Hacker News Reader

Comment by 015a | Hacker News Reader

I've personally worked on issues stemming from this; situations where the date is important, but not the time, such as correctly reporting when it is someone's birthday.

One of the earliest axioms we drill into programmers is "use UTC, use date-time types". Don't store dates as strings, use your database's DateTime type.

If your database has a Date sans-Time type, use that. But many databases don't, and that's where the axiom starts breaking down, because its very natural to say "well, I'll use the next best thing, a DateTime type". But this is wrong! Its far, far better to simply use a YYYY-MM-DD string. I cannot stress how much better this is, but its not obvious why its better to someone who has had that axiom drilled into their heads, and has never ran into this situation before.

View on HN

Worst case of this I've personally had to fix was an airline where you couldn't purchase tickets if you lived anywhere with a negative time zone offset. They're a small European carrier so it was never a very big issue for them, and all the developers and management were in Europe so they never noticed it first hand.

Basically when you tried to purchase a ticket, their frontend would take the date-of-flight, and adjust it by the user's timezone. So a ticket for 2015-01-30 00:00 would become 2015-01-29 20:00, or whatever. The dates in the request did not match what was expected for your booking code, and 500 error.

As a developer in Australia, I have made these sorts of errors that I didn't notice until my colleagues in the US picked them up. I am actually very happy that America is in a negative timezone, because I believe these errors would be a lot more prevalent if they were not.

Yeah, "always use UTC" is overrated. It's useful for things that have happened, but quickly becomes unintuitive for things that will happen - including birthdays, meetings, ...

It can even become wrong for future dates. Always use UTC is a broken, bad rule for future dates, it only makes sense for things that have already happened.

As for how it breaks, you have a booking for 8pm the day after the start of summer time. You stored the date/time in UTC. Government changes the start of summer time. Your database now thinks it's at 7pm because it uses UTC.

Here's some examples:

https://codeofmatt.com/on-the-timing-of-time-zone-changes/

Also the EU is seriously debating getting rid of clock changes, and it comes up in the UK semi-regularly.

Your database will not fix your bad data when that happens.

But would using local time help in this case? For example, what if the future event is something like the solstice?

It's usecase dependent so it seems like the application should deal with this, at which point persisting in UTC is still a reasonable choice.

When you schedule an event in the future like “meet me at 14:00 in cafe foo”, this means local time and it’s going to mean 14:00 local time even when the timezone changes.

If you store this in UTC and the timezone changes, you end up converting to the wrong time.

So for future events you should store local time with a location you can map to a timezone.

The solstice is based on astronomical phenomena. If you store it in UTC it will always be right.

Unless there is a leap second scheduled on short notice or you install the updated tz database rather late. For astronomical events or long term measurements (think CERN) there is astronomical time, which does not add leap seconds.

But if you want to have precise date time far in the future you must not use UTC based milliseconds representations unless you want to check for necessary database migrations every time the time zone definitions change.

If you store zoned time in UTC then the zones or UTC change then the time can be formatted correctly in the future.

My point in mentioning the solstice is that while some events are at “2pm”, others are “when something happens”.

Sure, UTC is not suitable for conducting astronomy. It is suitable for working with human friendly times, since it is the foundation of timezones.

Yep! It sounds like overkill, but the suggestion I've seen that seemed to cover all bases was to store:

(1) The time in the local timezone (+ the timezone itself) (2) The UTC conversion (3) The version of the IANA database used to compute (2)

Then you can use (2) the UTC time for most computations/queries, but if a timezone changes then you can use (3) to tell which UTC dates need to be updated, and (1) to recompute the new correct version of (2).

That seems like overkill for most apps. Better to just store the date/time without timezone, and either the windows or linux timezone name of the location (which has far more specificity than the ISO 8601 offset). Then calculate the actual time on the fly when pulling it out of the DB.

No mucking around with IANA databases, everything's easy to program, server usually keeps itself up-to-date.

Even more common:

* Your meeting is scheduled for 2pm tomorrow.

* Oh oh! Your meeting got bumped to the next day.

* Daylight savings switched overnight.

Some things are measured in relative time. Some things are measured in absolute time. Never confuse the two in your database.

Yes, timezones do not at all work with reservations for physical objects.

I work on a registration and reservation system for rowing boats in my rowing club https://github.com/dsroklub/roprotokol

The only time that matters is whatever the clock in the clubhouse shows (Danish time) If you sign out a boat from the UK (or more realistically from a laptop that thinks it is in the UK) at 14:00 it is still 14:00 Danish time. This is surprisingly difficult to achieve with JS Date object, browsers and popular JS components.

So the Temporal Plain time looks promising.

And especially the explicit timezones. Because even without timezone you need to handle DST. For example my system needs to know that if you start a 3 hour trip at midnight a certain day in the spring, you are expected back at 04:00.

I just hope that browsers and components will support is properly.

> Your database will not fix your bad data when that happens.

