UTC leap-seconds (again!)

[ChrisBarker-NOAA]

Topic for discussion

From #297, it's clear that there are still confusions / problems with CF's handling (or not handling) of leap seconds in UTC time. I am starting this discussion to, yet again, hash some of this out.

First the non-topic:

In #297 it was proposed that the TAI calendar be added to CF -- that will or will not happen, but does not need to be discussed here. What I would like to discuss is how UTC is handled/talked about in CF.

A few definitions:

Time axis: The abstraction for the passage of time. Relativity aside, the time axis captures the passage of monotonically increasing time.

Timestamp: a human-readable expression of a "point on the time axis", e.g. 2024-03-22Y09:55:32 (in ISO format, but that's a different issue). What a given timestamp means depends on the calendar used.

Calendar: A way to denote passage of time in human-friendly way: e.g. years, months, days, hours. As such, the
calendar defines the set of timestamps (year-month-day-hour-minute-second) that are permitted, and how the different calendars can be mapped to / from each other.

Timedelta: A duration of time, e.g. 23 seconds, 5 years, etc.

Epoch: The starting point for a time specification -- e.g. Unix epoch is January 1st, 1970 at 00:00:00 UTC.

The problem:

(This is all a re-hashing of what I'm sure most readers know, but to put it in one place)

As we all know, the UTC calendar is widely used, and essentially universal, and thus we have no choice but to use it in CF (not to mention years of legacy). UTC also includes "leap seconds", which are used to keep UTC in sync with the sun -- e.g. local noon. NIST discussion of leap seconds This is not so different than leap years, except one critical difference -- there is no way to know when leap seconds may occur in the future. Combine that with the fact that, well, it's only a few seconds, which is a precision that most frequently doesn't matter, and there is very little (none??) software that handles leap seconds properly. And this means that when once does computation with timestamps and timedeltas (e.g. seconds since a timestamp) -- it may be off by some number of seconds from the correct result. (up to 37 seconds, I think, at this writing).

This is a problem for CF because in CF you encode time coordinates in "timedelta since a timestamp" format, e.g. "seconds since 1970-01-01T00:00:00", which requires computing the timedelta(s) to/from timestamps, and with the UTC calendar that can only be done with seconds precision if you include leap-seconds, which most software does not.

What this means is that many (most?) actual files in the wild are not actually correct UTC down to the leap second. For most applications, it really doesn't matter -- if you have a model with a 1hr timestep running for 10 years, second-precision does not matter. If you have data collected by an instrument -- hopefully, you have used an epoch close to the time of the data collection, and thus not crossed a leap second boundary (or only one?)

Another issue is that Computers most often use UTC time, and re-set themselves to UTC periodically (usually using the NTP system. What this means is that if you ask the computer what the UTC time is now, it will usually provide the correct UTC time. it does this without accounting for leap seconds by adjusting the Unix EPOCH so that it results in the correct time now, which would be a bit off for times in the past. (and a bit off right around the leap second itself -- see the RedHat post for details).

All these complications / limitations are out of the hands of CF, of course. All we can do is use the best specifications we can so that used know clearly what they are getting.

The current state of the issue in CF:

• The UTC calendar is used by epochs, but not called out as supported.
• It is recommended to not use months or years as units in the time coordinate, as they are not a well-defined timedelta
• leap seconds are addressed here:

When a time coordinate value is calculated from a date/time, or the reverse, it is assumed that the coordinate value increases by exactly 60 seconds from the start of any minute (identified by year, month, day, hour, minute, all being integers) to the start of the next minute, with no leap seconds, in all CF calendars. This assumption has various consequences when real-world date/times from calendars which do contain leap seconds (such as UTC) are stored in time coordinate variables:

Any date/times between the end of the 60th second of the last minute of one hour and the start of the first second of the next hour cannot be represented by time coordinates e.g. 2016-12-31 23:59:60.5 cannot be represented.

A time coordinate value must not be interpreted as representing a date/time in the excluded range. For instance, 60 seconds after 23:59 means 00:00 on the next day.

A date/time in the excluded range must not be used as a reference date/time e.g. seconds since 2016-12-31 23:59:60 is not a permitted value for units.

It is important to realise that a time coordinate value does not necessarily exactly equal the actual length of the interval of time between the reference date/time and the date/time it represents.

Is this good enough?

I don't think so -- the OP in #297 will probably be satisfied with a TAI calendar, but I think we can do better with handling of UTC. Some thoughts:

@JonathanGregory wrote elsewhere:

The elapsed duration doesn't have primacy in this convention.
...
The time coordinate value is primarily an encoded date-time, not an elapsed duration.

If you are going to use a "timedelta since" encoding, rather than a timestamp encoding, it makes a LOT more sense for the "timedelta" to be, well, an actual duration of time, not a calendar concept. And we've gone halfway there by recommending that years and months, for instance, not be used, as they are defined as a specific duration, NOT a calendar year or month. Granted, that was inherited from UDUNITS, but it makes sense, and it's a lot more consistent with most software, and how people use these data.

Most software (particularly scientific software) works with time in actual time duration units, not calendar times (seconds since, microseconds, etc.)-- e.g. unix time, etc. This is both for computational efficiency (storing time in a regular data type), and because elapsed durations are usually more important than timestamps. And most CF - processing code (at least the ones I deal with) transform the CF time coordinate assuming that the values DO mean an actual elapsed duration, so a timestamp change.

In fact, with UTC and leap seconds, the accuracy of the elapsed duration may not be correct -- but I think that should be considered an error in the encoding, NOT a correct way to encode UTC times!

