Python Enhancement Proposal 495: Local Time Disambiguation | Hacker News Reader

Python Enhancement Proposal 495: Local Time Disambiguation

python.org

51 pointsphilipn35 comments

I'm not sure the rationale behind this. That is, the rationale section of the proposal does a poor just of explaining any case where this is actually a problem.

In every case where I've seen this problem, it's a matter of people either not storing the timezone along with the local time, or not storing in UTC time. A local time with a timezone is a unique time, it does not occur twice. A UTC time additionally does not occur twice. Store a time zone along with the date and time or store in UTC and convert on use.

Note: If there are instances where a second is repeated, it's rare special occurrence that developing a formalized interface for seems like overkill.

Agreed - and more succinctly put than I managed - clearly it's the wrong time for me :-)

I believe the use case is for scheduling systems that need to do something at (say) 1:30am. On every day at 1:30am, please run the accounting job.

In the US, once a year you either have to have some locking system set up to avoid the second run (leave a trace that the job has already been started for the day). I believe the person proposing this thinks that it would make this determination easier.

The problem is that it only solves half the trouble caused by dual timezones. There's another day of the year when 2:30am (in the US) doesn't happen at all. If something is scheduled to occur once per day at 2:30am, then that day it is not going to happen.

There are other workarounds available such as avoiding the magic hours around 2am (in the US). But it seems to be a common problem that everyone seems to keep re-solving.

Storing the timezone with the time doesn't really solve the above issue.

> I believe the use case is for scheduling systems that need to do something at (say) 1:30am. On every day at 1:30am, please run the accounting job.

In the instance where someone wants something to happen at 1:30 AM, if they aren't specific in the specification, then they should expect that it may happen twice or not at all at at certain times of the year. This is an imprecise specification problem, not a problem in representing time in structures that can and do contain timezones. That is, it's a failure of cron, or the user specifying the time, take your pick. What it's not is a failure of dates, times and timezones, which specifically address this problem. Timezones or UTC (which is just timezone offset 0), are what we have to deal with this specific problem.

For example, specifying the originating timezone would disambiguate the time, as would specifying it in UTC, (or automatically converting to and using the UTC equivalent on entry).

> Storing the timezone with the time doesn't really solve the above issue.

It doesn't address the "do this thing at this local time daily" problem when someone chooses a time that has special behavior, but it does address the "do this thing at this offset from UTC daily" which may be the best you can expect when specifying a time for a recurring action and not taking into account timezones. If you want to use local time, you have to deal with either the actually time the job runs possibly shifting slightly throughout the year in some locations, or possibly running twice or not at all.

My real problem with the proposal is that adding a .first() method doesn't solve anything, and really just makes half the problem (it doesn't work if the time doesn't exist), and in a way that's already easily solved, since you can't get a valid result from .first without knowing the timezone already.

Storing it as an offset from UTC has the effect that the task will run at the wrong time 1/4 of the year.

Now, you'll probably claim that I shouldn't expect time to behave this way. But this is exactly the way everybody expect time to behave, thus, no I won't agree with you.

Anyway, yes, that still does not solve the entire problem. Time intervals are really messy.

> Now, you'll probably claim that I shouldn't expect time to behave this way. But this is exactly the way everybody expect time to behave, thus, no I won't agree with you.

You shouldn't expect time to behave that way. Cron-like programs also should try to do the right thing. What's really missing is an up-front understanding of what users are to expect when entering a time into a program that supports recurring time. This is a failure of most programs to account for common misconceptions.

This proposal doesn't really help this, as it only detects the instances where we duplicate time, and only where you already have a time zone applied.

This whole working in UTC and transforming the time 'at the last minute' into the timezone is only the current accepted practise precisely because of this problem of the time changing. Ideally, we should be able to just set tzinfo to the timezone location that we want when we construct the time and then we work with that object. Yes, even if we do arithmetic in local time, it should just work.

