Current implementation of hash group by (GrouperFastImpl class) converts input ExecBatches to row-oriented format,
then hashes and compares rows as if they were a single column.

It is more efficient to avoid relatively costly encoding and instead compute hashes of individual columns in column-oriented format mixing them together, and similarly comparing column-oriented data to row-oriented data in the hash table without converting.

Encoding only happens for a subset of input rows that are inserted into the hash table - they introduce new groups.
Keys in hash table remain stored as row-oriented.

There are 3 components that changed:

• key_compare.cc - contains implementation of one column at a time comparison between column-oriented and row-oriented data
• key_hash.cc - implements hashing individual columns of column-oriented key data and mixing the individual results together for a final multi-column hash
• key_encode.cc - implements encoding in row-format only a selection of input rows; previously all input rows were encoded without support for selective encoding

Performance before and after:
Here are selected results (in millions of rows per second) of arrow-compute-aggregate-benchmark for 0.01% nulls in input columns (Ubuntu VM, Clang 10.0, Intel I9-10980XE):