@JonathanGregory also write elsewhere:

... we could also define a UTC calendar, which would encode UTC as it should be (as you advocate), taking account of leap seconds.

When I first read this, I thought "but CF already uses UTC ?!?!", but then looked again, and indeed, it does not. And yet it does define UTC as the calendar for the epoch, which I think is an massive gap. So yes, we should specify the UTC calendar -- if no software supports it, that's OK, it would allow us to at least be clear that differences from UTC are errors (lack of precision) rather than expected.

A while back (#148, I think -- very long one that) both TAI And UTC was proposed, and IIRC, in the middle somewhere, a "kinda of like
UTC but without leap-seconds" calendar was proposed. I think that should be revived:, not because it's an actual good idea, but because it's what, in reality, many already written, and going to be written, files are encoded with.

It would actually define what the values mean, rather than simply saying "It is important to realize that a time coordinate value does not necessarily exactly equal the actual length of the interval of time between the reference date/time and the date/time it represents." -- without any way of knowing what it represents.

Essentially, the current status-quo is:

Time coordinates may be off by some unknown number of seconds, up to 37 as of this writing -- good luck with that!

Doesn't seem ideal .....

So what how would "kinda-like-utc-but-without-leap-seconds" be defined?

1. The epoch would be UTC. This is important, because computers (and presumably instruments) reset themselves so that "now" is correct UTC, even though they are (often) encoding it as milliseconds since 1970

2. it is not the same as TAI, because TAI is slowly differing from UTC over the years, so this would only be the same as TAI if the epoch was before June 30, 1972.

3. Any time after the epoch would be calculated with the Gregorian calendar, and not using leap seconds.

What this would allow is:

A) a time coordinate value does exactly equal the actual length of the interval of time between the reference date/time and the date/time it represents.

B) a clear way to go to/from the time coordinate and timestamps that would be correct, and the same with data producers and data consumers.

I think this is actually pretty much the state of affairs with most existing data files and processing software -- we would simply be codifying it.

Let the lengthy discussion ensue!

[sethmcg]

Nice summary @ChrisBarker-NOAA! Thanks for getting the ball rolling on this.

I agree that what we are actually doing is encoding time durations, and that the spec should reflect that. I think the correct way to think about it is that time durations are real geophysical quantities, and datatimes are just arbitrary labels that you apply to time durations according to a scheme defined by a calendar and an epoch associated with them.

I also agree with your assessment of the current state of affairs. Our standard calendar effectively is what you've described as "kinda-like-utc-but-without-leap-seconds", yes?

Now, I may be missing something because I haven't had time to think about it deeply, but it seems to me that the only difference between a proper TAI calendar and the standard/kinda-utc calendar is that the epoch is defined with a UTC timestamp rather than a TAI timestamp.

(I had been thinking that standard == TAI, but I hadn't taken UTC epochs into account, and I agree with your point about the need to accommodate that in existing real-world data and instruments.)

So what about this idea: We define a TAI calendar in the very straightforward way we've been discussing. We then declare that the default/standard calendar is a TAI calendar, BUT with the epoch declared in UTC. All you need to do to convert from a UTC epoch to a TAI epoch is count the number of leap seconds up to that epoch, which is a known and fixed quantity that you can easily look up, so that gives us a clear interoperability path between all of our existing calendars and TAI. Going forward, people could start using TAI epochs if they want to make things simpler, but all the existing data with UTC epochs is still easy to handle.

Then, separately, we define a UTC calendar that has all the caveats about leap seconds, and for which we recommend against using a unit longer than seconds. Add some discussion about how some units become variable-length under certain calendars, and now we've clarified and codified the existing situation without needing to make any big changes. Plus, anybody who has data where the time coordinates do properly handle leap seconds can fix it just by changing the calendar to "UTC". To convert from UTC to any other calendar, you need to take leap seconds into account, but anybody who's actually operating in for-real UTC (not just kinda-utc) is probably already dealing with that.

Would that work? Does that match up with both the existing state of things and data we expect to be produced in the future?

[ChrisBarker-NOAA]

the only difference between a proper TAI calendar and the standard/kinda-utc calendar is that the epoch is defined with a UTC timestamp rather than a TAI timestamp.

yes, I think that's it exactly.

Our standard calendar effectively is what you've described as "kinda-like-utc-but-without-leap-seconds", yes?

well, yes, but the difference is that "klubwls" specifically is based on a UTC epoch -- so weird -- epoch has leap seconds, but anything after that doesn't.

time durations are real geophysical quantities, and datatimes are just arbitrary labels that you apply to time durations according to a scheme defined by a calendar and an epoch associated with them.

• Exactly -- though the current spec is kinda the opposite :-(

In either case though, there is an encoding/decoding process, going from a time coordinate to datetimes and back again. And depending on how the file is generated, either one could be the "original" correct, precise, value. As long as the transformation between the two is defined, then BOTH are correct, and it doesn't matter -- so that's the state I'd like to get to.

But if we can't get there (I think we can), I'd rather that we say that the durations exprtessed in the time coordinate are correct, but that the timestamps that result from a decoding may not be exactly correct, if done wrong, rather than the opposite, which is what we have now.

[sethmcg]

I agree completely with everything you've said here.

[JonathanGregory]

Dear @ChrisBarker-NOAA and @sethmcg

Thanks for the discussion. I agree with a lot of what you both say. I probably disagree with you about which calendars I am happy to use, but that's not as important as our agreeing on what calendars CF should support and their exact definitions. I'll make some brief statements here as clearly as I can, and perhaps you can say whether you think they're correct.

