[
https://issues.apache.org/jira/browse/HDFS-13769?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16559171#comment-16559171
]
Yiqun Lin edited comment on HDFS-13769 at 7/27/18 2:58 AM:
-----------------------------------------------------------
{quote}
Also clear checkpoint in trash is a typical situation of deleting a large dir,
since the checkpoint dir of trash accumulates deleted files within several
hours.
{quote}
Agree. We also met this problem. There is a big chance the checkpoint dir being
a large dir. As [~kihwal] mentioned, for the safe deleting, it might not be a
atomic operation. But it should be okay using for clearing trash dir.
{quote}
Wei-Chiu Chuang, Agree! getContentSummary is a recursive method and it may take
several seconds if the dir is very large. getContentSummary holds the read-lock
in FSNameSystem rather than the write-lock. Also we need a way to know whether
a dir is large. If there is a better solution I don't know, please tell me, and
I think it need not to be very accurate.
{quote}
I am thinking for this, We can skip invoking expensive call
{{getContentSummary}} for in first level dir since there will be a large chance
as a big dir. For the child paths, we can do as current patch did. This might a
better way I think.
was (Author: linyiqun):
{quote}
Also clear checkpoint in trash is a typical situation of deleting a large dir,
since the checkpoint dir of trash accumulates deleted files within several
hours.
{quote}
Agree. We also met this problem. There is a big chance the checkpoint dir being
a large dir. As [~kihwal] mentioned, for the safe deleting, it might not be a
atomic operation. But it should be okay using for clearing trash dir.
{quote}
Wei-Chiu Chuang, Agree! getContentSummary is a recursive method and it may take
several seconds if the dir is very large. getContentSummary holds the read-lock
in FSNameSystem rather than the write-lock. Also we need a way to know whether
a dir is large. If there is a better solution I don't know, please tell me, and
I think it need not to be very accurate.
{quote}
I am thinking for this, we don't really need a limitation value
{{FS_TRASH_SAFE_DELETE_ITEM_LIMIT_KEY}}. I mean if users enable the
safe-deleteion trash policy, we are assuming the trash dir will have a big
chance being a large dir. And we just use safe deletion way in
{{deleteTrashInternal#safeDelete}}. And no need to invoke expensive call
{{getContentSummary}} to get the counts.
> Namenode gets stuck when deleting large dir in trash
> ----------------------------------------------------
>
> Key: HDFS-13769
> URL: https://issues.apache.org/jira/browse/HDFS-13769
> Project: Hadoop HDFS
> Issue Type: Improvement
> Affects Versions: 2.8.2, 3.1.0
> Reporter: Tao Jie
> Assignee: Tao Jie
> Priority: Major
> Attachments: HDFS-13769.001.patch
>
>
> Similar to the situation discussed in HDFS-13671, Namenode gets stuck for a
> long time when deleting trash dir with a large mount of data. We found log in
> namenode:
> {quote}
> 2018-06-08 20:00:59,042 INFO namenode.FSNamesystem
> (FSNamesystemLock.java:writeUnlock(252)) - FSNamesystem write lock held for
> 23018 ms via
> java.lang.Thread.getStackTrace(Thread.java:1552)
> org.apache.hadoop.util.StringUtils.getStackTrace(StringUtils.java:1033)
> org.apache.hadoop.hdfs.server.namenode.FSNamesystemLock.writeUnlock(FSNamesystemLock.java:254)
> org.apache.hadoop.hdfs.server.namenode.FSNamesystem.writeUnlock(FSNamesystem.java:1567)
> org.apache.hadoop.hdfs.server.namenode.FSNamesystem.delete(FSNamesystem.java:2820)
> org.apache.hadoop.hdfs.server.namenode.NameNodeRpcServer.delete(NameNodeRpcServer.java:1047)
> {quote}
> One simple solution is to avoid deleting large data in one delete RPC call.
> We implement a trashPolicy that divide the delete operation into several
> delete RPCs, and each single deletion would not delete too many files.
> Any thought? [~linyiqun]
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