Because it’s not bad data, HaHaOnlySerious. I tell developers that if the database attribute accepts a value it must be in the domain of allowable values. Often people refuse check constraints in order to avoid expressing business logic in the database.

The GP's example doesn't involve invalid datetimes. The datetimes aren't out of bounds or invalid in any way that a check constraint would detect. They've just become factually incorrect ("bad data"), because they are derived data that wasn't updated when the derivation rules changed (i.e., regulatory changes).

If you're storing future datetimes that semantically represent wall clock time, you need to store the locale time plus the full time zone (such as America/New_York) so that your program does the right thing in response to any common regulatory changes that happen after you store the value. Storing the time zone abbreviation (e.g., EST) is inadvisable, as computers sometimes care whether you asked for EST vs EDT. Storing the time offset (e.g., -500) is incorrect, as it has the same pitfalls as storing UTC - you're precomputing the locale's expected time offset at storage time, and your data won't automatically be corrected if time regulations change.

If you're storing historical timestamps, UTC is fine because you can safely convert it to whatever time zone you want to display, knowing that changes to time zone / DST regulations tend not to affect the past.

> If you're storing future datetimes that semantically represent wall clock time, you need to store the locale time plus the full time zone (such as America/New_York) so that your program does the right thing in response to any common regulatory changes that happen after you store the value.

At this point in the process first normal form flies out the window. Trying to generalize too much can lead you down some weird garden paths. If it looks like you need a function to validate an prospective column value then you probably need to model the value as a relation corresponding to the function parameters. Then you can make it into a foreign key constraint and get on with your project.

I truly appreciate the efforts of those who attempt to expand the utility of datetime value representation to perfect a wider variety of denotational semantics. But with a relational model it may be better to delegate to simpler abstraction sufficient to the specific case.

Personally I solve this problem by just doing everything in UTC. All my clocks, calendars, phones, everything read UTC.

"Always use UTC" works if you really, really always use UTC.

What about interfacing with the outside world, which doesn't always use UTC?

Translate. But frankly I often still use UTC to coordinate with people, especially when coordinating meetings with people in 3 timezones.

Always use UTC is not overrated. It is true, it may be overrated if you don't store time with your date or don't have a datetime type in your database (that makes me think WTF). You are always better using UTC to store dates if you have a datetime type.

Always use UTC for timestamps.

Never use UTC for dates. Like OP says, a simple YYYYMMDD string is better.

This is wrong depending on the use case and where the user is and where the user might be going.

This might be silly, but I kind of wish Date-with-Timezone types existed. It would be nice to be able to take a date and ask "Is does timestamp X fall within date Y", where Y is a calendar day in a particular place.

You don't need a zoned date. You can invert the question and ask if a zoned datetime falls within a plain date.

Temporal is very well designed thanks to building on years of prior art. Give it a try and you'll realize that it leads you to ask those questions the right way.

This is why i use microtime (epoch with miliseconds) over any database supported DateTime thingy. It is supported everywhere out of the box.

I've had this discussion with many confused developers.

Seems like everyone has to stub their own toe on this to get it.

I don't get why inventing some other ways of storing dates that are not RFC 3339 strings.

Multiple reasons. First, the DateTime type in a typical database isn't even an RFC-3339 string; its a unix timestamp. That's it; an unsigned integer. This is very space efficient; far more-so than a string. So, then, why the indirection; why not just use the integer type directly? It communicates to the database engine the intent of the data stored, so the database can do useful things like, for example, time-based queries, interfacing with that integer in more human friendly ways when writing queries, displaying it in a more friendly way... its a low-cost abstraction, generally.

Except, there is the cost intrinsic to this discussion, where going back to strings makes sense. That doesn't mean that decision comes with no cost. If you store someone's birthday as "1990-02-05", then need to answer the question "who has birthdays in the next month", that's actually essentially impossible to do in practically any database query engine. You have the problem of both ignoring the year field of that string (ok, so, maybe we just store "02-05" because the year doesn't matter for birthdays, moreso just 'dates of birth', we can denormalize)... but even beyond that, the closest most if not all engines could get is generating an array of every upcoming day in application code (02-05, 02-06, 02-07, etc) then matching on each. Alternatively, just sort on the field (sans year) then filter in application code, but that's also not happening in the database, which has negatives, and moreover, this would suddenly start failing across year boundaries, but you can special-case that by running a second query starting at 01-01 if the first one didn't return enough results...

There's no easy solution. Time is just really hard. Its good to have as many tools as possible in your belt, but with that you have to know when to reach for one versus another. There isn't one solution to any of these problems.

This. I have through much pain embraced this approach.

It's better to use a native date/time with a CHECK that ensures that the time portion is 0.

You'll have to do a CHECK on the YYYY-MM-DD string anyway to protect your database from bad data, so why not just do it on the native type?

The native type is likely to be more compact, performant, consistent across the database (with strings, some columns may be YYYY-MM-DD, some may be YYYY.MM.DD etc.) and more readily convertible to the client's type system.

Reply on news.ycombinator.com