• Chris defined "calendar" as a "way to denote passage of time in human-friendly way: e.g. years, months, days, hours, ...". That's true, but I think we should be explicit about what makes one calendar different from another in CF: The calendar defines the set of timestamps (year-month-day-hour-minute-second) that are permitted.

• A time coordinate value can be uniquely converted to a timestamp, and vice-versa, given the time-unit and reference timestamp (which are both in the units attribute) and the calendar. By "encode" I mean conversion from timestamp to time coordinate variable, by "decode" the reverse.

• In all the existing CF calendars, "day", "hour" and "minute" have the lengths that UDUNITS gives them, namely 86400 seconds, 3600 seconds and 60 seconds respectively. They are not SI units, but they are accepted for use with SI, and defined by SI with those lengths (Table 8 of https://www.bipm.org/documents/20126/41483022/SI-Brochure-9-EN.pdf). CF recommends not using "month" and "year" so let's not consider those.

• In all existing CF calendars except the standard default calendar, the time coordinate value, in the given time units, equals the length of the time interval between the reference timestamp and the timestamp encoded by the time coordinate value.

• This last statement is true in the standard calendar as well, except that any leap seconds which occurred between the two timestamps are not counted. That's a possibly more helpful version of the statement which Chris quoted from Sect 4.4.1, and which @claashk asked his question about, It is important to realise that a time coordinate value does not necessarily exactly equal the actual length of the interval of time between the reference date/time and the date/time it represents. It's also important to realise that this statement shouldn't be taken out of context! It refers only to the standard calendar, and only in that particular case where there are intervening leap seconds. If you look up when leap seconds were inserted, you can work out the time interval by adding however many are needed.

• Most data that has been written in CF for with real-world (as opposed to model world) timestamps uses the standard calendar, despite its limitations that (a) the time coordinate is not always exactly equal to the time interval, and (b) the standard calendar does not permit timestamps that refer to leap seconds i.e. 60-61 seconds after the start of the minute. The reason it's commonly used is that it's the encoding that most software uses (including UDUNITS).

• @claashk has a use-case for the TAI calendar. It would not be difficult to introduce it, and we agree it should be done. The TAI calendar allows the same set of timestamps as the standard calendar, but a given timestamp in the TAI calendar and the standard calendar does not refer to the same instant in the real world because the TAI calendar does not contain leap seconds. The same encoding is used in the TAI and the standard calendar (i.e. a given timestamp, unit and reference timestamp convert to the same time coordinate in both calendars), but in the TAI calendar the time coordinate is always equal to the length of the time interval between the timestamp and the reference timestamp.

• If there is a use-case for it, we should also introduce a UTC calendar. Unlike the standard calendar, the UTC calendar would allow timestamps that refer to leap seconds, which the standard calendar does not. Apart from those ones, the two calendars permit the same set of timestamps, and a given timestamp refers to the same instant of time in the two calendars. In the UTC calendar, unlike in the standard calendar but like all other existing CF calendars and the TAI calendar, the time coordinate would always be equal to the length of the time interval between the timestamp and the reference timestamp.

What do you think?

Happy weekend

Jonathan

[JonathanGregory]

An addition to the above list of points:

• As I suggested above, a distinguishing feature of a calendar is the set of timestamps that it permits. I realise there's another distinguishing feature as well, which is that In different calendars, a given timestamp can identify different instants of time, or equivalently, A given instant of time can have different timestamps in different calendars. The Gregorian and Julian calendars have different timestamps for a given day, because they have different sets of possible timestamps, so they're distinguished in both ways. However, the standard and TAI calendar (if introduced) have the same set of possible timestamps, which refer to different instants of time in the two calendars. Model calendars refer to different worlds, but when they're compared, it's assumed that a given timestamp means the same instant in time in two calendars, as far as that's meaningful. [I edited this after posting it, because I'd forgotten about the Julian calendar at first.]

Cheers, Jonathan

[ChrisBarker-NOAA]

Model calendars refer to different worlds, but when they're compared, it's assumed that a given timestamp means the same instant in time in two calendars, as far as that's meaningful

I added some text to the definitions in the initial post -- did I capture this adequately?

[Armin-RS]

On most *nix systems, the "tzdata" package will contain a list of leap seconds, e.g. as "/usr/share/zoneinfo/leapseconds".
NIST publishes a list here: https://www.nist.gov/pml/time-and-frequency-division/time-realization/leap-seconds
The International Earth Rotation and Reference Systems Service (IERS) publishes a list here: https://hpiers.obspm.fr/iers/bul/bulc/Leap_Second.dat

But at least on my Cent OS system, the "udunits2" package does not require the "tzdata" package, which explains why udunits is not aware of leap seconds.

[ChrisBarker-NOAA]

I am not aware of ANY software that handles leap-seconds properly (or at all). Which doesn't mean it doesn't exist, but it does mean it's not commonly available to CF users :-(, so that's a reality that we have to deal with.

I'm still confused about why not -- leap seconds are inherently difficult, as we can't predict when they will occur in the future -- but they are quite tractable (if annoying code to write) for the past and near-future. But none the less -- for our perspective, we need to accommodate the fact that most folks do not have leap-seconds aware software -- which I think is what this conversation is about.

