> On March 9, 2016, 8:13 p.m., Aravindan Vijayan wrote:
> > I think Point 1 will not be needed since we execute the ams-env.sh in 
> > ambari-metrics-collector.
> > 
> > # execute ams-env.sh
> > if [[ -f "${COLLECTOR_CONF_DIR}/ams-env.sh" ]]; then
> >   . "${COLLECTOR_CONF_DIR}/ams-env.sh"
> > else
> >   echo "ERROR: Cannot execute ${COLLECTOR_CONF_DIR}/ams-env.sh." 2>&1
> >   exit 1
> > fi
> > 
> > 
> > Point 2 is a bug in UpgradeCatalog and the changes look fine. 
> > 
> > In any case, after https://issues.apache.org/jira/browse/AMBARI-15170, 
> > these properties are not used in the ambari-metrics-collector startup 
> > script.
> 
> Shantanu Mundkur wrote:
>     Aravindan,
>     
>     That is right. We don't really need any setting of the variables as we 
> expect ams-env.sh to be sourced. I had half a mind to just remove the 
> hardcoded lines for these but then that would be assuming that ams-env.sh 
> would definitely set the necessary variables. So besides setting the 
> variables after ams-env as been executed as an extra pre-caution chose to set 
> them to what used to be hardcoded before. This is just in case and these 
> statements are anyway no-ops if the variables were already set.
>     
>     BTW when we set FIFO compaction policy to True we have to be certain that 
> the bundled HBase version supports it. Also, I thought the support for that 
> was still experimental and so thought a default of False might have been 
> preferable but Ambari chooses it as True out of the box. Thoughts?
>     
>     Thanks.
> 
> Aravindan Vijayan wrote:
>     Shantanu,
>     
>     I am fine with the changes to the startup script as well. As I mentioned 
> though, these will not be used. 
>     
>     We control what version of HBase is bundled with AMS in every Ambari 
> version. That way we make sure the compatiblity between the HBase version and 
> FIFO compaction policy. AMS HBase is not part of the stack that Ambari is 
> managing on a cluster.
>     
>     Regarding FIFO compaction policy, we decided to go with a TRUE default 
> value after consulting with a few HBase devs. We are still refining when and 
> how we invoke these HBase config changes.
> 
> Sid Wagle wrote:
>     FIFO compaction gives us significant perf improvement, the normalizer is 
> more of a question mark from AMS point of view since it does merge Region 
> boundaries.
>     We recently did test out both of these features on a 900+ node test 
> cluster and got positive results with only a single node AMS. (HA work is 
> still on the table)

Yes indeed. Performance boosts are much needed as we have a number of reports 
of issues in this context for AMS.

So then as far as the changes are concerned do we only push these changes to 
trunk or should we consider making them in 2.2.2? Also if you do give this a 
Ship It!I'd request one of you to push the changes in for me as I do not have 
committer status. Thanks for the feedback.


- Shantanu


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/44525/#review122790
-----------------------------------------------------------


On March 9, 2016, 8:59 p.m., Shantanu Mundkur wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/44525/
> -----------------------------------------------------------
> 
> (Updated March 9, 2016, 8:59 p.m.)
> 
> 
> Review request for Ambari, Aravindan Vijayan and Sid Wagle.
> 
> 
> Bugs: AMBARI-15331
>     https://issues.apache.org/jira/browse/AMBARI-15331
> 
> 
> Repository: ambari
> 
> 
> Description
> -------
> 
> AMBARI-15331: AMS HBase FIFO compaction policy and Normalizer settings are 
> not handled correctly
> 
> A user could change the defaults for these settings via configuration updates 
> but at least two problems would result.
> 
> 1) Updates to variables AMS_HBASE_FIFO_COMPACTION_ENABLED and 
> AMS_HBASE_NORMALIZER_ENABLED will not be honored. Same would be the case for 
> the recent addition to ams-env - AMS_HBASE_INIT_CHECK_ENABLED.
> 
> These are hardcoded in ambari-metrics-collector rather than using what was 
> propagated via ams-env.sh.To fix we'll try to use the values we expect to be 
> determined by ams-env (which would also account for any user modification to 
> the configuration). To be conservative we'll use some default value (same as 
> it is today) but only if the variable was found not set.
> 
> 2) During upgrade to Ambari 2.2.0 these settings will not be correctly 
> handled due to incorrect variable names and leads to exceptions e.g if 
> AMS_HBASE_FIFO_COMPACTION_ENABLED was expected to be set to 'false', for 
> instance if the AMS HBase version did not support FIFO compaction policy.
> 
> HBASE_FIFO_COMPACTION_POLICY_ENABLED is used instead of 
> AMS_HBASE_FIFO_COMPACTION_ENABLED
> HBASE_NORMALIZATION_ENABLED is used instead of AMS_HBASE_NORMALIZER_ENABLED
> 
> Due to this and due to the hardcoded-values in ambari-metrics-collector 
> undesired behavior (collector not starting up, attempts to startup AMS 
> leaving hanging processes around) can result especially if the underlying AMS 
> HBase version does not support FIFO compaction policy.
> 
> 
> Diffs
> -----
> 
>   
> ambari-metrics/ambari-metrics-timelineservice/conf/unix/ambari-metrics-collector
>  64a7848 
>   
> ambari-server/src/main/java/org/apache/ambari/server/upgrade/UpgradeCatalog220.java
>  ac6b3c5 
>   
> ambari-server/src/main/resources/common-services/AMBARI_METRICS/0.1.0/configuration/ams-env.xml
>  78b8999 
>   
> ambari-server/src/test/java/org/apache/ambari/server/upgrade/UpgradeCatalog220Test.java
>  8263001 
> 
> Diff: https://reviews.apache.org/r/44525/diff/
> 
> 
> Testing
> -------
> 
> 1) Results :
> 
> Tests run: 3930, Failures: 0, Errors: 0, Skipped: 33
> 
> 2) Updated UpgradeCatalog220Test testcase.
> 
> 3) Manually tested:
> 
> i) Set timeline.metrics.hbase.fifo.compaction.enabled to false and ensure it 
> was effective after deployment. AMS components were all    started and 
> running successfully with dashboards and heatmaps being displayed as expected.
> ii) Upgrade from Ambari 2.1.0 to Ambari 2.2.0 and ensured the property 
> timeline.metrics.hbase.fifo.compaction.enabled was false after deployment. 
> AMS components were all started and running successfully with dashboards and 
> heatmaps being displayed as expected.
> iii) Updates to hbase.normalizer.enabled were effective in the configuration 
> for both the new install and upgrade cases.
> 
> 
> File Attachments
> ----------------
> 
> AMBARI-15331.patch
>   
> https://reviews.apache.org/media/uploaded/files/2016/03/08/243b5224-b255-404a-bd24-1b3ce5f79649__AMBARI-15331.patch
> 
> 
> Thanks,
> 
> Shantanu Mundkur
> 
>

Reply via email to