> Vector search, on the other hand, requires reading a lot of small vectors, which might create a large overhead. It is especially noticeable if we use binary quantization, where the size of even large OpenAI 1536d vectors is compressed down to 192 bytes. Dataset size: 2M 768d vectors (~6Gb Raw data), binary quantization, 650Mb of RAM limit. All benchmarks are made with minimal RAM allocation to demonstrate disk cache efficiency.
Nonsense.
There is no performance benefit that immutable structures offer that cannot be had by mutable ones. It is asinine to assert otherwise, especially when literally every single measurement ever done demonstrates that immutable structures perform orders of magnitude slower.
You do not need immutability to create a slow moving cache. Immutability is not where any semblance of performance is coming from here.
If they didnt succumb to idiotic bullshit nonsense, they wouldn’t have even needed this post. If you see any of your senior architects reading medium, fire them immediately.
Immutable structure can, in a single use case, meet mutable ones for performance and that single case is:
-read speed on aligned, flattened data
But, that ignores that fact that getting to a point of aligned, flattened data with immutable structures is incredibly slow.
