Github user jerryshao commented on the issue:
https://github.com/apache/spark/pull/14887
@zhaoyunjiong , the fix you made may introduce a situation where recovery
data will be existed in multiple directories, I'm not sure if this will
introduce recovery issue or others, since now the recovery data may not be
consistent.
IMO I think here based on SPARK-14963, we could change to enable Spark's
shuffle service recovery as a configuration:
1. If it is not enabled, then Spark will not persist data into leveldb, in
that case yarn shuffle service can still be served but lose the ability for
recovery.
2. If it is enabled, then user should guarantee recovery path is reliable.
Because recovery path is also crucial for NM to recover.
3. Also this configuration should be consistent with NM's recovery enabled
configuration.
4. If this shuffle service is running on a lower version of Hadoop where
there's no NM recovery
* If Spark's shuffle service recovery is enabled, refer to 2.
* If it is not enabled, then refer to 1.
Just my two cents, may have some missing parts. Basically I think to solve
your problem (also considering recovery) it might be better to make Spark's
shuffle recovery mechanism as configurable.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
