[vlovich123]

I get that it’s more correct, but it assumes that Europe/Paris is a constant representation of how to apply the timezone-specific stuff but that’s incorrect. For example, ‘2025-06-20T17:00:00+02[Europe/Dublin]’ is a very different time if it’s created today vs if it were created in 1760 [1]. That’s a very extreme example, but timezone rules change and dates created from before the change was announced would be interpreted differently from ones created after they were announced. It’s interesting to me the standard doesn’t embed the creation time of the timestamp in UTC as well.

https://en.wikipedia.org/wiki/Time_in_the_Republic_of_Irelan...

[jrochkind1]

It does not assume that -- implementing libraries need to -- and do, in practice right now -- implement the rules for jurisdictions change the rules for offsets.

That's part of why they are tied to a certain city -- time zone rules are unlikely to bisect a city, although if they did I guess they'd have to deprecate it as a timezone name and use something else! Not sure if this has ever happened.

All of this is kept track of in the IANA Time Zone database, and labels like `Europe/Paris` are from keys into that database, not arbitrary. https://www.iana.org/time-zones

Sometimes when jurisdictions do weird stuff like changing their rules for (say) when Daylight Savings starts with no notice (effective tomorrow!), the libraries can take a bit of time to catch up and be correct again (and have the new version be distributed to all users).

But keeping track of (say) that America/New York on March 25 2024 is UTC-4 but March 2025 1990 is UTC-5 hours (they changed when Daylight Savings started in between those years) is absolutely something current (eg OS) time libraries do.

As well as keeping track of the gregorian correction in (depending on country) October 1582 (skipping over 10 days in the calendar!), when calculating historical intervals. They really do this, already!

That's why you say "Europe/Paris" or "America/New York" instead of "UTC-5", to let the library figure out the rules for offsets at that location on the time specified.

I assume Temporal will do the same. JS environments are usually running on OS's that will already provide this service, the browser or other execution environment won't have to implement it from scratch. Although I think moment.js did it from scratch, and distributes a timezone database with moment.js packages.

[burntsushi]

> As well as keeping track of the gregorian correction in (depending on country) October 1582 (skipping over 10 days in the calendar!), when calculating historical intervals. They really do this, already!

Which libraries do this? Libraries usually implement proleptic calendars, including Temporal[1], which specifically do not account for shifts like this. And indeed, the Temporal docs even call out this specific example.

(I agree with the rest of your comment!)

[1]: https://tc39.es/proposal-temporal/docs/calendars.html#handli...

[jrochkind1]

Ah, I guess not as many as I thought!

Ruby DateTime does it, I hadn't realized it was unusual, if it is!

Here it is particularly called out in documentation with example taking account that April 23rd 1616 in England was not the same day as April 23rd 1616 in Italy, and DateTime knows that! https://ruby-doc.org/stdlib-2.6.1/libdoc/date/rdoc/DateTime....

(That class however is using weird "Date::ENGLAND" constants for locale instead of ISO timezone locales, which is a legacy mistake!)

(I work in archiving cultural history, so probably deal with pre-19th century dates a lot more than the average coder)

update: I was curious what would happen if I gave DateTime a non-existent date...

DateTime.iso8601('1752-09-05', Date::ENGLAND) => invalid date (Date::Error)

it knows! (England skipped some days in September 1752 as part of Gregorian adjustment)

[burntsushi]

Ah yeah, those look like opt-in calendars to me I think. Definitely other datetime libraries do that. They just also require opt-in AFAIK.

[jrochkind1]

Ruby date/time handling is confusing, with too many overlapping classes, developed at different times with odd interfaces and/or semantic edge cases (not unusual on other platforms either), but I believe the things people use as default/recommended standard these days on ruby actually do support gregorian/julian changeover!

But I guess ruby is unusual there! Or I'm wrong.

But good to know most have it as opt-in, anyway!

[burntsushi]

I would need to see the APIs and examples to be able to evaluate it honestly.

(FWIW, I'm the author of a datetime library in Rust called Jiff.)

[mason55]

> That's part of why they are tied to a certain city -- time zone rules are unlikely to bisect a city, although if they did I guess they'd have to deprecate it as a timezone name and use something else! Not sure if this has ever happened.

It's actually easier to create this problem than by bisecting a city, and the easier way is even more complex than bisecting a city.

You obviously can't put every hamlet, town and village into tzdb, for a lot of reasons. So, if you're trying to represent a time in a place that isn't in tzdb, you have to pick the nearest location that is in tzdb. And it's quite possible that between when you enter your time and when that time comes to pass, the location you were specifying for changes it's rules in a way that's different from the original place you chose.

