Timestamps done right | Hacker News Reader

Timestamps done right

getkerf.wordpress.com

64 pointsmollmerx71 comments

When you build that syntax into the language / tokenizer, you're making a bunch of choices for the user.

As many have said elsewhere in the thread, what is `m` supposed to mean, minute or month? If the smallest value you can represent with this special syntax is 1s, you're excluding people who work with sub-second timestamps. Unless, you want them to do like 1m0.0000415s or something like that, but now you have the same problem as you started with (long series of digits being hard to read for humans).

More problems: different cultures have different concepts of a "month" (because they have different calendars). Also, the units that you want to use depend heavily on the application. `1y` is different from `52w` is different from `365d`, but all of those concepts are very similar to users. I expect people to create bugs where these tokens are mixed. When you write `now + 1y` do you mean "this time on this date, one year from now" or do you mean "this instant plus 606024*365 seconds" (which is slightly greater/less, I forget which way)? How do you communicate that to the user? If they can't control which sense ("human" time vs "physical" time) the larger units represent, then those people have to calculate the time offsets by hand.

Just pick nanos or micros since some Epoch (the Unix one is one reasonable choice), use them everywhere internally, and let the user pretty-print them. If you want to support high-energy timeseries where the time scale is smaller, either make your nanos fractional or use femtos/attos/zeptos/plancks.

The ambiguity of "now + 1y" caused an outage for Azure:

https://azure.microsoft.com/en-us/blog/summary-of-windows-az...

> I mean, just try it in Java.

For ages JodaTime actually nailed it, and the Java 8 date API was based off this.

> Not an add on type as in R or Python or Java.

Again, let's talk about the modern version of the language and not act like prior screw ups are the end all for a language.

Also

> 2012.01.01 + 1m1d

How is that more clean than:

> new DateTime(2012, 1, 1).plusMonths(1).plusDays(1)

By my count it's over 30 characters less "clean". The Java syntax obscures some meaning and requires a lot more boilerplate in favor of less magical (and thus complex) syntax.

I agree that this sort of first-class datetime-type representation may not be appropriate for every language, but myself, I find it refreshing and brilliant, and I'd love to see more languages support this sort of syntax instead of overloading strings or using complicated objects or APIs.

It's like comparing the power and ease of using regular expressions in Perl or Ruby versus in Java or Python. In Perl and Ruby, regexes are built-in to the syntax of the language itself. They're a truly first-class type, like strings and integers are in all four languages, and like lists and dicts/hashes/associative arrays are in Perl, Ruby, and Python.

I'd love to see datetime objects promoted to similar first-class native syntax support in this way in more languages. It won't be appropriate everywhere, but in the right language it'd be amazing.

The Java is clean because it's perfectly understandable. I don't have to think about "+ 1m" meaning month, minute, or milli-. Verbose, yes, but the meaning is 100% unambiguous which I think makes it a better API.

> overloading strings

I'm pretty sure writing "+ 1m" is more of an overloading of a string than ".plusMinutes(1)".

> By my count it's over 30 characters less "clean".

If character count were really the ultimate measure of cleanliness, then we'd all be programming pointlessly (https://en.wikipedia.org/wiki/Tacit_programming) and using single-character names for any variables that slipped through. (That's not to say that shorter is never better, but rather that, when it is better, its brevity is not the only reason.)

How is plusMonths(1).plusDays(1) obscure in any way? You could show that to my grandparents and they would be able to guess what it does. "+1m1d" on the other hand they wouldn't have a clue.

Length, in either direction, does not correlate to "clean". Clarity and intent does. Clean Coder (http://www.amazon.ca/Clean-Code-Handbook-Software-Craftsmans...) does a fantastic job talking about this, it's worth picking up if you haven't read it in the past.

"Length, in either direction, does not correlate to "clean". Clarity and intent does."

Exactly. So, the first example starts with a date in a way of representing dates that will register immediately for even a lay person. The developer intends to add time to that date. The example does this with an addition operator then a value with letters representing recognizable units of time. Matter of fact, this was so obvious that I knew what the author was doing before I read the explanation. I'd probably do "min," "sec," "hr," etc to aid intuition, though. Esp avoid confusion on months vs minutes for m.

Then, there's the other example. It appears to create an object. It then calls a method on that clearly adds one month. It also calls a method of that method, that object... idk that language so I don't really know the semantics of what it's doing... to add 1 day.

One is definitely more clear and intuitive than the other. It also has the rare property of being easier to type. Epic win over whatever the other thing is. Not to say the other one was bad: still pretty clear. Just not as much as a straight-forward expression.

