v6(proposal): use float[] and float[][] to represent vectors instead of boxed Float[] / Float[][]

[bevzzz]

Investigate if it has a significant memory overhead. If yes, make class Vectors use primitive arrays.

[bevzzz]

Btw, the reason we use float and not the more-precise double is because Weaviate would only send vectors with 32-bit precision, see corresponding gRPC schema here.

[yazan-abu-obaideh]

I see the change has already been made - great! I've investigated the difference in memory use using VisualVm. A 10,000,000 size array of type float[] takes up 40,000,000B, which makes perfect sense at 4B per float, and an array of type Float[] and of the same size takes up 4 times as much memory. Thought you'd like to know!

And that's beside the GC overhead too ofc...

[bevzzz]

Thanks for putting the number on this change @yazan-abu-obaideh 👍
We figured there wasn't any benefit to be gained from using Float[] as Weaviate vectors aren't meant to store nulls + the GC overhead from cleaning all those objects you mentioned.