[ 
https://issues.apache.org/jira/browse/SPARK-27603?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Chenzhao Guo updated SPARK-27603:
---------------------------------
    Description: 
ShuffleClient resides in BlockManager, it's the client to read other executors' 
shuffle files.

ShuffleManager is pluggable, however, ShuffleClient is restrained to be either 
a `NettyBlockTransferService` or an 
`ExternalShuffleClient`(&`MesosExternalShuffleClient`).

When a custom ShuffleManager is used, the current 3 implementations of 
ShuffleClient may not be enough. For example, a shuffle manager leveraging 
remote storage may not need either of these 3 ShuffleClients. It may require 
the network transferring shuffle data index(the ranges of blocks) from the 
executor who wrote them, and then read directly from the globally-accessible 
remote storage, instead of transfering data through network as the 3 
implementations do. Also, there is need for non-Netty based block transfer 
service.

 

 

  was:
ShuffleClient resides in BlockManager, it's the client to read other executors' 
shuffle files.

ShuffleManager is pluggable, however, ShuffleClient is restrained to be either 
a `NettyBlockTransferService` or an `ExternalShuffleClient`.

When a custom ShuffleManager is used, the current 2 implementations of 
ShuffleClient may not be enough. For example, a shuffle manager leveraging 
remote storage may not need either of these 2 ShuffleClient. It may require the 
network transferring shuffle data index(the ranges of blocks) from the executor 
who wrote them, and then read directly from the globally-accessible remote 
storage, instead of transfering data through network as the 2 implementations 
do. Also, there is need for non-Netty based block transfer service.

 

 


> Make ShuffleClient pluggable
> ----------------------------
>
>                 Key: SPARK-27603
>                 URL: https://issues.apache.org/jira/browse/SPARK-27603
>             Project: Spark
>          Issue Type: Improvement
>          Components: Spark Core
>    Affects Versions: 2.4.0
>            Reporter: Chenzhao Guo
>            Priority: Major
>
> ShuffleClient resides in BlockManager, it's the client to read other 
> executors' shuffle files.
> ShuffleManager is pluggable, however, ShuffleClient is restrained to be 
> either a `NettyBlockTransferService` or an 
> `ExternalShuffleClient`(&`MesosExternalShuffleClient`).
> When a custom ShuffleManager is used, the current 3 implementations of 
> ShuffleClient may not be enough. For example, a shuffle manager leveraging 
> remote storage may not need either of these 3 ShuffleClients. It may require 
> the network transferring shuffle data index(the ranges of blocks) from the 
> executor who wrote them, and then read directly from the globally-accessible 
> remote storage, instead of transfering data through network as the 3 
> implementations do. Also, there is need for non-Netty based block transfer 
> service.
>  
>  



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to