In the former, typing '2012.01.01' implicitly creates a date object (probably), and '1m1d' implicitly creates some sort of duration object (probably) and uses an operator to combine them.

The Java was isn't really that different. Just more verbose, but again, I don't mind trading a few key strokes for clarity.

What clarity? That's what I mean. The first example starts with a date then adds values to it. Second does same thing. I use neither language nor do tjme series but still knew tge intent. So all you're getting is extra keystrokes with same clarity.

Length itself does not; but there is an obvious benefit to a syntax that is both short and clear -- when reviewing the code, it takes much less time to figure out what's happening.

Take matrix algebra for example: what takes less time to process (e.g. when debugging code) --

`a = b'*c+d` or

`a = matrixAdd(matrixMult(matrixTranspose(b),c),d)`

or perhaps

`a = b.inverse().times(c).add(d)`? What if there are hundreds of operations like that?

Now, there _might_ be people for whom syntax (3) is the clearest -- for example, these could be people who know some programming, but are not familiar with matrix data or operations. However, their convenience, or that of your grandparents, doesn't really matter: if it's a one-off job for them, figuring it out would take a very small fraction of their time and is not worth optimizing for, and if it's not, extra time to learn the syntax would more than pay for itself when they have to regularly work with it. We don't use words to describe such expressions in print for exactly the same reason! :)

The java "syntax" doesn't matter.

The IDE provides context-sensitive cues as one types so that you don't have the cognitive paper-cut from having to think for even a second if "1m" means "one minute" or "one month".

At a deeper level, chaining method calls to build-up an object is perfectly understandable way to modify something like a date.

Personally, I think there is a more subtle conceptual idea that it glosses over that makes the more explicit Java version more "clean".

> D + 1m1d != D + 1d1m

Mixing time units like days, months, years (where units are intransitive) is, in my opinion, a bad idea.

Should be able to type or sanity check them like anything else. What specific issue are you worried about?

2011-02-28 + 1m1d = 2011-03-29

2011-02-28 + 1d1m = 2011-04-01

(edit: stupid leap-year!)

Why does that happen? I'd think they were equivalent. I don't do time-series or anything to be clear. I would just expect the a + b = b + a principle to apply for the m/d value.

because 1m = [28d,29d,30d,31d] depending on which month you are in. By adding 1d, pre or post, it can alter the definition of 1m.

Sure (assuming that's how a "month" is defined in this context -- I haven't actually read the whole article, so I'm not sure); but why do you find the Java syntax less confusing? They look pretty equivalent, except for one is shorter than the other.

By splitting the addition up into two discrete calls, the ordering is made much more explicit, there's no room for ambiguity.

In C#, the largest part of a TimeSpan is a day, you can't have a TimeSpan of "one month and one day", because adding that to a DateTime would add a different amount of absolute time depending on the DateTime. You can have a TimeSpan of "32 days", and adding that to any DateTime yields a consistent result.

1h4m is 1 hr then 4 min. Order is explicit unless language rules obfuscate it.

Ih ok. Sounds like we can keep first style just make order of operations clear and with rules to catch this. Second syntax woukd appear to require same checks. Again, equivalent intent and problems with more verbosity in 2.

2011-02-28 + 1m1d = (2011-02-28 + 1m) + 1d = 2011-03-28 + 1d = 2011-03-29

2011-02-28 + 1d1m = (2011-02-28 + 1d) + 1m = 2011-03-01 + 1m = 2011-04-01

Also, you have these overflow rules:

2011-03-31 + 1m = 2011-04-30

2011-04-30 - 1m = 2011-03-30

February 2012 has got 29 days.

This is why everyone hates working with time and date functions. :-)

Oh it's been a pleasant trip down memory lane for me let me tell ya. Probably why I haven't done that sort of thing in a long time haha.

Because one is an expression like humans think of it and one is a pile of OOP. The former is more desirable. The data types are also obvious.

How is "1m" more obvious that "plusMonths(1)"? I mean really, the latter is basically and English phrase. I know it's cool to shit all over OOP and Java these days, but if Haskell/Clojure/Whatever had a function (plusMonths dt 1) or something that, you wouldn't call it a pile of FP.

I said elsewhere id add a letter to distinguish months, minutes, etc. Otherwise, you missing larger picture: it's a time series analysis language with an expression adding a value to a date. So the value's letters are a datatype or unit of time. That simple.

They are both about equally clear, but the former syntax will always take less time to process (unless perhaps you are a Java programmer who is familiar with the latter syntax but unfamiliar with the former, in which case it's not a valid comparison anyway). If it's an analysis-oriented package with most of the work done at a REPL, less time to type, too.