As far as I read it this pep should eventually lead to a better pytz api. Currently, we have to use the normalize() and localize() functions in order to handle the ambiguous times that happen twice a year when the clocks change, this is ugly and hard to remember to do. I think that the 'first' flag should eventually allow us (once suitable timezones are created in pytz) to do arithmetic with local times and automatically transform between the summer and winter timezones.

I don't think this is going to fix pytz. There's a difference in philosophy - Python expects a tzinfo object that's intelligent and responds to the datetime that's passed into its calls, while pytz expects to be configured once then provides a fixed offset after that.

Seconds are not repeated AFAIK. Leap seconds are only added and the "seconds" field goes 58, 59, 60, 00, 01...

I also completely second your point. Naive datetime objects should only ever be used to describe UTC. Anything else should have an explicit timezone attached to it. Doing otherwise is asking for trouble.

> Seconds are not repeated AFAIK.

In POSIX time, seconds can repeat[1],

> [POSIX] is neither a linear representation of time nor a true representation of UTC […] The Unix time number increases by exactly 86400 each day […] Observe that when a positive leap second occurs (i.e., when a leap second is inserted) the Unix time numbers repeat themselves.

The upside to this is that days are always 86400 seconds "long": computing the start of the next day is simple. (Computing the length is not so much, and computing to-the-second elapsed time is also harder.)

This is somewhat relevant to Python, as the datetime module "ignores" (which I interpret to mean, "repeats the prior second"; I've never watched to see what really happens) leap seconds.

[1]: https://en.wikipedia.org/wiki/Unix_time#Leap_seconds

> A local time with a timezone is a unique time, it does not occur twice.

If you define timezone as an IANA timezone, this is incorrect: a whole slew of local times repeat during a DST fallback event: you'll have a (1:30 AM (dst=True), America/New_York), and then a (1:30 AM (dst=False), America/New_York); that "dst=True|False" bit is the only difference, and that needs to get stored. If you consider "America/New_York" to be the TZ, then storing that bit on the TZ isn't appropriate, as it depends on a particular timestamp.

If you've ever worked with PyTZ, there's a sort of rule of "just call normalize() always"; otherwise, you'll get funny answers to some introspections on the datetime instance: things like the offset being not what a local would say the offset should be. My understanding is that pytz stores the dst flag on the timezone instance itself; things get funny because the timezone instance is not given a chance to update after arithmetic on the datetime instance.

(Really, I feel like the whole thing would work better if there was a separate class for "instant in time" and a function for, "convert this instant in time to Gregorian year/month/day/etc. in this TZ", which then returned a broken-out-type. (And a reverse, of course, for building "Instant" instances.))

UTC datetime + IANA TZ (if relevant) is the way to go. Alas, not all data is so nice.

Note that the proposal doesn't fix arithmetic. Adding a 60 minute timedelta to 1:30 on the transition day will give the same result (2:30, first=True) whether first=True or first=False.

Really I was thinking of it as a distinct timezone that must be tracked. Whether the DST and non-DST versions label themselves as such, the representation used to track time must distinguish whether DST is active or not to display the correct local time. Really, when I say store local time + timezone, I mean local time plus identifier that gets you to the same unique timezone representation in your medium (python, in this case).

Personally, I just always convert to UTC and store that. It changes the problem from one of data fidelity to display or computation annoyance, and annoyances are easy to reduce or eliminate with tooling.

> Really I was thinking of it as a distinct timezone that must be tracked. Whether the DST and non-DST versions label themselves as such, the representation used to track time must distinguish whether DST is active or not to display the correct local time. Really, when I say store local time + timezone, I mean local time plus identifier that gets you to the same unique timezone representation in your medium (python, in this case).

