You are viewing a plain text version of this content. The canonical link for it is here.
Posted to common-dev@hadoop.apache.org by "Amar Kamat (JIRA)" <ji...@apache.org> on 2009/06/18 10:20:07 UTC
[jira] Created: (HADOOP-6078) Ability to re-configure hadoop
daemons online
Ability to re-configure hadoop daemons online
---------------------------------------------
Key: HADOOP-6078
URL: https://issues.apache.org/jira/browse/HADOOP-6078
Project: Hadoop Core
Issue Type: New Feature
Components: mapred
Reporter: Amar Kamat
Example :
Like we have _bin hadoop mradmin -refreshNodes_ we should also have _bin hadoop mradmin -reconfigure_ which re-configures mr while the cluster is online. Few parameters like job-expiry-interval etc can be changed in this way without having to restart the whole cluster.
Master, once reconfigured, can ask the slaves to reconfigure (reload its config) from a well defined location on hdfs or via heartbeat.
We can have some whitelisted configs that have _reloadable_ property.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
[jira] Commented: (HADOOP-6078) Ability to re-configure hadoop
daemons online
Posted by "Steve Loughran (JIRA)" <ji...@apache.org>.
[ https://issues.apache.org/jira/browse/HADOOP-6078?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12721776#action_12721776 ]
Steve Loughran commented on HADOOP-6078:
----------------------------------------
* You could do much of this just by taking down the specific nodes and bringing them up again.
* Hot reconfig is tricky, for as todd points out, some values are cached everywhere.
* Yet I'd like my TTs to be able to rebind to a JT that has just been brought up on a different address.
At the very least, all the loops where workers spin waiting for their masters should reread their configuration values every iteration. That way anything that has subclassed Configuration to give live data could provide updated locations
FWIW, although I've subclassed JobConf for my configuration, I stopped trying to do live updates, as the objects end up being serialized and reread. I think it would be really hard to do live reconfiguration with the current design. Node restart is cleaner.
> Ability to re-configure hadoop daemons online
> ---------------------------------------------
>
> Key: HADOOP-6078
> URL: https://issues.apache.org/jira/browse/HADOOP-6078
> Project: Hadoop Core
> Issue Type: New Feature
> Components: mapred
> Reporter: Amar Kamat
>
> Example :
> Like we have _bin hadoop mradmin -refreshNodes_ we should also have _bin hadoop mradmin -reconfigure_ which re-configures mr while the cluster is online. Few parameters like job-expiry-interval etc can be changed in this way without having to restart the whole cluster.
> Master, once reconfigured, can ask the slaves to reconfigure (reload its config) from a well defined location on hdfs or via heartbeat.
> We can have some whitelisted configs that have _reloadable_ property.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
[jira] Commented: (HADOOP-6078) Ability to re-configure hadoop
daemons online
Posted by "Todd Lipcon (JIRA)" <ji...@apache.org>.
[ https://issues.apache.org/jira/browse/HADOOP-6078?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12721421#action_12721421 ]
Todd Lipcon commented on HADOOP-6078:
-------------------------------------
I like this idea. Here are a couple thoughts:
- In addition to having a whitelist of reloadable properties, we should generate loud warnings about any config properties that are modified but are *not* whitelisted. Something like:
ERROR: configuration property fs.default.name modified on disk but not is not reloadable. Retaining old value.
These errors would ideally spit out on the console of the operator, not just in the log4j output on the server side.
- We should also probably have diagnostic output for any changes that did take effect. For example:
INFO: Configuration property foo.bar changed from "blah" to "blahblah"
- Does this introduce the need for additional synchronization inside Configuration?
> Ability to re-configure hadoop daemons online
> ---------------------------------------------
>
> Key: HADOOP-6078
> URL: https://issues.apache.org/jira/browse/HADOOP-6078
> Project: Hadoop Core
> Issue Type: New Feature
> Components: mapred
> Reporter: Amar Kamat
>
> Example :
> Like we have _bin hadoop mradmin -refreshNodes_ we should also have _bin hadoop mradmin -reconfigure_ which re-configures mr while the cluster is online. Few parameters like job-expiry-interval etc can be changed in this way without having to restart the whole cluster.
> Master, once reconfigured, can ask the slaves to reconfigure (reload its config) from a well defined location on hdfs or via heartbeat.
> We can have some whitelisted configs that have _reloadable_ property.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.