PERF: much faster to_datetime performance with a format of '%Y%m%d'

[jreback]

In [1]: rng = date_range('1/1/2000', periods=10000, freq='D')

In [2]: strings = Series(rng.year*10000+rng.month*100+rng.day,dtype=np.int64).apply(str)

In [3]: %timeit pandas.tslib.array_strptime(strings.values,"%Y%m%d")
10 loops, best of 3: 42.9 ms per loop

In [4]: %timeit pd.to_datetime(strings,format="%Y%m%d")
100 loops, best of 3: 9.21 ms per loop

In [5]: (pd.to_datetime(strings,format="%Y%m%d") == pandas.tslib.array_strptime(strings.values,"%Y%m%d")).all()
Out[5]: True

[ghost]

Does this apply to read_csv as well? I don't think it calls into to_datetime, does it?
we should get our fast paths consistent everywhere

[jreback]

I don't believe so, this should be wrapped up in tslib.array_to_datetime / tslib.try_parse_dates

or prob better is to have read_csv use to_datetime (or a related call) directly; the conversions are done after parsing anyhow (so all python and not cython code)

so maybe fix to_datetime (in #5490) and then have another issue to wrap back into read_csv...

[ghost]

Sounds good, open an issue for 0.14?

Added note: 433ddc3
... and fixed the typos: 38e3b90. sigh

[jreback]

yep