Dates for booleans

Here’s a little maxim for data modeling: use dates instead of booleans. In a lot of situations, this lets you capture more information without losing anything compared to a boolean.

A common example is implementing a status column with a datetime foobared_at column instead of a boolean is_foobar. If it’s null, then that row is not foobar. If the column is filled, then you not only know that it is foobar, but also since when.

Having that extra information is often useful for debugging, but can also be good for implementing business logic. For example, you might want to email all users who subscribed to your newsletter in the last week to send them an introductory offer. You can’t do that with a boolean subscribed column, but it’s easy with a datetime subscribed_at.

A further improvement in that data model would be to have a subscriptions table with a foreign key to your users, with the created_at or updated_at column on the subscriptions table holding this information.

This approach is often taken to implement soft deletes with a deleted_at column, but I tend to agree with this StackOverflow answer explaining how soft deletion is usually an antipattern. Soft deletion is convenient and supported by many database management libraries, but causes complexities in all queries and can lead to various bugs and annoyances as described there.

In any case, having a preference for date columns instead of booleans means you can gather extra information while it’s available so that it’s there if you need it later. Looking at it another way, boolean flags are often an [antipattern]( of-the-humble-boolean/), and using a datetime instead can mitigate that. Just make sure that there isn’t a better solution that lets you avoid this kind of flag logic altogether.