Chao Sun commented on HDFS-12975:

Thanks [~shv] for taking a look! I've fixed the changes for the 1) and 3), and 
currently is working on a way to let observer participate in failover. There 
are some complexities:

According to the design, during failover we want to choose SBN over Observer if 
the former is available. Therefore, the logic in {{ZKFailoverController}} will 
need to change.

First, for graceful failover, I'm thinking about associating each target with a 
flag indicating whether it is a "preferred" target (e.g., SBN).
 In {{ZKFailoverController::doGracefulFailover()}} we will then check the flag 
of the local target against all the flags of the other nodes, and determine the 
action (e.g., when local target is non-preferred and there exists at least one 
preferred target in other nodes, quit the election).

Second, for the case of "ungraceful" failover, such as ANN crash, it seems more 
difficult. The current mechanism is to haveĀ all the available candidates to 
compete for the ZK lock, so there is no way for them to communicate and 
coordinate such that SBN will be preferred.

I'll continue to look into this. But let me know if you have any good 

> Changes to the NameNode to support reads from standby
> -----------------------------------------------------
>                 Key: HDFS-12975
>                 URL: https://issues.apache.org/jira/browse/HDFS-12975
>             Project: Hadoop HDFS
>          Issue Type: Sub-task
>          Components: namenode
>            Reporter: Konstantin Shvachko
>            Assignee: Chao Sun
>            Priority: Major
>         Attachments: HDFS-12975.000.patch
> In order to support reads from standby NameNode needs changes to add Observer 
> role, turn off checkpointing and such.

This message was sent by Atlassian JIRA

To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org

Reply via email to