Hi Experts,
      When I`m reading spark code in version 3.0.0, when external shuffle 
service is enabled:
      ShuffleBlockFetcherIterator -> 
            fetchHostLocalBlocks ( there is some logic, when there is no record 
in cache, then it need to use hostLocalDirManager.getHostLocalDirs to send 
message to external shuffle service) 


      My question is :
      1. As the executors are in the same host, the local dir should be sure to 
be same. Why it does so much job here( do some cache and RPC), just to fetch 
local dirs?
      2. Should there be a config to let user just do the same logic as 
LocalBlocks?


      Thanks a lot.


Best,
Tao

Reply via email to