viirya commented on pull request #30812:
URL: https://github.com/apache/spark/pull/30812#issuecomment-750781778


   > I'm wondering whether this would solve the issue. There are so many 
factors involved. I feel the initial even task distribution (assuming the 
executors are free so the task scheduler will respect the preferred locations) 
would become uneven quickly after some micro batches, caused by, such as, 
uneven partitions, executor lost, concurrent queries, etc... Did you verify 
that this would make a real workload that didn't work before become working?
   
   During running stateful streaming queries recently, it caused some troubles 
by bad initial locations of stores. I agree that this is not ideal, but it 
solves the problem I saw during I tested SS recently. It is simple and 
shouldn't have bad impact/regression to SS queries. Combined with task locality 
configuration, it makes SS queries more stable in my local test.
   
   > If the issue is about memory, a low memory state store implementation such 
as https://github.com/qubole/spark-state-store is a better solution.
   
   The problem is, we have a built-in in memory store. And, I don't think 
Rocksdb-based state store is the answer to all cases. Even with Rocksdb-based 
store, is it good to have skew stores on few executors? Then local disk space 
might be the next issue.
    
   
   


----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
[email protected]



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

Reply via email to