BUG: to_timedelta still raises with errors='ignore'

[jorisvandenbossche]

From the added docs in #13425:

In [10]: pd.to_timedelta(['apple', '1 day'], errors='ignore')
...
ValueError: unit abbreviation w/o a number

In [12]: pd.to_timedelta(['apple', pd.Timedelta('1day')], errors='ignore')
...
ValueError: unit abbreviation w/o a number

It's the same error as you get with errors='raise' (so not captured somewhere), errors='coerce is working correctly.

[gfyoung]

The reason is because errors='ignore', is well, ignored. I think I would just wrap the current try-except logic as follows:

try:
    try:
        for i in range(n):
            result[i] = parse_timedelta_string(values[i])  # always raise
    except Exception:
        for i in range(n):
            result[i] = convert_to_timedelta64(values[i], unit, is_coerce)
except Exception:  # when is_coerce=False
    if errors == 'raise':
        raise
   elif errors == 'ignore':
        return values
return iresult

The reason is that one could consider the original try-except block as the process through which we convert to timedelta, then all three options for errors are satisfied:

1. errors='raise'

Try to convert via method 1, which fails, so we try with method 2. If that fails, we raise the exception.

2. errors='coerce' --> is_coerce=True

Try to convert via method 1, which fails, so we try with method 2. In method 2, if converting one of the elements fails, we return NPY_NAT.

3. errors='ignore'

Try to convert via method 1, which fails, so we try with method 2. If that fails, we just return the original values array as promised in documentation.

[jreback]

yeah this prob needs some logic like that

this is similar to how datetime parsing works