IMHO, working on timestamps in code like this is just asking for edge case to bite you in the ass.

The only sane way I've found to work with time is to convert any timestamp into a seconds-since-the-epoch value when doing any internal work and then covert back to the timestamp format for display. As an added bonus your code won't get super messy when you start getting timestamps from different sources that are formatted differently. Everything gets normalized to the internal representation before you do work on it.

You can't perform the operations "add one month" or "add one year" if you're working with seconds since the epoch.

For storing a timestamp, absolutely, you should use an integer-based format, counting discrete somethings since somewhen.

For working with timestamps, you need all the nuance, you need something that can manipulate the different parts of it independent of each other.

And finally, for displaying or reading timestamps, you need all the localization and parsing crap to figure out what "020304" means. Fourth of March 2002? Third of February 2004?

Yes you will need to have the localization logic when you convert the timestamps, but when you store them internally as timestamps you need that localization crap every single place you use them.

Adding 1 month or 1 year to the current date is usually a mistake. Quick question, what do you expect to happen when you code "Jan 31 + 1m"? What do you expect to happen when you code "Feb 29 2016 + 1y"? If you are thinking about doing this, ask yourself if it wouldn't make more sense to define your time by days instead. Jan 31 + 30 days, or Feb 29 + 365.25 days. Of course days are easy to implement on epoch time as well ( time + 30 * SECONDS_PER_DAY ).

> when you store them internally as timestamps you need that localization crap every single place you use them

Which is why you shouldn't do it, which is just what I said. :-)

> Adding 1 month or 1 year to the current date is usually a mistake.

No, I can think of many use-cases where this is useful. For example, what should Siri do if you tell it to "move today's 1 o'clock to the next month" ?

Jan 31 + 1m = Feb 28/29

Feb 29 + 1y = Feb 28

SECONDS_PER_DAY is not a constant, because of leap seconds. (Or it is, and shifting between UTC and UT1 causes them to appear. I forgot. It's messy no matter how you slice it.)

That's one reasonable interpretation, except that it causes problems where Jan 30 + 1m == Jan 31 + 1m and also Feb 28 - 1m != Jan 30 or 31.

Luckily most people can ignore leap seconds, just like pretty much every system time library. Because they don't happen on regular intervals it is impossible to code a fixed rule for dealing with leap seconds so very few things even attempt it.

Time is hard enough to deal with already and few human scale things care about 1 second differences that happen once every 3-5 years or so.

> what should Siri do if you tell it to "move today's 1 o'clock to the next month" ?

This is an interesting question, because there are at least two valid options. If it is the 31st of the month, then maybe you want it to happen on the first of next month. Something a human secretary might intuit. On the other hand, you might mean moving the appointment back a whole month, unless it's the 31st and then you mean to have it a day earlier on the next month...

Like I said, this kind of logic gets you in ambiguous edge case hell in a hurry. FWIW, I don't think Siri even attempts to deal with a request like that.

A better solution is probably to require the person to be a bit more explicit in their request "Siri, move my 1 o'clock to next Tuesday".

Besides, these kinds of interactions aren't impossible with epoch time, they just require more work. One can argue that it would be good to make this sort of thing a little tougher as it will encourage the programmer to think harder about what they are doing and reconsider if it is a good idea.

I agree with the timestamp argument for storage but you might also want to keep the timezone/locale along.

Operation like "1 day from this timestamp" is locale-dependent. And SECONDS_PER_DAY is not a constant.

SECONDS_PER_DAY is constant for human scale activities. People who fret over leap seconds are either overly pedantic about time or are working at high enough precision that they aren't going to be making crude adjustments like "plus 1 month from now".

Besides, I can practically guarantee that the libraries discussed here don't deal with leap seconds. They're going to get it just as wrong.

> SECONDS_PER_DAY is constant for human scale activities.

Except for Daylight savings time, which adds or subtracts a whole hour to the day.

This is just an aside, and not a knock on the article, but Java does have a built-in Timestamp type, [1] and has had it for some time.

Maybe it's not "first-class" (not sure what this means in context?), but it's definitely there and not in a third-party JAR or anything.

However, it's bad for other reasons, the first being that it extends java.util.Date (the Javadoc seems to admit this) and combined with the related java.sql.Date (which also extends java.util.Date) makes for a very confusing API.

For this reason, Oracle recommends just using the new Date APIs, [2] and mapping a SQL TIMESTAMP to LocalDateTime.

1. https://docs.oracle.com/javase/8/docs/api/java/sql/Timestamp...

2. http://www.oracle.com/technetwork/articles/java/jf14-date-ti...

