When you cross DST switchovers, for example, you may magically gain or lose an hour. A naive milliseconds calculation will miss that. It gets worse because DST in a country isn't a static thing either, but will change with the laws over time. So over decades, you need to have several lookup tables.
And people are ambiguous. Is January 31 plus 1 month the end of February or the beginning of March? What about during a leap year?
And a "day" isn't necessarily 24 hours (because of DST, again). Humans doing day math across those boundaries will just ignore (or really, never think about) the missing or gained hours, but computers have to explicitly account for it.
Then once you throw in different time zones it gets even crazier, especially for the half hour and 45 minute zones.
It's okay to store datetimes in epoch milliseconds (ideally with a separate field for the time zone, not just offset, which holds less information). But you can't easily/correctly do math on that without more specific instructions and cultural/locale adjustments.
