Better qdrant params

[generall]

Hi!

Sorry for late updates, just noticed that qdrant is included into the latest run. In general I would say that comparing client-server vector DBs with pure embedded implementations of algorithms might be misleading, but since other DBs are participating, we kinda have to as well.

In this PR I tried to do changes with the following goals:

• Minimize the overhead of data transfer: by default pydantic makes a lot of verification, that consumes CPU time. The PR changes it into raw gRPC data structures which doesn't have any validation
• Set the parameters for optimal RPS
• Try quantization

Hope you could re-run the latest iteration with those changes included.

Expected RPS increase with those changes: 2-3x

[erikbern]

Awesome. Thanks a lot for the contribution!

I think this is a bit too late for me to include in the benchmarks I'm publishing in #399 – I have to draw a line somewhere. But I'm planning to re-run everything in just a few weeks anyway. I don't intend to promote the results much until then.

[generall]

Hi @erikbern, is there any ETA for the next run?