mbrette

What is your take on existing merge optimization #12050 ? The approach seems very effective (*), however it does not work if there are deleted documents, which is likely to...

Indeed, looking at the [TieredMergePolicy](https://lucene.apache.org/core/9_7_0/core/org/apache/lucene/index/TieredMergePolicy.html), it seems that it will priviledge merge with the lowest skew, while we would want the reverse for hnsw merge, or at least have an...

I agree on having a specific merge strategy as you describe. As the graph construction is O(n logn), the relative cost (per node) of merging in a small segment is...

An idea, instead of trying to merge the subgraph, is to do a union of subgraphs: When we merge, we build a disconnected graph which is the union of all...

Having a look at the first paper you shared [On the merge of knn graphs](https://arxiv.org/pdf/1908.00814.pdf): they proposed 2 algorithms, the second one is called Joint-merge, and is exactly what we...

@jmazanec15 What is the current way to measure Lucene knn recall/performance this days ? I tried to reproduced your test from https://github.com/apache/lucene/issues/11354#issuecomment-1239961308, but was not able to (recall = 0.574).