CLN: blocks, not series, should contain logic for arithmetic

[shoyer]

See discussion here: #9848 (comment)

This would likely help clean up some of the Series/DataFrame discrepancies in arithmetic operations (mostly for datetime types).

[max-sixty]

Would this also help performance materially, if DataFrame ops could iterate over Blocks rather than Columns?
ref #11623, #11618

[jreback]

@MaximilianR most things are already dispatched already by blocks.

[max-sixty]

Ah, fair, cheers. The ones that are applied across columns (which happen to have been the ones I've been using recently), seem to only be quantile & round.
(There are others on the DataFrame object such as rank, but don't iterate over columns in python)

[jreback]

yeh round is almost a display thing so doesn't really count (and could be daily broadcast)

quantile is bad impl ATM

[max-sixty]

I could do a PR for quantile if it's not too complex - what should that look like? Could it go straight to np.nanpercentile [link]?
Looking at the current code, it's doing a lot though.

[jbrockmendel]

@jreback closeable? We’ve recently finished getting all arithmetic out of blocks.