Random image

Epigno Tech Blog

July 4, 2018

Malik Olivier Boussejra

Timezones are hell

trivia python postgres

Timezones are hell

Timezones are hell. This is really the kind of hell you don't want to get into, unless you really need to.

After getting timezone issues, I started wishing that all countries switch to UTC!

Alas, timezones, like many things, are here to stay. Until the 19th century, each city had its own local time calculated from the sun's movements. However, interconnections between communities forced the people to agree on a common time, a.k.a a time zone.

Unfortunately, as with all human agreements, timezones do not correlate very well with natural phenomena, here, Earth's geometry. While both are continent-size countries, the US shares more than 5 timezones, while China only has one single timezone. Then some countries include energy-saving time changes. They even move 1 hour in some countries during specific periods (e.g. Morocco during the month of Ramadan). Or did you know that an island in the Pacific changed date after its leaders decided to move to the other side of the international date-changing line?

Hopefully a lot of experienced developers thought really hard to solve these issues, both for database management systems and programming languages.

Timezones in Python

Let's take a look at Python's most known solution to the timezone mess.

Let's fire up a Python REPL:

>>> from datetime import datetime
>>>
>>> datetime.now()
datetime.datetime(2018, 7, 4, 7, 48, 7, 17268)

So it's now July 4th 2018 7:48:07 AM (and some microseconds). This is my local time.

Python has a handy function to get UTC time:

>>> datetime.utcnow()
datetime.datetime(2018, 7, 3, 22, 51, 17, 827155)

So it's now July 3rd 2018 22:51:17 (and some microseconds) in UTC time.

But where is the timezone information? How do I know, given a python datetime, what is the the timezone of the specific time? Well, that may sound harsh, but you cannot.

By default, all datetimes in python are said to be "timezone-unaware"!

Timezone-awareness in Python

As it appears, Python has two datetime types. One is said "timezone-unaware", while the other is said "timezone-aware".

Let's make a timezone-aware datetime. For that we will use the pytz module.

>>> import pytz
>>>
>>> # Create a timezone object
>>> tz = pytz.timezone('Asia/Tokyo')
>>> datetime.datetime.now(tz
datetime.datetime(2018, 7, 4, 8, 2, 39, 637167, tzinfo=<DstTzInfo 'Asia/Tokyo' JST+9:00:00 STD>)

Cool! As you see, the datetime.now function can take a timezone object as argument, if you do so, it will return a the current time in the given timezone as a timezone-aware datetime.

As you can see, if the tzinfo attribute is set, then a datetime becomes timezone-aware. You can easily check if a datetime is tz-aware as follows:

def am_i_tzaware(dt):
    return dt.tzinfo is not None

However, there is something you should never, never do (unless you know what you are doing): manually changing the tzinfo attribute of a datetime, either by setting the attribute or using the replace method. This WILL NOT do what you think it will do. You will get behavior that 99.99% of people do not want.

Let's try:

>>> from datetime import datetime
>>> import pytz
>>>
>>> tokyo_tz = pytz.timezone('Asia/Tokyo')
>>> paris_tz = pytz.timezone('Europe/Paris')
>>> tokyo_time = datetime.now(tokyo_tz)
>>> paris_time = tokyo_time.replace(tzinfo=paris_tz)
>>> paris_time
datetime.datetime(2018, 7, 4, 8, 12, 23, 903127, tzinfo=<DstTzInfo 'Europe/Paris' LMT+0:09:00 STD>)

Please take a clear look at the information after the tzinfo keyword. WTF??? What the heck is this 9 minutes offset??? Have you ever heard of a timezone named LMT+0:09:00???

Deeper look into pytz

I know what you are thinking: "Please explain what is this 9 minutes offset! If my program keeps on like this, calamity will happen."

Well, these are historical timezones (yes, they existed!). And people dealing with times from the first half of the 20th century may need them.

pytz can actually deal with the timezones of almost any region after the 1900s.

And if you only use the timezone object as is, it may show a historical offset for the area.

>>> tokyo_tz
<DstTzInfo 'Asia/Tokyo' LMT+9:19:00 STD>
>>> paris_tz
<DstTzInfo 'Europe/Paris' LMT+0:09:00 STD>

When you use the now function, the timezone object will update itself to use the correct offset that politics-driven people happened to choose at the current time in the selected region. The timezone object will not be updated if you directly use it with the tzinfo attribute. But how do I convert the time to another timezone then?

