[jira] [Commented] (REEF-594) Add NodeDescriptor, number of cores, and memory to construct complete EvaluatorManager for recovered evaluator

["Andrew Chung (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/REEF-594?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14698028#comment-14698028 ] 

Andrew Chung commented on REEF-594:
-----------------------------------

[~markus.weimer] [~yingdachen] Personally, I opt for keeping the information in the DFS, since logs are less frequently written than evaluator heartbeats and there are thus lower overheads. What do you think? This would require us to parse the evaluator log on recovery with a more complicated object serialization definition though (would comma separated work here?).

> Add NodeDescriptor, number of cores, and memory to construct complete EvaluatorManager for recovered evaluator
> --------------------------------------------------------------------------------------------------------------
>
>                 Key: REEF-594
>                 URL: https://issues.apache.org/jira/browse/REEF-594
>             Project: REEF
>          Issue Type: Sub-task
>          Components: REEF Driver, REEF.NET Driver
>            Reporter: Andrew Chung
>
> Currently, when we recover an evaluator, we cannot provide complete information of the evaluator back to the user because we do not persist the information anywhere. There are a few options here to keep the information. The first is to persist the information in the DFS, the second is to add the information in EvaluatorStatusProto.



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