Hi all!
We made some experiments with mmap and summarized in in the article "Minimal RAM you need to serve a million vectors" https://qdrant.tech/articles/memory-consumption/

We found that using htop for checking memory consumption is wrong.
So we did it in a hard way.
As a result, we could serve 1+ million vectors using just 135Mb of RAM!

But spoiler: ||you would need fast disks for that||

exist and benchmark or difference between swap memory and mmap? im using swap memory for my projects deployed in several servers and its work pretty well, with a high latency between servers but the search in a server its made in 10 - 15ms with 5 millions of vectors with 1024 features

I'm trying to evaluate how much filtering in queries can save disk reads in MMAPed collections. In the third experiment (HNSW in MMAP) of your article "Minimal RAM you need to serve a million vectors", would the RPS have been higher if the queries included filters? (that filtered out, say, 80% of the vectors)

i don't think 20% filtering could improve speed🤔 it will still use hnsw, but at the same time would need to check conditions while still doing the graph traversal. something close to 1% could help, but it would also require validation

@Andres Can you say more about your servers specs and RAM/swap usage?