kreczko
commented
3 years ago Summary from first look yesterday:
- there are no 1:1 equivalents for some array functions used (e.g. JaggedArray.fromparents), but there are alternative implementations
- in combination with multi-tree support and lazy loading it might be worth doing #32 for some of the parts. The only state left would be the cache.
Creating a dedicated issue for these breaking changes.
Python versions
The last three minor release versions are 3.9, 3.8, 3.7. Given that packages like Parsl require version >= 3.6, should a similar range apply here? For some reason the rule of 3 comes to mind, so supporting >= 3.7
uproot4
Multitree support still feels a bit odd, but lazy loading (with cache) is doable if you know the trees in advance (we do thanks to the configs). By default all keys are now strings, not bytestrings (yay!) - another reason to move
awkward1
This is the real major change and a risk at the moment. Not all numpy operations are available, which might make work with numexpr challenging. However, Jim is, as always, open to discussions and missing features can be requested or implemented by us. The big advantage of doing this well is access to GPUs!
The other interesting bit here is the new zip/unzip functionality. Essentially you can create a Tree/EventModel for you input data and define behaviors for it, e .g. https://github.com/FAST-HEP/scikit-validate/blob/0.4.0-dev/skvalidate/operations/_awkward.py#L4
@benkrikler if you are happy with the target version, I will start work during spare time. It would be good to merge the multi-tree support soon as well ;).