If you bisect a city, you could create two new names, so that if you encountered the old name you'd know that something needed to be reconciled. But if you chose the nearest place and then your rules changed, you'd have no way to know automatically that it needed to be revisited.

For example, parts of Chile decided not to do DST any more. To support this, a new timezone, America/Punta_Arenas, was added to tzdb. Before this, if you were in Punta Arenas, you would just put all your times as America/Santiago. And now you have no way of knowing if those times are really supposed to be Santiago or if they were Punta Arenas and Santiago was just the best you could do at the time.

Location-based tz's are the best we can do right now but even still they have intractable problems when things change.

[justingrantjg]

Yep. That said, for end users it's a fairly good story because consumer OSs have gotten very good at automatically adjusting clients' time zone based on geo-location. Just like when you fly to another country and the first thing that happens when you connect your laptop to the internet is you're offered to change your time zone to the local time zone. I assume that the same thing happens when you're in a place whose IANA time zone changes like your America/Punta_Arenas case or many others like it.

[jrochkind1]

On re-read, I understand the other aspects of what you mean better. True!

For dates in the past it isn't much of a problem. `America/[city in chile]` in the past (created before the change, refering to times before the change) still has a specific point-in-time meaning even when things change.

The problem is dates records created in the past but referring to times in the future. Which could now be ambiguous or wrong... and this is the first time I'm thinking about it, I'm not sure how easy it is to detect, I guess it should be detectable which dates may be ambiguous/wrong if you know the date of their creation (before the change was known), but it would take caring to write guards about it and having access to databases with sufficient info.

[mason55]

> For dates in the past it isn't much of a problem. `America/[city in chile]` in the past (created before the change, refering to times before the change) still has a specific point-in-time meaning even when things change.

Right - date-times in the past are always easy (at least until you have to take relativity into account). An event happened at some instant in the universe and you just need an agreed upon representation of that instant. UTC works fine for this - record the UTC-based instant at which the event happened and you can always translate it into any other representation without losing information. Recording it in your local timezone is fine too, as long as you also record the UTC offset or timezone along with the instant.

> I guess it should be detectable which dates may be ambiguous/wrong if you know the date of their creation (before the change was known)

Yeah - in theory, when a timezone is added, you could probably link it to timezones that users of the new timezone might have previously used. And then any future times that were saved using that timezone, you ask someone if they are still correct or if the timezone needs to be adjusted to the new one

For example, if a new timezone was added for southeast Colorado, you might ask someone about all times scheduled in both the Denver & Phoenix timezones, because you don't know which one people might have picked.

It gets complicated though because you need to keep track of which entries have been double checked and which ones haven't, and you need to keep track of the version of tzdb that you reconciled against, because there could be another change in the future.

[jrochkind1]

Right, I mean that if in 2030 for some odd reason half of Punta Arenas does DST and half does not, then 'America/Punta_Arenas' would not work as a timezone designator anymore! Obviously unlikely, I probably should not have mentioned it.

[burntsushi]

Can you provide a concrete example? Ideally using Temporal to show where it's going wrong. Like, if you created `2025-06-20T17:00:00+02[Europe/Dublin]` (the instant) in 1760, then its civil representation would be different, to account for the rules in place at the time. And then if you went to deserialize it today, Temporal would reject it. You'd get an error, because the rules in place when the string was serialized don't match the rules in place today for that instant.

To be clear, I might be misunderstanding what you're saying. So that's why I'm asking for a concrete example. That will cut through everything. And if you didn't, I would strongly suggest you take a look at https://tc39.es/proposal-temporal/docs/zoneddatetime.html and search for "conflict". I think that will help explain things.

> I get that it’s more correct

We can chase perfection, but perfection isn't the goal. All models are wrong, but some are useful. In other words, the question isn't whether Temporal's model of interaction with time is wrong (it is!), it's how wrong it is and what its impact is that matters.

[vlovich123]

I must be misunderstanding what you’re saying. How does Temporal know to reject something serialized from before a rule was changed when the creation time of the serialized representation isn’t encoded? You’re saying in 1760 [Europe/Dublin] would be a different string vs today? A more concrete example is normal time-zone rule changes - `2025-06-20T17:00:00+02[Europe/Dublin]` would represent different instants if Dubling passed a new DST rule adjusting the clock back by 15 minutes at 2025-06-20T16:50:00+02[Europe/Dublin] - then the meaning of the instant is different because now there’s 2 different `2025-06-20T17:00:00+02[Europe/Dublin]` and which one you get will depend on when you deserialize.

