pytz: The Fastest Footgun in the West

Whenever I give a talk about time zones, someone comes up to me afterwards and tells me that they have broken code currently in production, because they misunderstood how pytz works. This is because pytz uses its own non-standard interface for handling time zone information that is partially but not entirely compatible with the way Python's datetime library was intended to work, which leads to a lot of confusion from people naively using pytz as a time zone provider. This incompatibility is why, as of Python 3.6, the tzinfo documentation recommends dateutil.tz rather than pytz as an IANA time zone provider. [1]

In this post, I will cover both time zone models and if I cannot convince you to switch to dateutil.tz, at least provide some intuition about the differences between pytz and the standard time zone model.

Python's time zone model

In the datetime module, Python provides support for time zones rather than time zone offsets - which is to say that a datetime.tzinfo object is expected to provide not a fixed offset and name but a set of rules for what the time zone information is as a function of the datetime. This is so that something like this will work:

from dateutil import tz
from datetime import datetime, timedelta

NYC = tz.gettz('America/New_York')
dt_winter = datetime(2018, 2, 14, 12, tzinfo=NYC)
print(dt_winter)
# 2018-02-14 12:00:00-05:00

dt_spring = dt_winter + timedelta(days=60)
print(dt_spring)
# 2018-04-15 12:00:00-04:00

If NYC were a static, fixed offset, you'd need to attach a different tzinfo to each datetime depending on whether or not you're in standard or daylight time, and any time you did any math on your datetime, you'd have to redo the calculations in case the offset had changed. Thus, Python's model is that any tzinfo subclass should implement the following three methods:

  • tzname(self, dt): The name of the offset at the given datetime (e.g. EST, PDT)
  • utcoffset(self, dt): The offset from UTC at the given datetime
  • dst(self, dt): The difference between the current offset and the zone's "standard offset" [2]

These values are not only implemented as a function, they are also invoked lazily, so there are no hooks in the datetime constructor that call these – they are only invoked when a user wants to know one or more of these pieces of information.

pytz's time zone model

The biggest mistake people make with pytz is simply attaching its time zones to the constructor, since that is the standard way to add a time zone to a datetime in Python. If you try and do that, the best case scenario is that you'll get something obviously absurd:

import pytz
from datetime import datetime, timedelta

NYC = pytz.timezone('America/New_York')
dt = datetime(2018, 2, 14, 12, tzinfo=NYC)
print(dt)
# 2018-02-14 12:00:00-04:56

Why is the time offset -04:56 and not -05:00? Because that was the local solar mean time in New York before standardized time zones were adopted, and is thus the first entry in the America/New_York time zone. Why did pytz return that? Because unlike the standard library's model of lazily-computed time zone information, pytz takes an eager calculation approach. Whenever you construct an aware datetime from a naive one, you need to call the localize function on it:

dt = NYC.localize(datetime(2018, 2, 14, 12))
print(dt)
# 2018-02-14 12:00:00-05:00

Each pytz time zone contains a list of possible fixed offset "time zone" objects that are valid at different times in that zone, and the localize function figures out which one is valid at that local date and time and attaches it. In this case it detects correctly that 2018-02-14 should be EST with offset -05:00 and DST -00:00. Now what happens when you perform the arithmetic on a localized datetime?

from datetime import timedelta

dt_spring = dt + timedelta(days=60)
print(dt_spring)
# 2018-04-15 12:00:00-05:00

Since the localize function eagerly attached EST to the datetime, the offset is not updated in response to the arithmetic. In order to fix this error, any time you do any arithmetic on a pytz-aware datetime, you need to call the normalize function:

print(NYC.normalize(dt_spring))
# 2018-04-15 13:00:00-04:00

This is, again, just eagerly doing the calculation that would be done lazily by a dateutil.tz time zone object [3].

