Skip to content

Conversation

@kaivalnp
Copy link
Contributor

@kaivalnp kaivalnp commented Aug 28, 2025

Description

@mikemccand pointed me to this bump in indexing throughput for "~1 KB Wikipedia English docs, with KNN Scalar Quantized Vectors" -- see bump after JB from ~32 GB/hr to ~37 GB/hr (~15%):

Screenshot 2025-08-28 at 11 14 21 AM

Corresponding run: https://benchmarks.mikemccandless.com/2025.07.29.18.04.08.html
The only commit was apache/lucene#14874

This speedup seems plausible, because the same scoring function is used to score quantized float vectors (we initially thought the issue to be scoped to non-quantized byte vectors)

Copy link
Owner

@mikemccand mikemccand left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yay, thanks @kaivalnp! Go benchmarks! (Uncovering an impact, positive thank god, we didn't know about).

@mikemccand mikemccand merged commit 4a9221e into mikemccand:main Aug 28, 2025
@kaivalnp kaivalnp deleted the 14874 branch August 28, 2025 16:08
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

2 participants