ParallelLeafReader has a getTermVectors(docId) implementation that loops over each field it has in a loop and calls getTermVector(docId,fieldName). But the implementation of that will load all term vectors for all fields in that reader, yet ParallelLeafReader only wants one. The effect is an O(n^2) where 'n' is the number of fields, when we could get O(n) if we do it right. PLR should call getTermVectors(docId) (not referring to a specific field) for each of it's readers and then aggregate them.
This wouldn't be such a problem if our term vector API/Codec was improved to not load all term vectors for all fields from disk at once.
Found via randomized-testing of IndexWriter auto-picking ParallelAtomicReader along with a test I have that asserts TVs aren't fetched for a doc more than once.