[FLINK-3479] Add hash-based strategy for CombineFunction

• Type: Sub-task
• Status: Closed
• Priority: Minor
• Resolution: Won't Do
• Component/s: Runtime / Task
• Assignee: Unassigned
• Reporter: Fabian Hueske
• Created: 23/Feb/16 09:34
• Updated: 08/Jul/19 13:51
• Resolved: 08/Jul/19 13:51

Description

This issue is similar to FLINK-3477 but adds a hash-based strategy for CombineFunction instead of ReduceFunction.

The interface of CombineFunction differs from ReduceFunction by providing an Iterable<T> instead of two T values. Hence, if the Iterable<T> provides two values, we can do the same as with a ReduceFunction.

At the moment, CombineFunction is wrapped in a GroupCombineFunction and hence executed using the GroupReduceCombineDriver.
We should add dedicated two dedicated drivers: CombineDriver and ChainedCombineDriver and two driver strategies: HASH_COMBINE and SORT_COMBINE.

If FLINK-3477 is resolved, we can reuse the hash-table.

We should also add compiler hints to `DataSet.reduceGroup()` and `Grouping.reduceGroup()` to allow users to select between a SORT and HASH based combine strategies (HASH will only be applicable to CombineFunction and not GroupCombineFunction).