pytz localize vs datetime replace

I'm having some weird issues with pytz's .localize() function. Sometimes it wouldn't make adjustments to the localized datetime:

.localize behaviour:

>>> tz
<DstTzInfo 'Africa/Abidjan' LMT-1 day, 23:44:00 STD> 
>>> d
datetime.datetime(2009, 9, 2, 14, 45, 42, 91421)

>>> tz.localize(d)
datetime.datetime(2009, 9, 2, 14, 45, 42, 91421, 
                  tzinfo=<DstTzInfo 'Africa/Abidjan' GMT0:00:00 STD>)
>>> tz.normalize(tz.localize(d))
datetime.datetime(2009, 9, 2, 14, 45, 42, 91421,
                  tzinfo=<DstTzInfo 'Africa/Abidjan' GMT0:00:00 STD>)

As you can see, time has not been changed as a result of localize/normalize operations. However, if .replace is used:

>>> d.replace(tzinfo=tz)
datetime.datetime(2009, 9, 2, 14, 45, 42, 91421, 
                  tzinfo=<DstTzInfo 'Africa/Abidjan' LMT-1 day, 23:44:00 STD>)
>>> tz.normalize(d.replace(tzinfo=tz))
datetime.datetime(2009, 9, 2, 15, 1, 42, 91421,
                  tzinfo=<DstTzInfo 'Africa/Abidjan' GMT0:00:00 STD>)

Which seems to make adjustments into datetime.

Question is - which is correct and why other's wrong?

Solution 1:

localize just assumes that the naive datetime you pass it is "right" (except for not knowing about the timezone!) and so just sets the timezone, no other adjustments.

You can (and it's advisable...) internally work in UTC (rather than with naive datetimes) and use replace when you need to perform I/O of datetimes in a localized way (normalize will handle DST and the like).

Solution 2:

localize is the correct function to use for creating datetime aware objects with an initial fixed datetime value. The resulting datetime aware object will have the original datetime value. A very common usage pattern in my view, and one that perhaps pytz can better document.

replace(tzinfo = ...) is unfortunately named. It is a function that is random in its behaviour. I would advise avoiding the use of this function to set timezones unless you enjoy self-inflicted pain. I have already suffered enough from using this function.

Solution 3:

This DstTzInfo class is used for timezones where the offset from UTC changes at certain points in time. For example (as you are probably aware), many locations transition to "daylight savings time" at the beginning of Summer, and then back to "standard time" at the end of Summer. Each DstTzInfo instance only represents one of these timezones, but the "localize" and "normalize" methods help you get the right instance.

For Abidjan, there has only ever been one transition (according to pytz), and that was in 1912:

>>> tz = pytz.timezone('Africa/Abidjan')
>>> tz._utc_transition_times
[datetime.datetime(1, 1, 1, 0, 0), datetime.datetime(1912, 1, 1, 0, 16, 8)]

The tz object we get out of pytz represents the pre-1912 timezone:

>>> tz
<DstTzInfo 'Africa/Abidjan' LMT-1 day, 23:44:00 STD>

Now looking up at your two examples, see that when you call tz.localize(d) you do NOT get this pre-1912 timezone added to your naive datetime object. It assumes that the datetime object you give it represents local time in the correct timezone for that local time, which is the post-1912 timezone.

However in your second example using d.replace(tzinfo=tz), it takes your datetime object to represent the time in the pre-1912 timezone. This is probably not what you meant. Then when you call dt.normalize it converts this to the timezone that is correct for that datetime value, ie the post-1912 timezone.

Solution 4:

I realize I'm a little late on this... but here is what I found to work well. Work in UTC as Alex stated:

tz = pytz.timezone('Africa/Abidjan')
now = datetime.datetime.utcnow()

Then to localize:

tzoffset = tz.utcoffset(now)
mynow = now+tzoffset

And this method does handle DST perfectly