By default, velox enabled lazy read. Only the columns used by filter are decoded in table scan operator. The other columnes are saved as lazy vector. Data Fetch, decompression and decoder is triggered only when these columns are flatten. So part of the table scan time is actually counted into next operator.

While in WholeStageResultIterator::nextInternal function, we flatten the vector before it's returned to Spark which trigger the lazy read operator. The time is missing in our tracker.

The PR add metrics "time of loading lazy vectors" in NextInternal function to the last operator of the transformer.