due to some host's disk are busy, it will results failed in timeoutException while executor try to register to shuffler server on that host...
and then it will exit(1) while launch task on a null executor.
and yarn cluster resource are a little busy, yarn will thought that host is idle, it will prefer to allocate the same host executor, so it will have a chance that one task failed 4 times in the same host.
currently, CoarsedExecutorBackend register to driver first, and after registerDriver successful, then initial Executor.
if exception occurs in Executor initialization,
But Driver don't know that event, will still launch task in that executor,
then will call system.exit(1).
It is more reasonable to register with driver after Executor is ready... and make registerTimeout to be configurable...