ambari-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Robert Nettleton (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (AMBARI-11850) Blueprint export process is very slow in multi-node clusters
Date Wed, 10 Jun 2015 19:46:00 GMT

     [ https://issues.apache.org/jira/browse/AMBARI-11850?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Robert Nettleton updated AMBARI-11850:
--------------------------------------
    Description: 
The Blueprint export process (using the REST API to create a Blueprint based on a running
cluster) has been found to have some performance issues in multi-node clusters. 

Steps to reproduce:

1. Setup a 3-node cluster. 
2. Using the Ambari UI, start a cluster that includes HDFS, Yarn, MapReduce, Zookeeper, and
Ambari Metrics.  Accept all the defaults in configuration and host assignment. 
3. When the cluster deployment succeeds, attempt to export a Blueprint with the following
REST API URL:

http://host:port/api/v1/clusters/clusterone?format=blueprint

The Blueprint export will eventually succeed, but takes quite a while to complete.  With some
basic testing, I've determined that for a 3-node cluster, the Blueprint export process can
take between 7-16 seconds on average.  I've attempted the export multiple times to make sure
no initial startup issues were involved.

The workaround for this issue is to turn off the Ambari Metrics services prior to exporting
the Blueprint.  When Metrics is turned off, the export is much faster.  

The Blueprint export process needs to be updated, such that the Metrics and Alerting services
are not called by the REST framework during a Blueprint export.  These services add quite
a bit to the overall time it takes to export the Blueprint, but without any real benefit,
since the Blueprint does not make use of these services' data.  

I'm working on a fix for this, and will be submitting a patch shortly.  


  was:
The Blueprint export process (using the REST API to create a Blueprint based on a running
cluster) has been found to have some performance issues in multi-node clusters. 

Steps to reproduce:

1. Setup a 3-node cluster. 
2. Using the Ambari UI, start a cluster that includes HDFS, Yarn, MapReduce, Zookeeper, and
Ambari Metrics.  Accept all the defaults in configuration and host assignment. 
3. When the cluster deployment succeeds, attempt to export a Blueprint with the following
REST API URL:

http://host:port/api/v1/clusters/clusterone?format=blueprint

The Blueprint export will eventually succeed, but takes quite a while to complete.  With some
basic testing, I've determined that for a 3-node cluster, the Blueprint export process can
take between 7-16 seconds on average.  I've attempted the export multiple times to make sure
no initial startup issues were involved.

The workaround for this issue is to turn off the Ambari Metrics services prior to exporting
the Blueprint.  When Metrics is turned off, the export is much faster.  

The Blueprint export process needs to be updated, such that the Metrics and Alerting services
are not called by the REST framework during a Blueprint export.  These services add quite
a bit to the overall time it takes to export the Blueprint, but without any real benefit,
since the Blueprint does not make use of these servcies' data.  

I'm working on a fix for this, and will be submitting a patch shortly.  



> Blueprint export process is very slow in multi-node clusters
> ------------------------------------------------------------
>
>                 Key: AMBARI-11850
>                 URL: https://issues.apache.org/jira/browse/AMBARI-11850
>             Project: Ambari
>          Issue Type: Bug
>          Components: ambari-server
>    Affects Versions: 2.0.0
>            Reporter: Robert Nettleton
>            Assignee: Robert Nettleton
>            Priority: Critical
>             Fix For: 2.1.0
>
>   Original Estimate: 48h
>  Remaining Estimate: 48h
>
> The Blueprint export process (using the REST API to create a Blueprint based on a running
cluster) has been found to have some performance issues in multi-node clusters. 
> Steps to reproduce:
> 1. Setup a 3-node cluster. 
> 2. Using the Ambari UI, start a cluster that includes HDFS, Yarn, MapReduce, Zookeeper,
and Ambari Metrics.  Accept all the defaults in configuration and host assignment. 
> 3. When the cluster deployment succeeds, attempt to export a Blueprint with the following
REST API URL:
> http://host:port/api/v1/clusters/clusterone?format=blueprint
> The Blueprint export will eventually succeed, but takes quite a while to complete.  With
some basic testing, I've determined that for a 3-node cluster, the Blueprint export process
can take between 7-16 seconds on average.  I've attempted the export multiple times to make
sure no initial startup issues were involved.
> The workaround for this issue is to turn off the Ambari Metrics services prior to exporting
the Blueprint.  When Metrics is turned off, the export is much faster.  
> The Blueprint export process needs to be updated, such that the Metrics and Alerting
services are not called by the REST framework during a Blueprint export.  These services add
quite a bit to the overall time it takes to export the Blueprint, but without any real benefit,
since the Blueprint does not make use of these services' data.  
> I'm working on a fix for this, and will be submitting a patch shortly.  



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

Mime
View raw message