While working on
LUCENE-6652, I figured out that there were so many test with wrongly implemented TermsToBytesRefAttribute. In addition, the whole concept back from Lucene 4.0 was no longer correct:
- We don't return the hash code anymore; it is calculated by BytesRefHash
- The interface is horrible to use. It tends to reuse the BytesRef instance but the whole thing is not correct.
Instead we should remove the fillBytesRef() method from the interface and let getBytesRef() populate and return the BytesRef. It does not matter if the attribute reuses the BytesRef or returns a new one. It just get consumed like a standard CharTermAttribute. You get a BytesRef and can use it until you call incrementToken().
As the TermsToBytesRefAttribute is marked experimental, I see no reason why we should not change the semantics to be more easy to understand and behave like all other attributes. I will add a note to the backwards incompatible changes in Lucene 5.3.