hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ming Ma (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-7521) Refactor DN state management
Date Tue, 16 Dec 2014 23:41:13 GMT

    [ https://issues.apache.org/jira/browse/HDFS-7521?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14249167#comment-14249167
] 

Ming Ma commented on HDFS-7521:
-------------------------------

Thanks [~wheat9], [~yzhangal] for the comments.

* Regarding the state machine abstraction, the patch has DN's state modified asynchronously
by the dispatcher thread instead of synchronously by the caller thread. The motivation is
to have one thread modify DN's state and might help to simplify the lock management in NN.
But not sure if that is really worthwhile. Changing from async to sync is pretty straightforward.
* IMO, reusing existing state machine lib is beneficial. It declares how states transition
and any actions required. If you look at the state machine lib's internal implementation,
it is similar to we would have implemented. Another nice thing of reusing existing state machine
lib is you can generate the state machine diagram easily.
* For  DECOMMISSIONING states, AFAIK, [~andrew.wang] and [~zhz] made some changes in that
area. a) If the node is dead when decomm is requested, go directly to DECOMMISSIONED state.
b) If their is no under replicated block in the whole cluster, transition the node from DECOMMISSIONING
to DECOMMISSIONED. c) Work in progress in HDFS-7411 to track blocks at individual DECOMMISSIONING
node level. This patch doesn't change any of that.

> Refactor DN state management
> ----------------------------
>
>                 Key: HDFS-7521
>                 URL: https://issues.apache.org/jira/browse/HDFS-7521
>             Project: Hadoop HDFS
>          Issue Type: Improvement
>            Reporter: Ming Ma
>         Attachments: DNStateMachines.png, HDFS-7521.patch
>
>
> There are two aspects w.r.t. DN state management in NN.
> * State machine management within active NN
> NN maintains states of each data node regarding whether it is running or being decommissioned.
But the state machine isn’t well defined. We have dealt with some corner case bug in this
area. It will be useful if we can refactor the code to use clear state machine definition
that define events, available states and actions for state transitions. It has these benefits.
> ** Make it easy to define correctness of DN state management. Currently some of the state
transitions aren't defined in the code. For example, if admins remove a node from include
host file while the node is being decommissioned, it will be transitioned to DEAD and DECOMM_IN_PROGRESS.
That might not be the intention. If we have state machine definition, we can identify this
case.
> ** Make it easy to add new state for DN later. For example, people discussed about new
“maintenance” state for DN to support the scenario where admins need to take the machine/rack
down for 30 minutes for repair.
> We can refactor DN with clear state machine definition based on YARN state related components.
> * State machine consistency between active and standby NN
> Another dimension of state machine management is consistency across NN pairs. We have
dealt with bugs due to different live nodes between active NN and standby NN. Current design
is to have each NN manage its own state based on the events it receives. For example, DNs
will send heartbeat to both NNs; admins will issue decommission commands to both NNs. Alternative
design approach could be to have ZK manage the state.
> Thoughts?



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message