API: change default in to_datetime for errors to 'raise'

[jreback]

xref #8894
xref #10615
xref #10154

The idea is that if you have completely unparseable dates with or w/o a format specified, then the default will be to raise. Further, since coerce is by default False. You will get an immediate exception rather than have strings returned.

I think some tests will need to be adjusted and a simple example for the user in the whatsnew.

[shoyer]

Agreed, this is a good idea.

[chris-b1]

To summarize latest current behavior (in #10615) - here is what happens to invalid string data - same whether format specified or not. Default is (coerce=False, errors='ignore', format=None)

coerce	errors	Invalid
FALSE	'ignore'	Return obj
FALSE	'raise'	Raise
TRUE	either	Return NaTs

[vincentdavis]

@chris-b1 I was think about the grid you made above. Why not have 1 keyword arg that can be 3 different values, (raise, ignore, NaTs) Would it not simplify the code to not have to search for coerce or error handling but only whatever this new keyword would be. It would be possible to convert coerce and coerce to the new keyword for backward compatibility. The new keyword default would be 'raise'.

Should coerce=True error=raise be rejected?

[jorisvandenbossche]

I agree with @vincentdavis that the two keywords errors and coerce don't need to be two different ones, as they are actually mutually exclusive (which makes it even confusing what you should expect when combining those two, eg the example of coerce=True and error=raise).
I wanted to propose to just integrate coerce into the errors keyword, giving that one then an extra option: 'ignore'|'raise'|'coerce' (this should also be possible in a backwards compatible change).

And then the change of the actual default value is of course another discussion. But I am personally +1 to change it to 'raise'.

Also note that the actual stable behaviour (in the latest release) depended on whether you supplied the format keyword or not. The table above provided by @chris-b1 was valid for the default case, but when providing a date that could not be parsed following a provided format, this would raise regardless of errors='ignore'.
So this is another reason to change the default to raise to unify this behaviour.

[jorisvandenbossche]

As the discussion about coerce is not really linked to the change of the default value to'raise', I opened a separate issue for that: #10653

[jorisvandenbossche]

I also mentioned this discussion on the mailing list, as it backwards incompatible change: https://groups.google.com/forum/?fromgroups#!topic/pydata/Ef_xP7zQOTc

[bwillers]

Having been bitten by the silent failure in to_datetime on a few occassions, would love to see this change.