Hi Stéphane, Lucene's segments are made of an immutable "core", on top of which deleted docs and doc-value updates may be applied. These deleted docs and doc-value updates may change on every reopen (`DirectoryReader#open(IndexWriter)`) while the segment's core (inverted index, points, etc.) is guaranteed to be fully immutable. So if we took deleted docs into account in the query cache, cached entries would be invalidated more frequently (until the next reopen that brings more deletions vs. until the next merge that merges this segment away).
One could argue that applying deletes is fine because deletes may only accrue over time. But this is a bit dangerous, as nothing prevents users from running a search on an older point-in-time view of the index after a refresh occurs, and the benefits would likely not be worth the extra trappiness. So we're sticking to the simplest approach: the query cache completely ignores deleted docs, and deleted docs need to be applied on top of the cache, typically by a `BulkScorer`. On Thu, Jul 24, 2025 at 11:28 AM Stéphane Campinas < stephane.campi...@gmail.com> wrote: > Hello, > > In the bitset cache implementations (in LRUQueryCache), I am wondering > why livedocs aren't passed to the `score` methods used to create > bitsets. See [0] for an example. > My understanding would be that passing livedocs would benefit the > creation of the cached bitset: > > - faster search since some docs are deleted > - potentially smaller memory usage, e.g., with the roaring bitset > - potentially smaller cost of the FixedBitset > > I have checked the mailing list archive, as well as the commit history, > but found no explanation. Could you explain the reason why livedocs > isn't passed to the `score` method ? > > Best, > > [0] > https://github.com/apache/lucene/blob/7fe43de9f6678b41d8894258e1f99a9f38e87689/lucene/core/src/java/org/apache/lucene/search/LRUQueryCache.java#L539-L553 > -- > Campinas Stephane > -- Adrien
