Class ReaderDeltaParallel


public class ReaderDeltaParallel extends ReaderDelta
Parallel native Delta Lake matrix reader. Delta tables are stored as one or more parquet data files; this reader decodes those files concurrently (one task per data file) and then concatenates the per-file row-major buffers into the dense output in the original file order.

The expensive part of a Delta read is the parquet decode, which the kernel performs per data file; parallelizing across files is therefore the natural way to bridge the gap to the (near-raw) binary reader. A table backed by a single data file (the default for tables <= the parquet target file size) cannot be split this way, so the reader transparently falls back to the sequential ReaderDelta path in that case.