[
https://issues.apache.org/jira/browse/MAPREDUCE-3060?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Luke Lu updated MAPREDUCE-3060:
-------------------------------
Description:
When I was talking to Owen about MAPREDUCE-2600, we came across (again, talked
about it with Chris before) the shuffle dependency issue. NodeManager currently
has an implicit (hidden by the service plugin mechanism) dependency of a
specific version of mapreduce shuffle. While this works in many cases, as long
as we don't change shuffle headers and the usage of mapred security tokens,
it's a hack to make things work none the less. It's generally agreed upon that
nodemanager should only load generic services that are mapreduce framework
neutral.
In this particular case, the right solution seems to be a generic shuffle
handler that can serve data for a particular partition securely. The
ShuffleHandler currently only depends on mapreduce for task tokens and shuffle
header, which is only used for writing data, i.e., the shuffle handler has no
semantic dependency on mapreduce.
was:
When I was talking to Owen about HADOOP-2600, we came across (again, talked
about it with Chris before) the shuffle dependency issue. NodeManager currently
has an implicit (hidden by the service plugin mechanism) dependency of a
specific version of mapreduce shuffle. While this works in many cases, as long
as we don't change shuffle headers and the usage of mapred security tokens,
it's a hack to make things work none the less. It's generally agreed upon that
nodemanager should only load generic services that are mapreduce framework
neutral.
In this particular case, the right solution seems to be a generic shuffle
handler that can serve data for a particular partition securely. The
ShuffleHandler currently only depends on mapreduce for task tokens and shuffle
header, which is only used for writing data, i.e., the shuffle handler has no
semantic dependency on mapreduce.
> Generic shuffle service
> -----------------------
>
> Key: MAPREDUCE-3060
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-3060
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Components: mrv2
> Affects Versions: 0.23.0
> Reporter: Luke Lu
> Labels: shuffle
> Fix For: 0.24.0
>
>
> When I was talking to Owen about MAPREDUCE-2600, we came across (again,
> talked about it with Chris before) the shuffle dependency issue. NodeManager
> currently has an implicit (hidden by the service plugin mechanism) dependency
> of a specific version of mapreduce shuffle. While this works in many cases,
> as long as we don't change shuffle headers and the usage of mapred security
> tokens, it's a hack to make things work none the less. It's generally agreed
> upon that nodemanager should only load generic services that are mapreduce
> framework neutral.
> In this particular case, the right solution seems to be a generic shuffle
> handler that can serve data for a particular partition securely. The
> ShuffleHandler currently only depends on mapreduce for task tokens and
> shuffle header, which is only used for writing data, i.e., the shuffle
> handler has no semantic dependency on mapreduce.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
