Details
-
Documentation
-
Status: Closed
-
Minor
-
Resolution: Invalid
-
2.1.1
-
None
-
None
Description
Hi
This is not a bug but maybe a lack of documentation.
I have a job that produce a lot of blockmgr files... I do not understand why the shuffle writes so much on disk and not in the heap of nodemanager.
I wanted to increase spark.shuffle.memoryFraction to reduce the part of data on disk, but this parameter is deprecated in the version we use (https://spark.apache.org/docs/2.1.1/configuration.html)
How to increase the memory size allocated to shuffle in spark 2 ? Is there a non documented parameter ?
I do not use an external shuffle service and I'd prefer to avoid it for now...
Thanks in advance