BUG: pd.frequency.get_offset cache may be overwritten

[sinhrks]

Because frequencies.get_offset may be broken by user operation because of its cache.

https://github.com/pydata/pandas/blob/master/pandas/tseries/frequencies.py#L515

import pandas as pd
x = pd.tseries.frequencies.get_offset('D')
x.__dict__
# {'normalize': False, 'kwds': {}, 'n': 1, '_offset': datetime.timedelta(1), '_named': 'D', '_use_relativedelta': False}

# modify its property
x.n = 5

# NG, modified instance is returned
y = pd.tseries.frequencies.get_offset('D')
y.__dict__
# {'normalize': False, 'kwds': {}, 'n': 5, '_offset': datetime.timedelta(1), '_named': 'D', '_use_relativedelta': False}

It should return a copy of the cache.

[jreback]

yeh these are class properties, but get_offset should prob return a copy. Though need to check for perf impacts on this.

[sinhrks]

n is an instance property specify its length/span. I don't think user touch it actually though...

From what I've seen, there is no function uses freq taken from get_offset as it is. Applying numeric ops (like multiplying freq mult) will results in copy. Thus, user ops should not affect to internal cache.

pd.tseries.frequencies.get_offset('A')
# A-DEC
pd.tseries.frequencies.get_offset('A') * 1
# <YearEnd: month=12>

Above repr difference is caused by _named property, which looks to be attached by _make_offset. It is introduced in #5189 for repr backcompat. I don't think it is needed anymore (because user can't get freq from get_offset).

Maybe we can cleanup freq-retrieving functions in freqnencies.py and guarantee user can't touch the internal cache.