This JIRA proposes to improve vectorization for ACID by eliminating row-by-row stitching when reading back ACID files. In the current implementation, a vectorized row batch is created by populating the batch one row at a time, before the vectorized batch is passed up along the operator pipeline. This row-by-row stitching limitation was because of the fact that the ACID insert/update/delete events from various delta files needed to be merged together before the actual version of a given row was found out.
HIVE-14035 has enabled us to break away from that limitation by splitting ACID update events into a combination of delete+insert. In fact, it has now enabled us to create splits on delta files.
Building on top of
HIVE-14035, this JIRA proposes to solve this earlier bottleneck in the vectorized code path for ACID by now directly reading row batches from the underlying ORC files and avoiding any stitching altogether. Once a row batch is read from the split (which may be on a base/delta file), the deleted rows will be found by cross-referencing them against a data structure that will just keep track of deleted events (found in the deleted_delta files). This will lead to a large performance gain when reading ACID files in vectorized fashion, while enabling further optimizations in future that can be done on top of that.