ambari-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jonathan Hurley" <>
Subject Re: Review Request 36231: Revert, The Default hdfs-site.xml Should Have Client Retry Logic Enabled For Rolling Upgrade
Date Tue, 07 Jul 2015 15:29:15 GMT

> On July 6, 2015, 8:22 p.m., Jonathan Hurley wrote:
> > Won't reverting this patch cause problems with the long running tests?
> Alejandro Fernandez wrote:
>     Basically, HDFS has two layers of retries, and only of them should really be used
during HA. When one NameNode is up, setting dfs.client.retry.policy.enabled to false allows
it to quickly find the active (in the case of HA and when a client tried to connect to the
dead NameNode), or retry for up to 3 mins in non-HA.
>     Comment from Jing,
>     "dfs.client.retry.policy.enabled" is only used for client connecting to NameNode
so the change will not affect other components. In the meanwhile, after setting the conf property
back to false, the retry is controlled only by FailoverOnNetworkExceptionRetry, which can
retry up to 10 times and failover 15 times by default. In case that both NameNodes are down
because only one retry policy takes effect the total retry time will be much less compared
with the case we set "dfs.client.retry.policy.enabled" to true. If both NameNodes are in the
standby state the total retry time will not be affected.

>From the 2.2 to 2.3 runbook, section entitled "Cluster Prerequisites":

• Enable client retry properties for HDFS, Hive, and Oozie. These properties are not included
by default, so you might need to add them to the site files.

• For HDFS, set dfs.client.retry.policy.enabled to true in hdfs-site.xml on all nodes with
HDFS services.

So, is the runbook wrong?

- Jonathan

This is an automatically generated e-mail. To reply, visit:

On July 6, 2015, 7:58 p.m., Alejandro Fernandez wrote:
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> -----------------------------------------------------------
> (Updated July 6, 2015, 7:58 p.m.)
> Review request for Ambari, Jonathan Hurley and Nate Cole.
> Repository: ambari
> Description
> -------
> In the case of an HA cluster where the former primary NN was killed "dirty", by catastrophic
power-down or equivalent, and the cluster has successfully failed over to the other NN, a
client that first attempts to contact the dead NN takes 10 minutes to switch to the other
> In Ambari 2.0 and HDP 2.2, dfs.client.retry.policy.enabled was not set at all.
> Recently, in Ambari 2.1 for HDP 2.3, it was defaulted to true as part of AMBARI-11192.
> However, this causes problems during RU
> In an HA setup, our retry actually should be handled by RetryInvocationHandler using
retry policy FailoverOnNetworkExceptionRetry. The client first translates the nameservice
ID into two host names, and creates an individual RPC proxy for each NameNode accordingly.
Each individual NameNode proxy still uses MultipleLinearRandomRetry as its local retry policy,
but because we usually set dfs.client.retry.policy.enabled to false, thus this internal retry
is actually disabled. Then in case we hit any connection issue or remote exception (including
StandbyException), the exception is caught by RetryInvocationHandler and handled according
to FailoverOnNetworkExceptionRetry. In this way the client can failover to the other namenode
immediately instead of keeping retrying the same NameNode.
> However, here because we set dfs.client.retry.policy.enabled to true, the MultipleLinearRandomRetry
is triggered inside of the internal NameNode proxy thus we have to wait 10+ min. The exception
is finally thrown to RetryInvocationHandler until all the retries of MultipleLinearRandomRetry
> Diffs
> -----
>   ambari-server/src/main/java/org/apache/ambari/server/checks/ 5e029f4

>   ambari-server/src/main/java/org/apache/ambari/server/checks/
>   ambari-server/src/main/resources/stacks/HDP/2.2/services/HDFS/configuration/hdfs-site.xml
>   ambari-server/src/test/java/org/apache/ambari/server/checks/
> Diff:
> Testing
> -------
> Unit tests passed,
> ----------------------------------------------------------------------
> Total run:761
> Total errors:0
> Total failures:0
> OK
> I deployed my changes to a brand new cluster and it correctly set the hdfs-site property
dfs.client.retry.policy.enabled to false.
> Thanks,
> Alejandro Fernandez

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message