Time is difficult. Not only from the point of view of special and general relativity but also for simple calendars. Time zones, Gregorian vs. Julian calendar, leap years, and leap seconds make this awfully difficult. Time and date is a complex concept in our society, so we must capture this complexity in software. In general, a computer program has two easy-to-use approaches Work with timestamps or work with ISO-8601 date time strings. Both are perfectly valid approaches with different advantages and disadvantages. This article discusses issues introduced by leap seconds and the Python approach with the datetime package.

Leap seconds

First, let’s look at leap seconds. The Gregorian calendar introduces an additional day (February 29) for every year divisible by four, except for years divisible by 100. However, years divisible by 400 still receive the additional day. This brings the average year to 365.2425 days, which is very close to the length of a tropical year. I recommend reading Matt Parker’s Humble Pi book for a lightweight and funny introduction to this issue.

Earth rotation slows down and varies over time due to many effects. To keep our clocks (UTC) in sync with earth’s rotation, we need to squeeze in leap seconds from time to time. As Opposed to perfectly deterministic leap years (you can write a program to compute leap years for year 3000), leap seconds are based on astronomical measurements. We cannot reliably predict all leap seconds until the year 3000. Unix time stamps are entirely agnostic of the existence of leap seconds. The expression time() % 60 will always give you the number of seconds since the last minute tick. This makes it necessary to synchronize the Unix time stamp with UTC whenever a leap second is introduced. By making a Unix time stamp slightly longer, we can accommodate the leap second. For example, after 60 ticks of the Unix clock, 61 ticks were recorded at an atomic clock. Some applications stretch the time stamp over a full day, some only over a minute, and some just double the last second.

This approach is straightforward to use but can cause issues if you count external events over time. Assume an event occurs every second. If you use Unix time stamps to aggregate the count, the count will be off if leap seconds are introduced. There is one additional external event in the time frame that accommodates the leap second. Depending on the application,this could be an issue. On the other hand, time stamps are very easy to use. They work independently of local time zones. Two people in different time zones calling time() at the same time (i.e., talking on the phone) get approximately the same value. Practically every external system can store timestamps as a (hopefully) 64-bit integer.


An alternative approach to work within the ISO-8601 framework and Python’s datetime module. This is already where it becomes difficult. datetime objects come in two flavors: time-zone aware or unaware. By default the object is time-zone unaware storing local time information:

from datetime import datetime

now = datetime.now()

# 2022-07-19T22:19:36.925858

# 1658261976.925858

This might be perfectly fine for desktop applications but becomes problematic in distributed or online applications. In these cases, you should always prefer time-zone aware datetime objects with pytz. The time stamp printed in the last line would also work as an alternative, as described in the previous section.

Interestingly, there is also datetime.utcnow(). Let’s look at its properties.

utcnow = datetime.utcnow()

# 2022-07-19T20:23:24.130360

# 1658255004.13036

diff = now.timestamp() - utcnow.timestamp()
# 6972.795498132706

The first print gives the current UTC time, a two-hour offset for my local time. Earlier, I said, time stamps work independently of time zones. However, here we see utcnow is offset by approximately 7200 seconds or 2 hours. How can this be? The issue lies with the Python implementation. utcnow() returns the time in UTC time zone, but as a time-zone unaware object. There is no difference between the object returned by utcnow() and an object returned by now() two hours earlier.

I think this is a very sublet pitfall, and I think utcnow() should never be used in the first place. Use time-zone aware objects, local unaware objects, or time stamps, but do not mix!