> On Oct. 28, 2015, 4:52 p.m., Alejandro Fernandez wrote:
> > ambari-server/src/main/resources/common-services/HDFS/2.1.0.2.0/package/scripts/zkfc_slave.py,
> > line 72
> > <https://reviews.apache.org/r/39731/diff/1/?file=1111687#file1111687line72>
> >
> > I thought we started ZKFC using the same component in hdp-select as
> > namenode.
> >
> > Otherwise, it may be risky to change hadoop-client this early on during
> > RU.
I think ZKFC never upgrades in a rolling upgrade. Take a look at the workflow
in RU:
<group name="CORE_MASTER" title="Core Masters">
<service-check>false</service-check>
<service name="HDFS">
<component>JOURNALNODE</component>
<component>ZKFC</component>
<component>NAMENODE</component>
</service>
It's after JN, but before NN. Since ZKFC wasn't bound to JN in our mappings, it
must have always started on the old version. I can change this, but if I change
it to NameNode, that could be a problem still.
- Jonathan
-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/39731/#review104345
-----------------------------------------------------------
On Oct. 28, 2015, 3:31 p.m., Jonathan Hurley wrote:
>
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/39731/
> -----------------------------------------------------------
>
> (Updated Oct. 28, 2015, 3:31 p.m.)
>
>
> Review request for Ambari, Alejandro Fernandez, Jayush Luniya, and Nate Cole.
>
>
> Bugs: AMBARI-13615
> https://issues.apache.org/jira/browse/AMBARI-13615
>
>
> Repository: ambari
>
>
> Description
> -------
>
> During an express upgrade, components are stopped ahead of time. Before
> {{restart}} is invoked, the following task runs updating all hdp pointers:
>
> {code}
> <group xsi:type="cluster" name="RESTORE_CONFIG_DIRS" title="Restore
> Configuration Directories">
> <direction>DOWNGRADE</direction>
> <execute-stage title="Restore configuration directories and remove HDP
> 2.3 symlinks">
> <task xsi:type="execute">
> <script>scripts/ru_set_all.py</script>
> <function>unlink_all_configs</function>
> </task>
> </execute-stage>
> </group>
> {code}
>
> After this, all components begin to restart. However, restarting involves a
> {{stop}} and a {{start}} command. The components are already stopped and most
> of them have logic that says if the PID says it's not running, then don't
> stop it twice.
>
> However, some components like ZKFC and HBase Master don't have this logic and
> try stopping it regardless. The problem arises when a JVM is spun up to stop
> the process:
>
> Initially it was though that moving the {{hdp-select set all}} to after the
> {{restart}} groups would solve the problem. As it turns out, moving the
> {{hdp-select set all}} doesn't work since the {{params.py}} it always taking
> the new version and building conf/lib/bin directories with it. Additionally,
> some components have upgrade bugs which calling {{hdp-select set all}}
> corrects.
>
>
> Diffs
> -----
>
>
> ambari-common/src/main/python/resource_management/libraries/functions/hdp_select.py
> 5f05777
>
> ambari-server/src/main/resources/common-services/HDFS/2.1.0.2.0/package/scripts/utils.py
> 97ad424
>
> ambari-server/src/main/resources/common-services/HDFS/2.1.0.2.0/package/scripts/zkfc_slave.py
> e9037d8
>
> ambari-server/src/main/resources/stacks/HDP/2.1/upgrades/nonrolling-upgrade-2.3.xml
> 25fd6ab
>
> ambari-server/src/main/resources/stacks/HDP/2.2/upgrades/nonrolling-upgrade-2.2.xml
> 63f9f8d
>
> ambari-server/src/main/resources/stacks/HDP/2.2/upgrades/nonrolling-upgrade-2.3.xml
> 44413d3
>
> ambari-server/src/main/resources/stacks/HDP/2.3/upgrades/nonrolling-upgrade-2.3.xml
> 407b22b
> ambari-server/src/test/python/stacks/2.1/HIVE/test_hive_metastore.py
> 2800224
> ambari-server/src/test/python/stacks/utils/RMFTestCase.py ab4eed4
>
> Diff: https://reviews.apache.org/r/39731/diff/
>
>
> Testing
> -------
>
> Upgrades from 2.2 to 2.3, and 2.3 to 2.3+ with ZKFC.
>
> mvn clean test
>
>
> Thanks,
>
> Jonathan Hurley
>
>
