Consider the following code:

1
2
3
4
5
6
7
8
9

>>> import datetime as dt

>>> later_date = dt.datetime.now().replace(hour=18, minute=39, second=59, microsecond=0)
>>> earlier_date = later_date - dt.timedelta(days=1, hours=5, minutes=7, seconds=20)

>>> later_date
datetime.datetime(2025, 8, 31, 18, 39, 59)
>>> earlier_date
datetime.datetime(2025, 8, 30, 13, 32, 39)

The obvious direction

Now I want to get the difference between these two, so essentially the timedelta. The larger minus the smaller one gives me nearly what I expected:

1
2

>>> later_date - earlier_date
datetime.timedelta(days=1, seconds=18440)

Now my expectation would have been a day, $5$ hours, $7$ minutes and $20$ seconds. But since $18440$ seconds is the same amount of time, this is fine and correct.

The non-obvious direction

What happens if we subtract the other way round?

My first intuition would be to get a timedelta(days=-1, seconds=-18440). Well, not really…:

1
2

>>> earlier_date - later_date
datetime.timedelta(days=-2, seconds=67960)

What happens?

After thinking about time a bit, this actually does make sense: $67960 + 18440 = 86400$, which is exactly the number of seconds of one day. So this timedelta can be understood as: “Take your first timestamp, subtract $2$ days and then re-add $67960$ seconds to the result”. Which actually is the same as saying “Subtract one day and then subtract $18440$ seconds”.

Danger!

And now here’s what I find dangerous:

1
2

(earlier_date - later_date).seconds
67960

Which means: As a developer, I have to be aware that I have got a negative timedelta, altough the seconds return a positive value!

But a simple solution

It is actually sufficient to just check total_seconds():

1
2

(earlier_date - later_date).total.seconds()
-104840.0

This again gives me something I can work with.

Again a learning here: Note that the total_seconds method returns a floating point number, as I can have more precision than seconds in a timedelta!