PostgreSQL's missing DateDiff function

DATEDIFF is an extremely useful function for analytical queries. Given two timestamps it'll tell you how many days, weeks, months, hours, etc they are apart. Here we show an implementation of this function for Postgres.

PostgreSQL's missing DateDiff function

Photo by Daniele Levis Pelusi / Unsplash

Data warehouses like Redshift and Snowflake have a super useful DATEDIFF function – given two timestamps and a date part (hour, year, week, etc) it'll return how far apart they are. For example,

DATEDIFF('week', '06-01-2021', '06-28-2021') returns 4

This function can be used to bucket times together, like when doing a cohort analysis. Unfortunately Postgres simply doesn't have it. For operational data it's probably not often used, but if you do analytical queries it can be pretty helpful.

If you want to use Postgres as a data warehouse you'll probably want it. A great blog by sqlines suggests an implementation, but I found it didn't quite match the functionality of most data warehouses.

Why is this even a function? Why not just use date_part? Because time rolls over: the last month of the year is 12 and the first is 1, so naively using date part would give us -11.

The DATEDIFF code

I'll jump straight to the code for those who like to see the answer first, and further down explain how it works

This gist creates a function in Postgres that implements the DATEDIFF function found in Snowflake, BigQuery, and Redshift.

This function take a time unit and two dates, and counts the number of date boundaries crossed between them. It will always return an integer, so it's very useful for grouping date differences together.

So what does counting date boundaries mean? It's best illustrated with an example

DATEDIFF('year', '12-31-2020', '01-01-2021') returns 1 because even though the two dates are a day apart, they've crossed the year boundary.

Similarly DateDiff('week', '05-02-2021', '05-03-2021') is 1, because 5/02/21 is a Sunday and 5/03/21 is a Monday

Note that Postgres uses ISO 8601 week numbering, so weeks will always start on Mondays. This isn't consistent across databases – Redshift uses Sunday, while in Snowflake it's configurable.

How DATEDIFF works

The code basically works in two parts. It computes year, quarter, and month boundaries by subtracting integers. It computes weeks and below by truncating.

Years, Quarters, and Months

Years is easy: 2021 - 2020 = 1

Months – just use DATE_PART to get the month as an integer (1-12). Subtract the two and add in the year expressed in months to handle rollover.

2021/2 - 2020/11 would be (2 - 11) + 1 * 12 = 3

Quarters is effectively the same as months (since DATE_PART gives us quarters as an integer from 1-4).

Weeks and Days

Weeks and days use a slightly different approach. Weeks are special because they don't fit evenly into months, years or anything bigger. They only fit evenly into days.

The way to count weeks is to truncate the start and end timestamps to the first day of the week, then subtract days. That will give us an integer that's a multiple of 7. Note that the 'first day of the week' is not uniform across databases. Postgres uses Monday.

RETURN DATE_PART('day', (DATE_TRUNC('week', end_t) - DATE_TRUNC('week', start_t)) / 7);

Subtracting the days returns an interval, so we use DATE_PART to get an integer number of days. Since this is always divisible by 7, we now have our number of weeks.

Days is computed the same way, without dividing by 7. Instead of the start of the week we truncate to the beginning of the day to ensure the interval is an exact number of days.

Hours on down

Hours, minutes, and seconds are built out just like months and quarters.

hours = days * 24 + (DATE_PART('hour', end_t) - DATE_PART('hour', start_t));

We extract the hour as an integer (0-23) for each timestamp, subtract from each other, and multiply by the days expressed as hours.

Minutes is the same calculation, but it uses hours * 60 since we already have it.

Seconds is also the same, using minutes * 60.

If you need to handle milliseconds or below it's easy to extend the code. I stopped at seconds.


Check us out on the Data Engineering Podcast

Find it on the podcast page or stream it below