Ambiguous datetimes

Why was pytz designed this way, given that it doesn't mesh well with Python's standard time zone model? Consider the scenario of an ambiguous datetime, which occurs during a daylight saving time transition, e.g. 2018-11-04 01:30-04:00, and an hour later, 2018-11-04 01:30-05:00. How would you write a function that takes the 2018-11-04 01:30 portion of that datetime, and returns the correct answer? You can't, because there are two correct answers.

pytz is able to solve this problem because during the localize step, the time zone could take in the additional information of whether you wanted to be on the DST or STD side of the transition:

dt_dst = NYC.localize(datetime(2018, 11, 4, 1, 30), is_dst=True)
print(dt_dst)
# 2018-11-04 01:30:00-04:00

dt_std = NYC.localize(datetime(2018, 11, 4, 1, 30), is_dst=False)
print(dt_std)
# 2018-11-04 01:30:00-05:00

This is because some of the information about the datetime (which side of DST it represents) is now encoded in the tzinfo that was attached to it. Using the standard Python interface, this problem was not solved until Python 3.6 with the introduction of PEP 495, which added the fold attribute to the datetime class. This allows the "which side of an ambiguous datetime do I fall on" decision to be encoded in the datetime itself, allowing for lazy calculation of ambiguous datetimes. If dt.fold is 0, ambiguous datetimes resolve to the first occurrence of a time in a given zone, if it's 1, they resolve to the second occurrence. So to represent the example above:

from dateutil import tz
NYC_du = tz.gettz('America/New_York')

dt_dst = datetime(2018, 11, 4, 1, 30, fold=0, tzinfo=NYC_du)
print(dt_dst)
# 2018-11-04 01:30:00-04:00

dt_std = datetime(2018, 11, 4, 1, 30, fold=1, tzinfo=NYC_du)
print(dt_std)
# 2018-11-04 01:30:00-05:00

Additionally, dateutil is able to backport this to earlier versions of Python by providing a tz.enfold function that will, if necessary, create a datetime subclass providing the fold attribute. So, on Python 2.7 you get:

>>> tz.enfold(datetime(2018, 11, 4, 1, 30), fold=0)
datetime.datetime(2018, 11, 4, 1, 30)

>>> tz.enfold(datetime(2018, 11, 4, 1, 30), fold=1)
_DatetimeWithFold(2018, 11, 4, 1, 30)

Now that this issue is resolved, pytz no longer has the advantage of being the best way to handle ambiguous datetimes, but retains the disadvantage of its somewhat clunky interface and eagerly-calculated time zone information.

Fastest footgun in the west

The title of this post claims that pytz is the fastest footgun in the west, by which I mean that pytz is a quite well-optimized library, and historically it has been faster than dateutil.tz. The performance gap has been significantly reduced in the last several releases [4], but there is still a persistent gap in performance for some use cases owing to the lazy vs. eager nature of the calculations. To demonstrate, I've timed pytz==2018.3 and python-dateutil==2.7.0 on Python 3.6, using IPython's %timeit magic, and put the timing numbers in comments after each function:[6]

import pytz
from dateutil import tz

from datetime import datetime

NYC_p = pytz.timezone('America/New_York')   # 1.53 µs
NYC_d = tz.gettz('America/New_York')        # 863 ns

dt_p = NYC_p.localize(datetime(2018, 11, 1))    # 35.4 µs
dt_d = datetime(2018, 11, 1, tzinfo=NYC_d)      # 1.38 µs

dt_p.utcoffset()        # 655 ns
dt_d.utcoffset()        # 13.9 µs

As you can see, pytz time zones are more costly to construct and the initial localize call is much slower than even dateutil's utcoffset() call, but pytz's utcoffset() call beats out dateutil's by a wide margin (because the result is cached). If you plan to construct a bunch of datetimes and query their time zone information relatively infrequently (an average of 2-3 times per datetime), it seems that dateutil beats pytz in "total time to first utcoffset call":

NYC_p.localize(datetime(2018, 11, 1)).utcoffset()   # 37 µs
datetime(2018, 11, 1, tzinfo=NYC_d).utcoffset()     # 15.2 µs

The story is similar for converting from one time zone to another (though pytz does better than I would expect):

