I've spent like a week trying to tune this query as a background task and it's driving me a little nuts I think.
Cobbling together pipelines on pipelines on conditional pipelines of distinct ranking models, fusing them together, and trying to do it on a CPU is so slow.
It's already some 5x or so better than prod, but there's a handful of corner cases that I know I should be able to treat.
I've spent like a week trying to tune this query as a background task and it's driving me a little nuts I think.
Cobbling together pipelines on pipelines on conditional pipelines of distinct ranking models, fusing them together, and trying to do it on a CPU is so slow.
It's already some 5x or so better than prod, but there's a handful of corner cases that I know I should be able to treat.