Andrew Kane
Andrew Kane
Hi @pashkinelfe, closing this out, but happy to revisit if the cause is identified.
Hi @pashkinelfe, thanks for the PR, and sorry for the delay. The build improvement looks nice. My main concern (besides maintenance of more distance functions) is accumulation happens at half...
Cleaning up stale PRs (but happy to revisit if there's a way to significantly improve performance without changing the accumulation precision).
Hi @mschmidt23, which instance type are you using, and what are the values of `maintenance_work_mem`, `max_parallel_maintenance_workers`, and `shared_buffers`?
My initial thinking is F16C + FMA dispatching might not be happening. Another contributing factor could be using less parallel workers due to a smaller table size (possibly one less...
@mschmidt23 Sorry for the delay. I just tried indexing 1M 384-dimension random vectors with 4 processes (`ivfflat`, 1000 lists, cosine distance) on a r5.xlarge EC2 instance and got the following...
@mschmidt23 If you're still seeing this issue, I'd recommend following up with AWS.
@amelboc It sounds like you're trying to create an invalid vector. If there's a value at index 102, the vector can't have less than 102 dimensions. For instance, `{3:99}/3` represents...
Hi @Hornobster, can you paste the full stack trace? (with any personal info removed)
Thanks. I don't really see a way to fix this without a breaking change (having vector columns return a `Vector` object instead of an `nd.array`), so will consider this for...