LA_p = pytz.timezone('America/Los_Angeles')
LA_d = tz.gettz('America/Los_Angeles')

dt_p.astimezone(LA_p)       # 7.53 µs
dt_d.astimezone(LA_d)       # 31.5 µs

dateutil is slower in this operation because it needs to calculate the UTC offset for both New York and Los Angeles, and pytz evidently is faster at creating localized datetimes from UTC than from naive datetimes (otherwise I would expect to pay at least the cost of another localize). If you are planning on doing only one time zone conversion per localized datetime, though, the full cost of the operation is:

NYC_p.localize(datetime(2018, 11, 1)).astimezone(LA_p)  # 44.8 µs
datetime(2018, 11, 1, tzinfo=NYC_d).astimzone(LA_d)     # 32.8 µs

In neither case is there a clear "overall" winner on performance. If you plan on localizing a datetime and then repeatedly query its utcoffset or convert it to other time zones, pytz may perform better, because the offset values are cached. [5] If you are only making an average of 2-3 utcoffset calls per datetime, you may get better performance out of dateutil.

I suspect that for most practical cases, the marginal cost of additional time zone calculations can be narrowed significantly with little memory overhead by a implementing even a modest Least Recently Used cache (update: not as feasible as I originally thought ) [7], but it is worth noting that pytz's design brings an effectively infinite cache "for free". In any case, the standard disclaimer applies — don't worry about these kinds of micro-benchmarks until you're certain that they are the bottleneck of your operation.

Conclusions

At the time of its creation, pytz was cleverly designed to optimize for performance and correctness, but with the changes introduced by PEP 495 and the performance improvements to dateutil, the reasons to use it are dwindling. As mentioned in the previous sections, there are some use cases where pytz's IANA time zones are faster, but common use cases where they are slower as well. Historically, they provided a "more correct" time zone implementation, but now they solve the ambiguous time problem in a way that is inconsistent with Python's time zone model.

The biggest reason to use dateutil over pytz is the fact that dateutil uses the standard interface and pytz doesn't, and as a result it is very easy to pytz incorrectly. Even if you now know the right way to use pytz, are you sure that you're not going to pass your datetime to a function expecting something that uses the standard tzinfo interface? Are you sure that anyone maintaining your code or consuming its outputs are going to know to avoid these mistakes?

Footnotes

[1]Scroll down to the "See also" section, which doesn't have its own anchor link.
[2]I'll note that the existence of a "standard offset" may be an unwarranted assumption of Python's time zone model. There is some identifiable "standard" offset in all time zones that I know about, but there's nothing stopping people from observing a strange time zone that switches between 3 equally valid "time zones" during the course of a year for reasons other than daylight saving time.
[3]By providing a completely different tzinfo object, this also interferes with Python's model for time zone aware arithmetic semantics, which depends on two datetimes in the same zone satisfying dt1.tzinfo is dt2.tzinfo.
[4]Among other improvements in version 2.7.0, dateutil has added a dateutil.tz.UTC singleton (previously dateutil.tz.tzutc() constructed a new object for each call) and started caching calls to tz.gettz and generally reduced the number of constructor calls to other time zone objects.
[5]It is possible, of course, to simply add caching to dateutil's time zone functions, but pytz has the inherent advantage that the tzinfo cache is built in to each datetime object – in order to implement a cache for dateutil's time zones, each tzinfo has to maintain a history of datetime values for which the offset has been calculated and map that to the proper lookup values. By contrast, pytz stores this mapping in the tzinfo argument, since each dateutil stores a reference to the offset that it was localized to, at the (memory) cost of proliferating one extra tzinfo object per possible offset (in most zones this will be 3-5 extra tzinfo objects).
[6]Keep in mind, both dateutil and pytz make extensive use of caching, so these numbers are the "asymptotic" behavior. These numbers are mainly useful if you end up doing basically the same thing over and over again in a tight loop.
[7]Update 2020-04-01: An LRU cache for datetimes is not as feasible as I had originally thought, because the UTC offset is a part of the datetime's hash calculation, so it is not possible to simply look it up in a dictionary without the expensive operation of constructing a new datetime. There are other strategies and cache designs that could work, and I proposed one such scheme in bpo-35723, but likely the need for these will be mooted by the acceptance of PEP 615, since every operation in the C extension in the reference implementation is considerably faster than pytz, even without any caching.