> I'll try the simplest query I can reduce it to with loads of memory and see > if that gets anywhere. Other pointers are much appreciated.
Looks like something I'm testing right now (to make the memory setting cost-based). https://issues.apache.org/jira/browse/HIVE-21399 A less "cost-based" solution would be to do set hive.map.aggr=false; And disable the feature. Cheers, Gopal