[
https://issues.apache.org/jira/browse/HDFS-15806?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17286871#comment-17286871
]
Jinglun commented on HDFS-15806:
--------------------------------
Hi [~ayushtkn], thanks your comments !
{quote}before this was there some kind of memory leak, or these threads were
getting cleared later?
{quote}
In Xiaomi we use the dead node detector feature only for hbase. The HBase
doesn't close the files system and the dfs client. So we haven't notice the
leak before. Recently we found the dead node detector won't remove alive nodes
from the dead node set, as described in HDFS-15809. So I started reviewing the
whole feature and found this leak bug.
{quote}Secondly, for the shutdown is there some specific order, or it is just
random
{quote}
It is random. Most of the threads are connected by queue(the producer-consumer
model). So the order of stopping the producer or the consumer won't be a
problem.
1) The DeadNodeDetector thread is responsible for add nodes from
_suspectAndDeadNodes_ set to _deadNodesProbeQueue_.
2) The _probeDeadNodesSchedulerThr_ is responsible for taking nodes from
_deadNodesProbeQueue_ and __ submit probe tasks to _probeDeadNodesThreadPool_.
3) The _probeSuspectNodesSchedulerThr_ is responsible for taking nodes from
_suspectNodesProbeQueue_ and submit probe tasks to
_probeSuspectNodesThreadPool_.
4) All the probe tasks submit getDatanodeInfo rpc calls in the thread pool
_rpcThreadPool_.
Some other thoughts: the thread model is a little complicated and could be
improved. For example I think we can do the rpc call at the probe task instead
of submitting to rpcThreadPool. I need first figure out the purpose of the
original design then may be start a new Jira for the thread improvement later.
> DeadNodeDetector should close all the threads when it is closed.
> ----------------------------------------------------------------
>
> Key: HDFS-15806
> URL: https://issues.apache.org/jira/browse/HDFS-15806
> Project: Hadoop HDFS
> Issue Type: Bug
> Reporter: Jinglun
> Assignee: Jinglun
> Priority: Major
> Attachments: HDFS-15806.001.patch
>
>
> The DeadNodeDetector doesn't close all the threads when it is closed. This
> Jira trys to fix this.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