Here is what you can do with a tzaware_localize function:

import datetime
import pytz


def tzaware_localize(dt: datetime.datetime, local_tz: str):
    """
    Takes a TZ-unaware datetime in local time *or* a TZ-aware datetime and turns
    it into a TZ-aware datetime in the provided timezone.
    :return: TZ-aware DateTime in the provided timezone
    """
    if dt.tzinfo is None:
        return pytz.timezone(local_tz).localize(dt)
    else:
        return dt.astimezone(pytz.timezone(local_tz))

Let's try to use it:

>>> # Localize TZ-unaware datetimes to the selected timezone
>>> tzunaware_in_tokyo = datetime.datetime.now()
>>> tzaware_in_tokyo = tzaware_localize(tzunaware_in_tokyo, 'Asia/Tokyo')
>>> tzaware_in_tokyo
datetime.datetime(2018, 7, 4, 8, 32, 32, 680085, tzinfo=<DstTzInfo 'Asia/Tokyo' JST+9:00:00 STD>)
>>>
>>> ### Convert a datetime to another timezone
>>> now_in_paris = tzaware_localize(tzaware_in_tokyo, 'Europe/Paris')
>>> now_in_paris
datetime.datetime(2018, 7, 4, 1, 32, 32, 680085, tzinfo=<DstTzInfo 'Europe/Paris' CEST+2:00:00 DST>)

Basically, if you add timezone information to a timezone unaware datetime, please use the localize method of the datetime object. Please use astimezone if you want to convert a tz-aware datetime to a different timezone.

Python ecosystem

As a side note to the Python world, numpy does not currently support timezone-aware datetimes at the time of writing, while pandas does!

Though, I would advise you to store and process all times as UTC. They should only be converted when you need to show them to the user.

Here is how you can convert a series to UTC tz-aware datetime in pandas:

from datetime import datetime
import pandas as pd

utc_timestamp = pd.to_datetime(datetime.now(), utc=True)

For communication with a database system, there are plugins for SQLAlchemy that handles tz-aware datetimes. Please consider sqlalchemy-utc for example.

Databases: Example with PostgreSQL

You should store all times as UTC in databases. PosgreSQL has two types to store datetimes:

  • TIMESTAMP WITHOUT TIME ZONE
  • TIMESTAMP WITH TIME ZONE

But don't be fooled. None of them actually store time zone information.

A TIMESTAMP WITHOUT TIME ZONE is just what its namesake is, a timestamp without time zone, while a TIMESTAMP WITH TIME ZONE is implicitly converted to the connection's timezone everytime the value is read. That's the only difference. In term of data representation, they store the same thing.

If the connection timezone is not set, a TIMESTAMP WITH TIME ZONE will be converted to whatever default your PostgreSQL instance is configured with (usually UTC).

You can check your connection's timezone with:

SHOW TIMEZONE;

By default, the query should return 'UTC'.

You can set your connection's timezone with:

SET TIME ZONE 'Asia/Tokyo';

Then SHOW TIMEZONE will return 'Asia/Tokyo', and all TIMESTAMP WITH TIME ZONE will be converted from UTC to the timezone (+9 hour offset for Japan for a post-war datetime).

You should especially be cautious when dealing with dates! What date a timestamp falls on is dependent on the timezone! The date now is 2018-07-04 in Tokyo, but 2018-07-03 in New York. That's the previous day!

SELECT * FROM my_table
-- Here, the UTC datetime will be converted to the date in the connection's timezone!
WHERE my_table.datetime_column::date = '2018-07-04'::date

If your business analytics software happens to connect to a PostgreSQL database, do not forget to set the connection timezone appropriatly! If not, you may get surprises when you query the database!

NB:

  • You can store timezones as a separate column in a database.

SQLAlchemy-Utils even has a dedicated type for it! So if need be, you may store the timezone of a timestamp in another column.

  • If you ever need to explicitly convert to a specific timezones, you can use the

"AT TIME ZONE" operator in PostgreSQL.

  • If you are wondering how pytz knows which offset is used at what time in

what area, they are using the official Time Zone Database.

Author

Malik Olivier Boussejra

CTO at Epigno