I guess that's my point: the IANA identifier is a well-known way to serialize a TZ, but doesn't include DST flags because they're not relevant. I think if you wanted to store something like a Python tzinfo object, the easiest way is just storing (local time, offset from UTC); (maybe (local time, offset from UTC, IANA TZ ID), if you want to keep the TZ)

tzinfo's don't really have a defining quality in Python, I've found. You can end up — depending on libraries used — with two tzinfos that both conceptually are "UTC", but don't compare equal…

Now that I've thought about it again, I'm not entirely sure that the DST flag + TZ name by itself is sufficient, mostly in the case of a TZ deciding to change their offset.

> just always convert to UTC and store that. It changes the problem from one of data fidelity to display

The right thing to do, and for the right reasons.

Terrible PEP, I hope it gets rejected. One-off flags like this are hacks that shouldn't be in stdlib. It simply stinks, in an area (time handling) where the stdlib does not really smell like roses already.

Dealing with time adjustments is the OS's job, not userland. If your job has to be scheduled exactly and cannot rely on the OS, and you refuse to deal with UTC, it's your own damn fault and you can always just use a long-running process with timers.

I agree. It's library baggage.

So :

Time in UK is currently UTC +1 (BST) At 2am on 25 OCT we will return to GMT / UTC. It will therefore become 1am, and for the next hour all times will have happened before

The idea is to put a bit flag that says "alreadyseenthistime"

It seems to me this is a solution to the wrong problem.

Store all strings as bytes, assuming UTF-8, store all times as longs assuming UTC

If we convert all python datelines to non-naive (ie embedded with a TZ) then we are forced always to choose an encoding just like in strings. The right encoding is to always assume incoming dates are UTC, to throw error if they are non naive, and to assume that local clocks are set correctly (which we do anyway)

I need to read it more carefully - but it seems the wrong solution

In general any solution to datetimes that doesn't involve the time being in UTC is solving thew wrong problem.

I would disagree. UTC is ambiguous and unpredictable. Computers should really be based on GPST (seconds since 1980-Jan-6 UTC).

If you want to be pedantic about getting rid of unpredictability in your time measurement, use TAI.

That's defined as TAI minus some seconds, just go for TAI. :)

When the clocks change, don't you shift timezone? There aren't duplicate times in BST, we just switch from GMT to BST and then back.

> In these situations, the information displayed on a local clock (or stored in a Python datetime instance) is insufficient to identify a particular moment in time.

Does the datetime instance not store the timezone?

Isn't it about summer and winter time? I'm pretty sure you're still in the same timezone in those cases.

That said, I don't see the reason for this PEP. You should only store a value in the DB that doesn't change like that, and then when displayed you can change it to fit the local time of the user.

In Britian during the summer we use the BST timezone and during winter we use GMT.

But once again this is one of those "assumptions about time that everyone gets wrong" just because we shift timezones in Britian doesn't necessarily mean that people in other countries do the same - they might just change what the timezone offset is and keep the timezone the same.

Absolutely agree with the "one of those assumptions" bit. See http://infiniteundo.com/post/25326999628/falsehoods-programm... for a great list.

(And related, see https://www.mjt.me.uk/posts/falsehoods-programmers-believe-a... for more about addresses.)

Another great resource is the tz database [0]. It includes some great vignettes as well as the time zone rules:

From Larry M. Smith (2006-04-26) re Wisconsin: http://www.legis.state.wi.us/statutes/Stat0175.pdf ... is currently enforced at the 01:00 time of change. Because the local "bar time" in the state corresponds to 02:00, a number of citations are issued for the "sale of class 'B' alcohol after prohibited hours" within the deviated hour of this change every year....

From Winston Churchill (1934-04-28): It is one of the paradoxes of history that we should owe the boon of summer time, which gives every year to the people of this country between 160 and 170 hours more daylight leisure, to a war which plunged Europe into darkness for four years, and shook the foundations of civilization throughout the world.

[0] http://www.iana.org/time-zones

> But once again this is one of those "assumptions about time that everyone gets wrong" just because we shift timezones in Britian doesn't necessarily mean that people in other countries do the same - they might just change what the timezone offset is and keep the timezone the same.