But this does have an impact on whether to add a UTC calendar -- it seems obvious to do so, it's a well-defined and universally used calendar. However, there is a danger: given that most software does not handle it properly , there would be a high likelyhood that folks would specify UTC, because they think that's what they are using, when they really aren't (quite) -- and very few clients would able to use it anyway.

[ChrisBarker-NOAA]

OK -- on to my point:

It is important to realise that a time coordinate value does not necessarily exactly equal the actual length of the interval of time between the reference date/time and the date/time it represents.
...
It refers only to the standard calendar, and only in that particular case where there are intervening leap seconds. If you look up when leap seconds were inserted, you can work out the time interval by adding however many are needed.

This is what I'm trying to wrap my head around -- a time coordinate represents a bunch of points in time on the Time Axis (now defined above). And each point is encoded as being a timedelta since an epoch. What I think is that it should, well, represent exactly what it appears to represent -- how much time has passed since that epoch. And I think that should be the canonical "truth". The challenge comes in when you want to decode the values to timestamps in some calendar (or encode them from timestamps). As Jonathan points out, this is a defined 1:1 process with all calendars except the "standard" one. So that's the only issue at hand. My point is that if:

"... a time coordinate value does not necessarily exactly equal the actual length of the interval of time between the reference date/time and the date/time it represents."

Then that means there was an error in the encoding of coordinate values, rather than an inherent property of the way we are storing the data -- and I think in CF we should express it that way. My point is that the coordinate values should represent specific points along the Time Axis -- how those points might be expressed in a given calendar is up to the decoding software -- the entire idea of a leap second (or a leap year, or ....) is utterly irrelevant to the time coordinate when encoded this way. -- well, not utterly, as you still need to express the epoch somehow.

Anyway, If everyone had leap-second aware software, we could all just use UTC "correctly" and all would be good -- but that is not the case, hence the 'standard" calendar. But I think we can do a bit better than just stating that leap seconds could mess things up.

Thinking out loud now:

How do we end up with these "wonky" time coordinates?

They can come about because of the challenge in how to encode / decode timestamps that are in UTC, without using leap-seconds aware software.

For instance:
An instrument measures something, and gives it a UTC timestamp (from the computer it's hooked up to, which presumably is keeping up with UTC via NTP or something).

1. If that measurement occurs right around a leap-second, the timestamp may be incorrect by up to a second, depending on how the computer is handling leap seconds and NTP -- but CF can't do anything about that, so let's focus on timestamps that are correct UTC.

2. When writing a CF-compliant file, the timestamp has to be encoded into timedelta-since-epoch form. If the software is, like most, leap-second unaware, it will compute the timedelta based on the standard Gregorian calendar, and the chosen epoch. So let's say the timestamp is after a leap-second, and the epoch is before a leap-second. Then the result will be a timedelta that did not include the leap-second. which means not that the timedelta is incorrect, but that the epoch is actually not-quite UTC, as it was back-calculated from a post-leap second. This is what (most?) Unix systems (and other computers?) do -- they adjust the epoch, so that the current time is UTC-correct, and times in the past may not be quite right. So all good.

The challenge (again, thinking out loud -- or thinking on computer) -- is when a time coordinate has a bunch of time-stamped measurements, some from before the leap-second, and some after. If we take it that all the timestamps are correct UTC, then when non-leap-second-aware software encodes them, there will be a discrepancy between the ones before and after the leap second -- this is the source of the "a time coordinate value [that] does not exactly equal the actual length of the interval of time between the reference date/time and the date/time it represents" -- either the ones after the leap-second are second long (or the other way around -- my brain hurts).

But: What CAN be done is to re-create the original timestamps, by decoding the time axis using the Gregorian calendar without leap seconds (i.e. standard software) -- which is the CF "standard" calendar.

OK -- I still think that an uneven time axis is an error, but an unavoidable one, given the software available now, so we do need to talk about it in CF.

Also: I mentioned that we might want a "kinda-like-utc-but-without-leap-seconds" calendar -- I realize now that that was reject back when because that's what the "standard" calendar already is. So no real change / addition required, but the language could be clarified.

More on that later -- it's dinner time for me.

[JonathanGregory]

Dear @ChrisBarker-NOAA

Thanks for your further thoughts. Picking up your last remark, if we can add more words to the convention in order to make it clearer and avoid confusion, that would be very useful. Given how much difficulty we have in discussions of this subject, I expect it would be beneficial to explain things in several ways, in order to suit different readers' expectations or preconceptions.

From my point of view (which I am afraid may baffle or enrage some others, maybe you - sorry about that), it seems that CF standard calendar doesn't result from any mistake. It's a simple way to deal with the reality of the situation, which is (a) that the time convention we mostly use i.e. UTC contains leap seconds, but (b) hardly any of the software we use is aware of them. I agree that if there had been plenty of widely used software which supported leap seconds, we would probably have used a strictly UTC calendar in CF from the outset.

The confusion comes from the syntax we have adopted from UDUNITS, of "time-unit since reference timestamp", because this suggests that the time coordinate is an elapsed time. I would like to suggest that a way to reduce the frustration is instead to begin from this statement:

A time coordinate value represents a date/time i.e. timestamp.

Let's not assume it means anything in the first place. It's just a number which encodes a timestamp. It's useful to store a timestamp as a number because it takes less storage, and numbers can be easily and reliably sorted into monotonic order, as required for coordinates.

The encoding we choose to use is as an elapsed time since a reference timestep. This is a very useful choice, because differences between time coordinates give us time intervals. However, it's tricky to use it in practice for UTC, because of not having convenient software. Therefore for UTC we define a different encoding, which ignores leap seconds. That means the time coordinate is not always exactly equal to the time interval, and timestamps applying to leap seconds can't be encoded.

