Me and @Aditya Singh were discussing how to fix snapshotting for the failing modules on GHA with some floating point number rounding or somesuch, but I saw that there is some activity on https://github.com/nictru/nft-anndata. Are you already working on implementing this for the pipeline, @Nico Trummer, or do you want some help?

Well, the details about the differences is not only rounding as I found out recently. I like to understand what exactly the reasons for the inconsistencies are, before implementing too lenient tests, which makes this process slow. But I'm working on it. My documentation on what I found for the PCA module is on the most recent PR, can't send the link as I'm on my phone

Ah, that's annoying... Let us know if you want assistance with anything!