Well, for a proposed change I'd like to see examples of the actual problem caused. The only example I can see is EST and EDT, which seems to be the same as the UK doing GMT->BST. Same for Andorra.

My question wasn't rhetorical, I was actually asking if there were places that don't do this.

Isn't that just a matter of having your library differentiate the timezones slightly,even if the canonical name is identical for the two different timezones (no matter what the country calls them, they really are different)? Additionally, there are ways to reference the timezone in the local time that are guaranteed to be unique. For example, offset from UTC. 1:15 AM UTC-09 is not the same as 1:15 AM UTC-08.

What localities do you know of that just change the hour and do not have a time zone assigned to deal with it? I'm fairly sure the point of time zones it to track all those changes, and a name would be assigned to deal with it.

That's not the way it works in the US at all. We keep the same zone all year long but change the offset and change the name slightly. The zones are geographically based: Eastern, Central, Mountain, and Pacific. They are abbreviated differently depending on whether Daylight Saving is in effect or not, for example Central is either CST (Standard, -06:00) or CDT (Daylight, -05:00).

The important part here is that the API to get a timezone usually doesn't include a Daylight Saving flag. For example pytz would be used as pytz.timezone('US/Central'). There needs to be a process that adjusts the tzinfo object to the proper offset; in pytz this is the localize method, but that's non-standard.

Yes, you could get the information from the timezone, but how would one do that in code?

The only time that one has the time fold is when you turn the clock backward (let's just call that shifting from daylight savings to standard timezone). And this would only affect code which used wall clock time (time as it's read, e.g. 1:30am PDT), and would also only affect code which wanted to run something only once at a time within that fold (e.g. 1:30am ... not on both 1:30am's).

So, using the timezone method, just check to see if your current 1:30am is in your daylight savings timezone. Hurray! You're in the clear. Go ahead and do that thing you wanted to do only on the first 1:30am.

But the next day you're going to run into a problem. The only 1:30am you're going to get is in the standard timezone. So now you have to check for this timezone change only on the day of the change, which is yet another piece of data you have to keep track of. On the day of the change, do this timezone comparison, and on every other day don't worry about it.

When the clock hits your interesting time of 1:30am, just check to see if today is the day of the change, check what the current timezone is, check what the daylight savings time zone is, check to see if those those two values are the same, and now do your thing. Otherwise, just do your thing.

All of the above also ignores that people change times at different times (11pm, 1am, 2am, 3am), some don't change a full hour, and some don't change at all.

The proposal gets rid of all of that convoluted logic in everyone's programs, and instead it provides a single boolean value: is this the second time I've seen this time because of daylight savings shenanigans.

You don't need to do any of that. Just get a unixtime, and move on.

Why are you doing time math in local time?

Simply do all the time math in universal time and be done with it.

> The proposal gets rid of all of that convoluted logic in everyone's programs

Does it? It doesn't cover the scheduling problem the other half of the year when the clocks move the other direction.

> So now you have to check for this timezone change only on the day of the change, which is yet another piece of data you have to keep track of.

If running a job twice is a problem, then why not check that the job has not already been run?

> is this the second time I've seen this time

Is this unambiguous? If it's 2015-10-25-01-30-00 GMT, have I seen that time before? In the UK, yes, in Mali no.

You are always in the same time zone but the zone time changes. BST and GMT are not time zones - time zones are (usually) named by geographic features like large cities and is Europe/London in your case - but zone times. Therefore when entering or leaving daylight saving time you remain in the same time zone but the zone time of your time zone changes.

It would be nice if those two things which have an important but very subtle distinction had names which were not just the reverse of each other. It seems like despite you explaining this, it would be very easy for one to mistake "zone time" for "time zone".

The more I'm exposed to other date/time libraries, the more impressed I become with Apple's date/time library.

https://developer.apple.com/library/mac/documentation/Cocoa/...

Reply on news.ycombinator.com