I'm not saying anything new here, just suggesting a different perspective. Instead of Timestamps should be encoded as elapsed time since a reference, and therefore the CF standard calendar is erroneous, I would describe it by saying Timestamps are encoded as numbers. The encoding used in CF has the benefit that the number equals the elapsed time since the reference, with exceptions that affect the standard calendar only.

Best wishes

Jonathan

[ChrisBarker-NOAA]

From my point of view (which I am afraid may baffle or enrage some others, maybe you - sorry about that), it seems that CF standard calendar doesn't result from any mistake. It's a simple way to deal with the reality of the situation, which is (a) that the time convention we mostly use i.e. UTC contains leap seconds, but (b) hardly any of the software we use is aware of them.

I don't think it's a mistake, but I do think that the result is an "incorrect" encoding (or maybe "limited" [*]) -- but it is incorrect in a defined way, so still useful. I think we can probably find wording the avoids editorial works like "mistake" or "incorrect" :-), and rather describe the situation precisely instead.

The fact is that folks for whom seconds precision matters should be using TAI, not UTC or the "standard" calendar at all.

Hopefully I'll get a chance to suggest some different language soon.

[*] -- limited -- while using UTC for the epoch, it is impossible to encode the point on Time Axis that is a leap-second in UTC. But that time does, in fact exist in the "real" world, so that's kind of a problem :-(.

[JonathanGregory]

Dear Chris

I'd certainly agree that the standard calendar is limited or restricted, because of the timestamps it can't represent. We can work on appropriate words to get the point across. I would like to move the statement about the time coordinate primarily being an encoded timestamp to the start of the section about time coordinates, because that is always true. It's almost always true that it's an elapsed time as well.

I agree with you that for applications where precision to the second is important the TAI calendar is needed, or a true UTC calendar (if someone asks for that). Since we already have a use-case for TAI, does everyone agree we should add it to CF?

Best wishes

Jonathan

[ChrisBarker-NOAA]

Since we already have a use-case for TAI, does everyone agree we should add it to CF?

I think so, and I don't think anyone involved with the recent discussion has said otherwise. I didn't start a PR, 'cause I thought someone that actually used TAI should probably do it :-). Or does it need an issue first?

[davidhassell]

I agree that a TAI calendar is needed. I also wonder if it might be a good idea to also introduce the UTC calendar at the same time, even without a use case (do we know that the remote sensing community wouldn't want this?), because it would help make it clear what the TAI calendar is, since it is defined in relation to UTC.

[larsbarring]

I support having a TAI calendar in CF. We have a clear usecase for it. And, actually, I was thinking the same as David, that a TAI calendar preferably should to be complemented with a UTC calendar ("explicit is better than implicit").

Having said that, and without diminishing my support, I am still curious about how the TAI calendar is intended to be used. I do realize that there is a need to very high precision timekeeping in satellite navigation and control. But as the CF time coordinate is seconds since DateTime reference, and the actual coordinate values are just a sequence of numeric values (of enough precision meeting the needs). And the TAI calendar kicks in only when converting from these numeric values to DateTime timestamps. But when are such DateTime timestamps actually necessary and relevant as opposed to the numeric count of seconds (and not going be confusing for all us humans living in UTC time)? I am just asking out of personal curiosity.

Many thanks,
Lars

[davidhassell]

Hi Lars,

The argument given by @JonathanGregory above is that the logical content in CF is in fact the timestamp, and that "seconds since DateTime reference" is just the encoding we use for sticking it in a file. (Please correct me if I'm wrong.)

[ChrisBarker-NOAA]

I also wonder if it might be a good idea to also introduce the UTC calendar at the same time, even without a use case

In theory, this is the obvious thing to do -- it's widely used, and clearly defined (if undefined for the future). By my concern for including it is that many people think they are using UTC (OK -- are using UTC), but if their software (both on the production and consumption side) doesn't support leap seconds, then they are not really using UTC, and there will be errors in the time axis. And unlike Jonathan's point about the standard calendar -- it would be actual errors.

So providing a UTC calendar would invite misuse -- is that a problem? not sure -- consenting adults and all that, and we can clearly note in the docs that the UTC calendar should only be used with software that handles UTC fully and properly. But if there's not actual use case, maybe we shouldn't provide the temptation.

[ChrisBarker-NOAA]

NOTE: in a five minute Google search, the only library I found that appears to handle leap seconds is this Java libary:

https://github.com/MenoData/Time4J

There's also this:
https://en.cppreference.com/w/cpp/chrono/utc_clock/leap_second_info

Though it's not clear to me if that's embedded in code that can be used to convert to/from timestamps to timedelta-since encoding.

Until / if there's readily available libraries (ideally supported for C, Fortran, Python, Java, others?) available for true UTC time, we probably shouldn't add it to CF.

[*] For intellectual curiosity, I took a quick look at the Python cftime library -- I don't think it would be a huge lift to add leap-seconds support for a proper UTC calendar, but someone would have to write the code, and that would only help a (subset of) python users. Now that I think about it, a small subset -- I think, for instance, that xarray / pandas uses numpy datetime64 internally -- so if it doesn't support leap-seconds, it may not work right -- though I'd have to think on that more.