The original Java date and time API was probably one of the worst APIs ever invented, as I describe in my answer on StackOverflow: http://stackoverflow.com/a/1969651

However the new javax.time APIs (created by Stephen Colebourne) are excellent and probably one of the best designed APIs. It's funny what twenty years difference can make :)

It's worth noting that much of the original java.util.Date class was based on the same misfeatures as POSIX C:

1. Leap seconds? Those don't exist, right? 2. Years date from 1900. 3. January is Month 0 (ignoring the fact that there is already a widespread convention of numbering January 1).

Of course, the "fix" of Calendar didn't attempt to fix any of the POSIX-did-it-first problems but instead mostly limited itself to supporting other locales by allowing non-Gregorian calendars.

Totally concur. But date/time is a hard thing to get right!

However, getting it wrong the first time is different than getting wrong again, as you pointed out with the Calendar type. 0 == JANUARY, but most other things are indexed from 1...

Third time's the charm, I guess :) Getting the input from Stephen Colebourne (of Joda-Time fame) was undoubtedly a key point in getting this done.

Numbering months from 0 is obnoxiously common in these time APIs. This is one of the few edge cases where IMHO we should index from 1, because people frequently just use the numeric version for output and it is way too easy to forget to add or subtract 1 from the month value. The memory lost by having an empty 0 spot in the month name array is absolutely not a concern in any sane project.

Why are we not using unix timestamps (seconds since the epoch) for timestamps on everything that doesn't require sub-second precision?

You still need to record the timezone somewhere and you might as well just stick to ISO8601/RFC3339 format then.

One approach is to store numeric timestamps in UTC format. Then if you do have data stored in a different timezone you have to include the timezone with it. That is the default mode of operation for many of the standard date formats.

If you get timestamp data from a system outside of your control though you always have to make sure you know what it means. At least half the time it seems like a date without a timezone isn't UTC, but in whatever the originating timezone was but the developers didn't include a timezone offset in the data...Timezones....the bane of my existence.

That approach has limits, when DST raises its ugly head (for example, in a calendar scheduler). It's interesting that some of these systems ask you for timezones like "GMT London" - it appears to the user that they're asking for the three letter timezone, 'GMT' - which the user probably knows - but what they're really asking is the TZDB timezone location - which the user probably doesn't.

No I don't. UTC.

They are difficult to read for humans. Without using a computer can you tell me when this timestamp was taken? 1453383978

Ultimately that was the point of the article, but the tone of the article about how they got it right and everyone else is wrong bugged me.

My approach has been to store everything in epoch form, do all of my calculations and manipulations from there, then build tools that make converting back to a human readable representation when and where it is needed. I think the idea that you have a problem completely solved though just prevents the search for any improvements.

Skimming the article doesn't make it clear whether they've addressed this, but, if you want to add units of time of non-constant size, then you cannot just use time-stamps. For example, `(today + 1 year) - today` is longer than `(today + 2 years) - (today + 1 year)` [0], so that one can't think of them as just `(today.unix + seconds_in_year).human` and `(today.unix + 2*seconds_in_year).human`.

[0] Assuming that today + 1 year is defined to be 2017-01-21, and today + 2 years is defined to be 2018-01-21; and, if not, then one faces other problems with intuition.

Ugh. Every time I see time code, I'm thankful I work with monotonic subsecond times.

Who cares how pretty the internal representation is? Only the processor sees that, you output pretty printed versions of the timestamp as you mentioned. You can even add the necessary bits to your debugger to automatically convert timestamps if necessary.

Dealing with sub-second precision isn't that hard. You just need either a second value that holds the microseconds/nanoseconds or 64 bit time that counts nanoseconds since the epoch instead of seconds.

We're getting into domains where sub-second precision is pretty important. Logs, for instance; on s aystem handling thousands of messages a second, storing log entries at a resolution of one second gets . . . irritating.

I like Windows NT's system of seconds-since-1600-or-so in increments of 100 nanoseconds. That system has served me well. 100ns isn't crazy to read directly and often from a hardware register (talk to a hardware engineer about latching a rapid counter sometime...), it covers a reasonable range for humans (handles everyone currently alive, for their whole lifetimes and most events they care about) and it fits pretty well with events that occur in multiprocessor systems (I'll be saying something different if I ever work on systems with terahertz clock rates, though).

I work with timestamps A LOT. In fact, I often have to deal with streams of 5 million + /second. Even worse, for most of them, microsecond precision is critical.

I'm convinced most languages do timestamps wrong. Specifically, they separate out the "time" component from the "date" component.

What's a more common operation? Counting how many events happened in a span of time? Or shifting every timestamp by 15 days?

