[
https://issues.apache.org/jira/browse/HDFS-12748?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16226534#comment-16226534
]
Weiwei Yang commented on HDFS-12748:
------------------------------------
Hi [~yangjiandan]
Thanks for reporting this issue, the leak must be fixed. I took a look at
{{NamenodeWebHdfsMethods}}, there are currently 2 places calling
FileSystem.get() without closing. Can we fix them both?
bq. which contains 7,844,890 DistributedFileSystem
First impression I was thinking the {{FileSystem#Cache}} is not working
properly, but after looking into the code, I think it is working as expected.
The cache stores FileSystem instances distinguished by {{FileSystem.Cache.Key}}
which includes a UGI field. According to HADOOP-6670, UGI is distinguished by
the {{Subject}}. In webhdfs, UGI is parsed by {{DataNodeUGIProvider}} per http
request, this way {{Subject}} is an different instance every time so it will be
new to the cache. That was why the cache has so many instances if not properly
closing.
> NameNode memory leak when accessing webhdfs GETHOMEDIRECTORY
> ------------------------------------------------------------
>
> Key: HDFS-12748
> URL: https://issues.apache.org/jira/browse/HDFS-12748
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: hdfs
> Affects Versions: 2.8.2
> Reporter: Jiandan Yang
>
> In our production environment, the standby NN often do fullgc, through mat we
> found the largest object is FileSystem$Cache, which contains 7,844,890
> DistributedFileSystem.
> By view hierarchy of method FileSystem.get() , I found only
> NamenodeWebHdfsMethods#get call FileSystem.get(). I don't know why creating
> different DistributedFileSystem every time instead of get a FileSystem from
> cache.
> {code:java}
> case GETHOMEDIRECTORY: {
> final String js = JsonUtil.toJsonString("Path",
> FileSystem.get(conf != null ? conf : new Configuration())
> .getHomeDirectory().toUri().getPath());
> return Response.ok(js).type(MediaType.APPLICATION_JSON).build();
> }
> {code}
> When we close FileSystem when GETHOMEDIRECTORY, NN don't do fullgc.
> {code:java}
> case GETHOMEDIRECTORY: {
> FileSystem fs = null;
> try {
> fs = FileSystem.get(conf != null ? conf : new Configuration());
> final String js = JsonUtil.toJsonString("Path",
> fs.getHomeDirectory().toUri().getPath());
> return Response.ok(js).type(MediaType.APPLICATION_JSON).build();
> } finally {
> if (fs != null) {
> fs.close();
> }
> }
> }
> {code}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