[taylor13]

Haven't had time to fully follow, but I am curious ... Is it true that

Someone recording the time of sunset today and 50 years ago would assign the same two TIMESTAMPs to these measurements, regardless of calendar. For example in a model with a 360-day calendar, you would assign a certain time stamp for today and the same time-stamp used in the model to identify a day 50*360 days ago. It's only when we compute the interval of elapsed time between two events that we get a difference, which depends on the calendar.

In models with a 360 day calendar, the elapsed time would be more than 200 fewer days, but in comparing models to observations for these dates, you would not care about the elapsed time, only the TIMESTAMPS, and the TIMESTAMPS would allow you to determine which times to compare between the model and the observed. A good model would have the sun exactly in the right place in the sky and set at the right time of day if the same time-stamp samples were compared.

Is this at all helpful or trivially irrelevant?

[JonathanGregory]

I think that is correct, Karl @taylor13. The first aim of CF is to enable comparison of data from different sources. For that aim, when comparing models and observations, you would often regard data with the same timestamp as "comparable". The timestamps themselves are important for comparison of data from different calendars, when they all refer to the real world for the purposes of the comparison. The encoding of the timestamps in CF as elapsed time since a reference is irrelevant for that purpose; it's simply an encoding.

[larsbarring]

Dear Jonathan,

Referring to you comment a few days ago, I am glad that we agree on many points, but as it seems not on all aspects regarding how to interpret/explain and use the CF standard calendar and its timestamps and "numerical encodings". As was the intention with my previous posts I now begins to understand where we might have different views. Your last point begins

• Possibly I disagree with you and Seth about whether the standard calendar is "imprecise", which you mention right at the end. I don't think there's any imprecision.

The calendar as such is of course as precise as a data producer choose when writing the data. What I meant with "imprecise" was more how it is used when a data user wants to combine data from various sources. That is, what I aimed at is more like what you wrote in your response to Karl. However, it is not possible to simply "cut off" part of a datetimestamp from the right to indicate some intrinsic [lack] of precision in the time coordinate of the data. I think that we largely agree on this.

Then you continue:

I think that timestamps in the standard calendar are to be understood as UTC. This is almost implied by the second paragraph of Sect 4.4.1, but it should be stated explicitly. They are UTC because that's the time people mean by default in the real world. I agree that real clocks (on the wall, on my wrist, in your computer) are imprecise, and differ more or less from UTC, but they are all intended to show UTC (with time-zones), and I believe it's UTC that people intend when they store real-world timestamps with the standard calendar.

Yes, I agree. And either any such clock is actually able to "display" the leap second time stamp when one occurs, in which case it is a proper UTC clock, or it will "absorb" the leaps by adjusting the internal tick mechanism (or something else), in which case the normal march of "displayed seconds" will show some kind of irregularity in connection to the leap second indicating that it is not a proper UTC clock. If I again use a table to illustrate where I think the problem is (this time as a screen clip)

To the left are the SI seconds, and to the right are how I interpret your suggestion for the CF timestamp, with the "excluded" leap second, and three alternatives for how it can be encoded as seconds since 1972-06-29 23:00:00. I do not see an alternative to momentarily stretching the second to fill the gap (marked by a ?) . Even though the left column represents the corresponding TAI encoded time, it is equally much just ticking off SI seconds. By removing the gap, i.e. basically deleting the whole row containing the UTC leap second (and thus the gap) we create a jump in the SI second timeline represented by the TAI time.

In an earlier comment you explain this by writing

However, as we have been discussing, the time coordinate is not equal to the length of time between the two instants if there are intervening leap seconds. It says it's 3600 s, but it's actually 3601 s. That is a disadvantage of the standard calendar. I think we ought to be clear that a second means an SI second in CF time units, just as it does in all other CF units.

I think the first of your sentences would be a good starting point for improving the nebulous (at least to me!) current CF sentence "It is important to realise that a time coordinate value does not necessarily exactly equal the actual length of the interval of time between the reference date/time and the date/time it represents.". But I would say, based on the table above, that the fact that the standard calendar ignores (=removes) leap seconds effectively means that momentarily the CF second is not a SI second. To me this is nothing to shy away from, rather a concrete and practical starting for explaining that applications requiring precise time coordinate data should look for the TAI calendar or GPS calendar.

Kind regards,
Lars

[davidhassell]

Hello,

Was there a consensus reached on whether time coordinates represent timestamps, or represent the distance along the time axis from the origin? Or are both interpretations "allowed". You might say that the discussion answers this question, but I would find the answer most useful in interpreting the discussion!

Lars wrote

that the fact that the standard calendar ignores (=removes) leap seconds effectively means that momentarily the CF second is not a SI second.

(Admitting I'm not sure where I am on this, yet) I'm not sure about this - if it were the case, then wouldn't it follow that the CF second is not an SI second in all of the current CF calendars (360_day, etc)? which I don't think we would say.

Cheers,
David

[larsbarring]

Hello David,

I do not think we actually reached a consensus, neither was there an outspoken disagreement .... I would argue that the time stamp is intrinsically linked to the calendar and that the time stamp in one calendar is not the same as the timestamp in another calendar.

See my comment, which only deals with differences in seconds (because of the overall focus of this issue), which often can ignore. But I would not want to equate 2023-06-21 (summer solstice) in the 360_day calendar to the same date in the 366_day calendar. Rather I would do some recalculation of the timestamp based on the length of the calendar year.

With respect to the current topic I would say that either we have a proper UTC calendar capable of representing the leap seconds, or we have something else. This "something else", which is what we are discussing, is not a proper UTC calendar only because it is not capable of representing the leap seconds, which means that since 1972-01-01 the number of datetimestamps is 27 fewer than that of a proper UTC clock. And this "something else" is exactly the CF standard calendar.