Timestamps should generally be designed for extremely fast and lightweight comparison, but keep enough information that a shift is doable. From my experience, all you need is: a unix timestamp, a microsecond (or nanosecond) offset, and the source timezone.

In this case, if you want to find elapsed time between two timestamps you simply subtract the unix timestamps and offsets. Very CPU friendly and easily vectorizable. You can do this even if the timestamps originated at different timezones (since everything is UTC under the hood).

What if you want to shift the date? Or group by date? Turns out computing the date on the fly is a very cheap operation. Easily can do hundreds of millions/second on a single high-end server core.

an use whatever calendar system floats your boat.

Any timestamp system that relies on year/month/day semantics is rarely going to be optimized for the most common operations users do with timestamps. Even worse, for simple comparison you run into all the weird edge-cases that you wouldn't have cared about if you stuck with a unix timestamp under the hood.

I'm surprised the author is so big on kerf but doesn't mention q/kdb+[1] at all, which kerf took almost all these ideas from and which has a much bigger user base.

1) https://kx.com/

This should enlighten you why: https://scottlocklin.wordpress.com/2015/12/15/an-introductio...

Basically, he's big on Kerf because he's involved in the development. The above blog post will tie it with Kx and other APLs.

https://getkerf.wordpress.com/ is the official Kerf blog. And Kerf is not meant to be free...

It's definitely an improvement in reader comprehension. I'll admit one of the reason I didn't learn those languages is that they looked like gibberish.

Doing timestamps right: use a library or class to abstract away the way the data is stored from the intent of the actions you want to perform on them.

What the article sez: GET KERF IT'S THE BEST AND ONLY WAY TO SOLVE THIS

  Numericaly sortable date [YYYY+][mm][dd].[HH][MM][SS],[NNNNNNNNN]
  date -u +%Y%m%d.%H%M%S,%N
  20160122.044145,052215000

  Absolute dates with nanos as seconds since epoch
  date -u +%s.%N
  1453437715.409682000

  Hyphenated,"chunks", with nanos and TZ offset.
  date -uIns
  2016-01-22T04:41:03,121501000+0000

This is interesting, and I can see that adding 1d10m is easier than adding 24 * 60 * 60000 + 10 * 60000 to a (millisecond) timestamp. Though I have to guess m is minute, not millisecond or month.

But...the display and storage of a timezone is a datetime, and it doesn't appear to have a timezone attached. Meaning I'd still rather just store/retrieve/work with millis since epoch, since that avoids any ambiguity about what timezone the timestamp takes place in. With just a datetime...I can ~assume~ it's GMT, but...is it? Millis since epoch are the same across all timezones, there is no ambiguity; datetimes vary, and I have to make assumptions, and make sure my library/driver/etc handles conversions correctly.

I wonder if a closed source language for data processing tasks can thrive on the long run. The concept for timestamps is nice though. But I do not want to spend time of learning proprietary languages.

I think that's a very interesting question, to which the answer probably depends on how you define 'thrive'.

It seems unlikely to me that any language could become mainstream without being open source. The expectation that a compiler or interpreter should have its source available is only growing.

At the same time, kx's kdb+ is an example of a product that sits in a niche and has generated significant revenue despite being closed source. kdb+ has achieved this by primarily targeting financial services, which is a sector that's less sensitive to closed source than most, and is able to spend money on whichever product solves their problem.

If kerf manages to gain an edge over its competitors for a particular set of problems then sure, it can thrive in the long run the way kdb+ has.

It could change in the future, but there are a plenty of languages that got pretty mainstream while closed source -- in terms of working with numbers and data, I'd definitely call matlab and sas mainstream, stata and mathematica are nearly mainstream, R of course is open source which is a huge reason for its popularity, but it's a clone of closed-source s+ which was fairly popular already. In the general-purpose world, c# and swift got most of their traction while still closed source, etc. Surely there are more examples in other domains... But yes, there is definitely a very welcome trend to open source stuff.

This post doesn't load at all for me without javascript.

I had a related comment earlier this week:

https://news.ycombinator.com/item?id=10930663

And it does not even manage to keep its formatting intact if you resize text. Why JS then?!

The post is quite readable in Firefox w/o Javascript once you know the secret, which is

   View -> Page Style -> No Style

Regarding timestamps: https://howtologatimestamp.info/

How's your treatment of timezones?

  date -u +'%FT%H:%M:%S %Z'

    date -u +%FT%TZ

    date -uIs

Shorter but sadly not RFC3339 compliant although it is ISO8601 compliant.

Reply on news.ycombinator.com