We previously optimized cases like this cases on the SubbinSizes helper class. Here is an example where this is not enough:

Bin into 50x50 output bins takes about 2 seconds, even for virtually no events:

sc.bin(a, erase=['detector_id'], edges=[theta_bins, wavelength_bins])

• Is there a way to ensure the optimization works better in this case?
• Note there are about 200 million mappings between bins, so performance hit is not unexpected.
• This issue might become much less relevant once we add multi-threading for reduction operations in transform. I think this should be the first thing to investigate, since it would also fix several other performance bottlenecks.

Note: A faster solution (for few events) is:

sc.bin(d.bins.concatenate('detector_id'), edges=[theta_bins, wavelength_bins])