[JonathanGregory]

Dear Lars

I'm happy to see that we are largely understanding each other now! That's wonderful. With this exposition, I see what you mean when you say that the standard calendar means a second is not an SI second. This is because UTC has 3601 seconds from 1972-06-29 23:00:00 to 1972-06-30 00:00:00, but the standard calendar says 1972-06-30 00:00:00 is 3600 seconds since 1972-06-29 23:00:00. These can be reconciled if the CF second for this interval is 3601/3600 SI seconds, or if the CF second is normally equal to the SI second, but between 1972-06-29 23:59:59 and 1972-06-30 00:00:00 the CF second equals 2 SI seconds.

I agree that's a possible way of seeing it, but not the way I think of it, which is that we don't have to take seconds since 1972-06-29 23:00:00 as a literal statement of fact. It's just a units string that indicates we encode the timestamp as a number in a certain way.

Alternatively, we could say that second does indeed mean SI second, but it's a count of seconds in an imaginary world which is just like the real world except that UTC does not have leap seconds (because the Earth is rotating at a constant rate, say, so they aren't needed). This interpretation is related to the point of Karl's question. We have GCMs which deal with imaginary worlds just like that.

Let's imagine there is a parallel universe which is exactly the same as this one in all respects except for constant rotation of the Earth in our era, and history and climate are all just the same there as here despite this small difference (which you expect would actually make everything totally different after a while, because of chaos, but never mind). Because everything is the same, it would be sensible for us to compare the events of 29 June 1972 in this Earth and the other Earth, down to the hour or minute, but we neglect the leap second. To make this comparison, we use timestamps. 1972-06-29 23:00:00 and 1972-06-30 00:00:00 refer to comparable instants of time in the history of the two Earths. For convenience, we decide to encode these timestamps as seconds since 1972-06-29 23:00:00 according to the version of UTC in the other Earth, not this one. It's just a convention.

Does that help, or is it too silly?

---

Dear David

I too don't think a consensus has been reached, because I don't agree with Lars's view. I would argue that the same timestamp in two different calendars can sometimes refer to instants of time which we regard as comparable (as in Karl's question, and my fanciful example above). In those situations, timestamps are primary. I think that 2023-06-21 in the 360_day calendar may well be comparable to the same date in the 366_day calendar, because climate models are set up in such a way that the insolation has the same annual cycle as in reality, as far as possible, in order to avoid distorting the annual cycle of climate, and the solstice is a probably a point they try to get right, I guess. Certainly, it's routine and sensible to compare monthly means (of June 1972 or April 2024), say, between two CMIP historical simulations done with models that have different calendars.

Best wishes

Jonathan

[davidhassell]

Certainly, it's routine and sensible to compare monthly means (of June 1972 or April 2024), say, between two CMIP historical simulations done with models that have different calendars.

Indeed so, which suggests to me that the amount of seconds during each of those months is immaterial, and the timestamps are the important thing.

[larsbarring]

??
This might be so for intensive quantities but certainly not for extensive quantities. Even for intensive quantities it might not perfectly valid because of which fraction of annual cycle the month cuts out. For example, for February:

non-leap year standard calendar:
starts 31/365 [=0.0843] and ends (31+28)/365 [=0.1616]

360_day calendar:
starts 30/360 [=0.0833] and ends 60/360 [=0.1667]

Admittedly not a big difference, but we are here discussing principles not approximations. Well, perhaps not principles, but more accurately how to express "things" (concepts and otherwise) as clearly as possible so that users of the convention can apply it in a way that is as correct as possible in relation to their use case, which often means that they have to make judgements regarding how best to do this. To best meet that purpose I do not think it is helpful that we make approximations and simplifications that are not really necessary, at least not without being clear and explicit about it and the reasons behind.

[larsbarring]

A practical example: Here where I live the day length increases about 6 minutes per day or more around the spring equinox. So being just a few days off does make difference if you are comparing data using different calendars and not being careful about getting the timing right in relation to the annual cycle.

Edit: added "per day"

[larsbarring]

Dear Jonathan,

Referring to your recentmost comment I fully agree with your first paragraph.

Regarding the second paragraph it seems that we do not agree. You write

I agree that's a possible way of seeing it, but not the way I think of it, which is that we don't have to take as a literal statement of fact. It's just a units string that indicates we encode the timestamp as a number in a certain way.

I am not sure I understand what you mean here. The second part indicates to me that duration since datetimestamp is SI seconds (because CF uses the SI system). any deviations from this, and the reasons why, should be made explicit. Moreover, I think that the datetimestamp part is something more than just a string; it is a an encoding of the number of time units (seconds, minutes, days, years, ...) since some historic zero point. This encoding is encumbered with many complications such as calendar gaps, uneven month length, variations in Earths's rotation, just to mention a few. But it is possible but for all practical purposes rather meaningless. More useful is however to relate this reference datetimestamp to a more recent "base reference" such as 1972-01-01, 1970-01-01, 1958-01-01, 1900-01-01.

Regarding your third and fourth paragraph, I fully agree in principle. But we are specifically dealing with the "real world" here, where those responsible for time keeping have decided to have leap seconds. And how to deal with these are precisely what we are trying to tease out here.

