Virtual Array TODOs

[ikrommyd]

With #1277 merged, coffea does support opening a file using virtual arrays. With scikit-hep/awkward#3475, it finally supports opening a file without needing to know the buffer shapes.
This is an issue just to keep track of the things that are pending to fully support virtual arrays well.

• Bring back the coffea 0.7-like executors.
• Investigate whether coffea should cache the original uncut events in the @original_array attribute because since they are virtual, the memory overhead is only the python objects and not real arrays. It's probably good to keep them there to avoid doing from_buffers when we need the original events.
• What should coffea's persistent and runtime caches do? It looks like one of them is storing callables in the case of virtual arrays and the other does not do anything.
• Can we do something about objects that are really regular arrays like LHEPdfWeights but uproot deserializes them as list offset arrays? That sounds like a useless offsets calculation to me but there may not be another way.

[ikrommyd]

@lgray @pfackeldey please edit if you think I forgot something.

[ikrommyd]

After having a chat with Peter, I rearranged them in order of priority in my opinion. The last thing:

Investigate the usage of coffea's caches to store deserialized offsets branches when opening up the file to avoid deserializing the same offsets more than one times (some objects might share offsets).

we're probably not gonna do because everything else should have solved the problem so there is no reason to add an extra solution to the same problem.

[ikrommyd]

Well scikit-hep/awkward#3475 reduces this issue by a lot so I edited it accordingly 😄