combined time taken for both calls is around 0.1 ms. (This patch
we applied in our production). This will just remove the entry from the map,
which will not affect the running containers. This solves the problem of
failing new containers.
2. The latest patch which checks if the resource path exists in one of the good
disks (basically some string comparison) has major performance implications. It
takes around 40 ms. No way we could incur that.
3. If the file does not exists or it is localized on a bad disk. We need to
keep track of those as well to remove them from the disk as suggested in the
Jason's comment. We can't delete blindly from the disk if refcount is greater
than one.
Can we logically separate the original problem and related problem of zombie
files and address them in separate JIRA?
> Resource Localisation on a bad disk causes subsequent containers failure
> -------------------------------------------------------------------------
>
> Key: YARN-3591
> URL: https://issues.apache.org/jira/browse/YARN-3591
> Project: Hadoop YARN
> Issue Type: Bug
> Affects Versions: 2.7.0
> Reporter: Lavkesh Lahngir
> Assignee: Lavkesh Lahngir
> Attachments: 0001-YARN-3591.1.patch, 0001-YARN-3591.patch,
> YARN-3591.2.patch, YARN-3591.3.patch, YARN-3591.4.patch, YARN-3591.5.patch
>
>
> It happens when a resource is localised on the disk, after localising that
> disk has gone bad. NM keeps paths for localised resources in memory. At the
> time of resource request isResourcePresent(rsrc) will be called which calls
> file.exists() on the localised path.
> In some cases when disk has gone bad, inodes are stilled cached and
> file.exists() returns true. But at the time of reading, file will not open.
> Note: file.exists() actually calls stat64 natively which returns true because
> it was able to find inode information from the OS.
> A proposal is to call file.list() on the parent path of the resource, which
> will call open() natively. If the disk is good it should return an array of
> paths with length at-least 1.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
