yimeng created YARN-9157:
----------------------------
Summary: Failed deletion dirs in yarn.nodemanager.local-dirs
causes accumulation lots of files under the path yarn.nodemanager.local-dirs
and causes operation systerm's Inode to to be depleted
Key: YARN-9157
URL: https://issues.apache.org/jira/browse/YARN-9157
Project: Hadoop YARN
Issue Type: Bug
Components: applications/distributed-shell
Affects Versions: 3.1.1, 3.0.1, 2.7.5, 2.8.3, 2.7.2
Reporter: yimeng
Attachments: image-2018-12-25-10-03-51-070.png
the Yarn task Excute failed , because excessive number of files under the path
yarn.nodemanager.local-dirs causes Inode to run out and calculates task failure
!image-2018-12-25-09-53-15-067.png!
check the NM Logs , found that many localized dirs delete failed because of
user not found in security Systerm.
actually the local dir files's size is 4.4GB, not 240859897B print in the log
!image-2018-12-25-10-03-51-070.png!
The user not found is because of our userInfo is saved in Ldap DB , when Ldap
Service have problem at some time , then get the user info will fail(not
because the user is deleted).When the Ldap Server recovery at some time , the
user info can get .
The problem is even we can get the user info later , the dirs that deleted
failed before will never be deleted later (it is deleted from the tracker list
), this cause the dirs accumulation .
I think NM ResourceLocalizationService should determine whether the file was
deleted successfully by Deletion Service Thread before deleting the directory
from tracker list and levelDB,if deleted failed ,we should add back it to
tracker list ,then delete the next dirs till the local dirs size is below
yarn.nodemanager.localizer.cache.target-size-mb
.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