[burntsushi]

Can you check out the Brazil example in the docs I linked? That really should clear everything up. It explains how a datetime in the future gets serialized, but after a DST change, that serialized datetime becomes invalid.

The way this works is by looking at offsets. Think of a time zone as a function mapping between civil time and physical time. Or, another way to think about it is a mapping from a civil time to an offset (which may not exist for gaps or may not be unique for folds) and a mapping from physical time to an offset (of which there is a bijection). With RFC 9557, you encode both the offset and the time zone into the serialized representation. Consider a case when you encode an offset corresponding to DST, and then DST is abolished. Your tzdb is updated. Then you go to deserialize this datetime. Its offset no longer matches the rules. This can be detected and Temporal reports an error. This is all explained in the docs.

Note that DST is only an example here, because it's a common manifestation of how this error arises. But it can happen with any change in offset.

So if you're in 1760 and you write down a datetime in the future using the rules of that time (of which, I'm not even sure that's a sensible question to ask), then you'd have a particular offset encoded in that future datetime. Now fast forward to the future and the rules have changed and are no longer consistent with the offset encoded in the datetime. Thus, you get an error.

Here's another take on the idea that I wrote for Jiff: https://docs.rs/jiff/latest/jiff/tz/enum.OffsetConflict.html

Think of it like this. In an RFC 9557 timestamp, you have an offset and you have a time zone. Those two pieces of information may be inconsistent with one another. For example, `2025-01-30T17:00+10[US/Eastern]`. When that happens, you can report an error. That's pretty much it.

> then the meaning of the instant is different because now there’s 2 different `2025-06-20T17:00:00+02[Europe/Dublin]` and which one you get will depend on when you deserialize.

If the region adjusted their close back by 15 minutes, then the offset would change. As for `2025-06-20T17:00:00+02[Europe/Dublin]`, it can only feasibly have two different interpretations: you can either interpret it as the instant `2025-06-20T17:00:00+02`, or you can interpret it as the civil time `2025-06-20T17:00:00` in the time zone `Europe/Dublin`, ignoring the offset. Or you can reject it. Temporal supports all of these modes, and it will by default reject such strings when the offset and time zone are inconsistent.

(There's a fourth mode, "prefer offset," that I won't get into here... It's not usually used for deserialization/parsing.)

[Macha]

So in my use cases, there's three types of dates that matter:

1. Past dates. These can be stored UTC, and just rendered in the appropriate timezone as a matter of formatting.

2. Future non-human dates: e.g. the next execution time of a job that runs every hour. These can just be UTC

3. Future human dates: I care about the human selected timezone so that events happen on the wall clock time the user expects. The UTC time and UTC offset are meaningless.

So in cases 1 and 2, having a non-UTC date is not required, while for case 3, the only thing that UTC offset is doing is adding information that could be inconsistent or confusing.

e.g. If the concert is on at 2026-01-31T18:00:00[Europe/Dublin] , that's all that matters, whether that ends up being 2026-01-31T18:00:00+00:00 or 2026-01-31T18:00:00+01:00 is unimportant for whether the building is going to be open at the time. So the system failing to give customers on the day of the concert a countdown because `2026-01-31T18:00:00+00:00[Europe/Dublin]` has become inconsistent because e.g. the EU actually did go ahead and abolish DST is suboptimal.

[burntsushi]

For that specific use case, sure! But Temporal isn't for Macha's 3 use cases. Not all future datetimes are only concerned with civil time. Some are concerned with the precise instant in time. So how do you choose which one? There is no one universal right answer, so IMO, the right default is to reject.

But if you know your use cases and know you always want to adhere to civil time even if it means a change in the precise instant, then Temporal supports that too:

    >> zdt = Temporal.ZonedDateTime.from("2025-06-20T17:00+08[US/Eastern]")
    Uncaught RangeError: Offset +08:00 is invalid for 2025-06-20T17:00:00 in US/Eastern
        InterpretISODateTimeOffset ecmascript.mjs:1467
        ToTemporalZonedDateTime ecmascript.mjs:1531
        from zoneddatetime.mjs:478
        <anonymous> debugger eval code:1
    >> zdt = Temporal.ZonedDateTime.from("2025-06-20T17:00+08[US/Eastern]", {offset: 'ignore'})
    >> zdt.toString()
    "2025-06-20T17:00:00-04:00[US/Eastern]"

> So in cases 1 and 2, having a non-UTC date is not required

If the only operation you need is formatting, then I agree, you can apply the time zone to the instant right before it's displayed. But there are many other operations (such as arithmetic or computing durations between datetimes) you might want to do that do required a time zone. You might still be able to get away with only storing a UTC date, but it really depends on what you're doing.