Paul Masurel

``` thread 'blocking-5' panicked at /Users/fulmicoton/.cargo/git/checkouts/tantivy-f70b7ea03dadae9a/b960e40/sstable/src/lib.rs:257:9: Keys should be increasing. ([111, 95, 112, 97, 114, 97, 109, 115, 1, 48, 34, 62, 60, 115, 99, 114, 105, 112, 116, 62,...

(Almost) minimum reproducible example on quickwit 2 documents: ```json {"\u0000B":"1"} {" A":"1"} ``` Index config: ```json { "version": "0.7", "index_id": "airmail", "indexing_settings": { "commit_timeout_secs": 30 }, "doc_mapping": { "mode": "dynamic",...

At the threshold we would have an average of 20 elements per chimera block. Maybe linear search should be considered too? Adding codecs could have some weird hidden switch-dispatch cost....

we cannot fix like this on serialization, because it can yield collision on serializaiton. We need to add unit test on serialization, and normalize upstream.

See https://github.com/quickwit-oss/tantivy/pull/2445

@LebranceBW I agree with both points. The `BooleanQuery -> BooleanWeight -> BooleanScorer` are precisely here so that while users define their query by instanting a `Query` object, this query can...

This sounds like a good idea. Did you have a look at how lucene does it too?

@PSeitz We'd also need to make sure the query language handles it though (filters should not impact scoring). It might be a pain. I'd go with @jpountz solution for simplicity.

> Also, it appears that the master branch is at version 0.8.3, but crates.io has 0.8.4. Is there some other repo somewhere > I've missed? That's actually a mystery. I...

(I put the comment on a different issue by mistake) I suspect the metric is only plugged in for ingest v2 and @fredsig is using ingest v1