I would be very happy if we could conclude something like
"the standard calendar exactly corresponds to the UTC date and time, except for when a leap second is inserted in the UTC time. The reason for this is that the standard calendar cannot represent leap seconds. What this means in practice depends on the context and what the datetimestamps represent and how they were generated. For example one interpretation is that the standardcalendar second immediately preceding the leap second is extend to actually cover two seconds, or that the minute immediate preceding the leap second is 1/60 longer than a standard minute, or that the hour immediately preceding the leap second is 1/3600 longer than a standard hour. There are other interpretations depending on for example the hardware used to generate the timestamps [some refs]. For applications where such approximations are not acceptable CF has the TAI calendar...."

Perhaps we could even include one of the tables above to illustrate impact of a leap second on different calendars.

Lars

[larsbarring]

Having thought a bit more about the recent direction the conversation has taken, I do think that we need to refocus on the actual issue at hand.

Clearly leap seconds are not an issue for current models. It would of course be possible to create a 360_day_with_one_in_a_million_random_leap_seconds calendar and implement it in some modelling experiment, but that is for another day to discuss. Likewise, I cannot see that leap seconds would be of any concern when dealing with monthly statistics coming from observations or model experiments (except for purely principal arguments if one would like to go down that road).

However, there are many observational datasets, either coming from operational data collection networks, or from specific field campaign having very high temporal resolution. Many observational system produce data at 1 Hz, and there are ultrasonic anemometers measuring at 100 Hz. While the latter probably require TAI time or GPS time, for the former the standard calendar is a plausible and convenient alternative.

I think that we should focus our conversation on how to explain and how to interpret leap seconds and the standardcalendar in the context of such use cases and needs. In doing so I believe that we will cover the need for explanation that may emerge in other more peripheral use cases.

[davidhassell]

... how to explain and how to interpret leap seconds and the standard calendar ..

Sorry, Lars, but I'm struggling to follow this. The standard calendar does not have leap seconds, so surely there is no such interpretation to be made. Similarly, imagine trying to define the interpretation of leap years in the 365_day calendar - I don't think you can because that calendar has no concept of leap years.

Your example on an extensive quantities is interesting, and to me supports the notion time coordinates represent timestamps. The month lengths are different in the two calendars, yet the monthly means are still comparable. This can only be the case when we regard the time information as timestamps rather than elapsed time.

Thanks,
David

[larsbarring]

The standard calendar does not have leap seconds, so surely there is no such interpretation to be made.

Well, if the interpretation of standardcalendar datetimestamps is that they are UTC datetimestamps, I beg to differ. I believe we all agree that now and then (rarely) there are UTC leap second timestamps, and that the standardcalendar does not have leap second datetimestamps. Moreover, it also clear that the UTC seconds, including the leaps seconds, are SI seconds.

How do we explain this discrepancy and what it means for a dataset that have one second time resolution?

The key question to answer is are all standardcalendar SI seconds, or only almost all?

In the first latter case we can say that all standard calendar timestamps are UTC timestamps, but temporarily give up absolute adherence to SI seconds in the vicinity of UTC leap seconds.

In the latter first case CF stays true to SI seconds and the standard calendar ends up as what I called a "local TAI calendar", which deviate from UTC once there is a leap second. Close, but no cigar.

Similarly, imagine trying to define the interpretation of leap years in the 365_day calendar - I don't think you can because that calendar has no concept of leap years.

Well isn't this exactly what we debating here? We, or at least some of us, want to shoehorn a UTC calendar into the standard calendar?

Anyway, I think we are not making any progress here and instead moving in circles and just rehashing old debates in new words. From my own perspective we are not much concerned with leaps seconds and it is pretty clear for us how to handle all this. Hence I will retire from this debate.

Notwithstanding that, I would still like to express my support for introducing a TAI calendar, and also a GPS calendar if that request reemerges.

Thank you,
Lars

[larsbarring]

Sorry, I turned some crucial sentences the wrong way. Now corrected in the above text.

[JonathanGregory]

@davidhassell and I have submitted a new conventions issue #542 with a proposal that attempts to address the difficulties discussed above and in previous issues. The summary of the proposal is as follows:

This proposal aims to reorganise and clarify the existing text, mostly in section 4.4, about time coordinates, with no change in meaning. It includes a new subsection on leap seconds and their implications for the CF standard calendar, with examples and a diagram, and defines a new use of the units_metadata attribute to remove ambiguity in the interpretation of leap seconds in the standard calendar. It introduces two new CF calendars: utc for UTC with leap seconds properly accounted for, and tai for atomic clock time, used for some satellite data.

We would welcome your comments on the issue. It would be marvellous if we could resolve this (after so many years and tears) in time for CF 1.12!

[JonathanGregory]

Please could anyone who wishes to comment on this proposal do so in issue #542, so that we keep the discussion in one place. Thanks.

[ChrisBarker-NOAA]

Thanks! this is an absolutely fantasic response to a very vexing problem -- congrats!

And clearly a lot of careful work -- kudos!

I put a number of comments on the PR -- but none of them are show-stoppers.

[ChrisBarker-NOAA]

My only real concern is that the UTC calendar is an "attractive nuisance", and there is very little software that handles it properly, and many people use "UTC" imprecisely. But the text is very clear about the leap seconds, so buyer beware, I guess.

[JonathanGregory]

I am glad of your support, @ChrisBarker-NOAA - thanks. I will copy your comments into issue #542, where we've made the proposal.

Please could anyone who wishes to comment on our proposal likewise do so in issue #542, so we keep the discussion in one place. Thanks.

[JonathanGregory]

Issue 542 is now concluded. The change will be in CF 1.12.