Description
- SCENARIO
Currently in carbondata data loading, during sort process step, records will be sorted partially and spilled to the disk. And then carbondata will read these records and do merge sort.
Since sort step is CPU-tense, during writing/reading these records, we can optimize the serialization/deserialization for these rows and reduce CPU consumption in parsing the rows.
This should enhance the data loading performance.
- RESOLVE
We can pick up the un-sorted fields in the row and pack them as bytes array and skip paring them.
- RESULT
I've tested it in my cluster and seen about 8% performance gained (74MB/s/Node -> 81MB/s/Node).
Attachments
Issue Links
- links to