Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
Description
In map side group aggregation, partial grouped aggregation is calculated to reduce the data written to disk by map task. In case of hash aggregation, where the input data is not sorted, hash table is used (with sorting also being performed before flushing). If the hash table size increases beyond configurable limit, data is flushed to disk and new hash table is generated. If the reduction by hash table is less than min hash aggregation reduction calculated during compile time, the map side aggregation is converted to streaming mode. So if the first few batch of records does not result into significant reduction, then the mode is switched to streaming mode. This may have impact on performance, if the subsequent batch of records have less number of distinct values.
To improve performance both in Hash and Streaming mode, a combiner can be added to the map task after the keys are sorted. This will make sure that the aggregation is done if possible and reduce the data written to disk.
