Please reconsider the Boolean evaluation of midnight
mail.python.org
mail.python.org
The ticket was reconsidered, reopened and classified as a bug. http://bugs.python.org/msg212771
Nick Coghlan's dissection of the issue here: https://mail.python.org/pipermail/python-ideas/2014-March/02... is pretty much perfect - wonderful piece of technical writing!
Donald Stufft has expressed an interest in making the patch for this happen, and assuming all goes as planned this usage will raise a deprecation warning in 3.5 and be fully fixed in 3.6.
News in brief: User raises issue. Issue gets resolved.
I'll make an equally aggressive counter point: Dammit guys and gals, stop using 2.x. Python 3 came out more than 6 years ago. You are holding back and splitting the community. And if you tell me 3.x has no perceived value then this is the value that is slowly stacking up.
That philosophy, taken to its logical conclusion, results in everything being broken forever.
if var:
...
if var is not None:
...Testing for "not None" also immediately tells other devs at least something about what's going on. Example: If var is an argument to a function and you see "if var," really understanding what's going on means looking at all invocations of that function to see what's being passed in.
Further, in that scenario, it's possible that different types are being passed for that argument. While "if var" could make this work where "if var is not None" would break functionality, in my opinion, it should break. Without getting into arguments about static typing, "truthiness" allows both sloppy code and the accidental introduction of bugs that might otherwise be caught during development due to faulty logic from vague conditional checks.
I suppose the hilarious way to argue it would be to say that there is an implicit "all other things being equal" in front of each of those.
All other things being equal, I see no reason to write code that needlessly introduces potential bugs and maintainability issues by ignoring the rule of thumb that explicit trumps implicit.
The documentation is pretty clear about it:
http://docs.python.org/release/3.3.4/library/stdtypes.html#t...
Which leads to this:
>>> a=object()
>>> bool(a)
True
>>> if var in (None, 0, "", [], {}, ...):
pass
In general, most of those false-y values make sense. Midnight being a false-y value does not make sense. This seems to be a case of "we represent midnight as zero internally, and zero is false-y, so midnight should be false-y." This logic does not make sense to me. If they represented noon as 0 instead, should that evaluate to false, just because?
I agree that midnight evaluating to false is ridiculous (I wasn't aware of this, thankfully I've never run into it). I'm not sure where you got the idea that I support this, but I don't.
def __bool__(self):
if self.second or self.microsecond:
return True
offset = self.utcoffset() or timedelta(0)
return timedelta(hours=self.hour, minutes=self.minute) != offset
Now that I look at the actual code, it makes less sense. It's not just midnight that evaluates to False. It's midnight UTC that evals to False. So if your timezone is PST, then 8AM evals to False. Since most datetime.time() objects are timezone 'dumb' by default, this distinction doesn't matter so much, but I can see this still biting some people hard, even if "Midnight is false-y" made sense.
Edit: Also, to the "explicit is better than implicit" crowd, notice this line (in core libs):
if self.second or self.microsecond:
return True
if self.second != 0 or self.microsecond != 0:
return True
Edit: s/EST/PST/ ^^;;
http://en.wikipedia.org/wiki/Principle_of_least_astonishment
self foo isNil ifTrue: [self defaultFoo]
later some implementations added a method on object to test for nil
self foo ifNil: [self defaultFoo]
if you had to test for nil before testing for a condition it would be a bit ugly.
self foo ifNil: [((self foo) isFooish) ifTrue: [self defaultFoo]]
In ruby which was developed about a bit later perhaps to late to influence pythons bool handling they got rid of this problem by treating either nil or false as Falsey and everything else as truthy.
In python everything is true except:
None
False
zero of any numeric type, for example, 0, 0L, 0.0, 0j.
any empty sequence, for example, '', (), [].
any empty mapping, for example, {}.
instances of user-defined classes, if the class defines a __nonzero__() or __len__() method, when that method returns the integer zero or bool value False. [1]
The upside to python behaving like this is that in many situations you don't care what kind of falsey value you may have and you can write
if person and person.name: instead of if person != None and person.name != None and person.name != "":
and
if students and "John" in students and students[John"].age and students[John"].friends: print "John has friends" instead of if students != None and "John" in students and students[John"].age != None and len(students[John"].friends) > 0: print "John has friends"
There are downsides to this and people may have a strong personal preference for personal scripts and you can argue that the ruby way or even the smalltalk way is conceptually nicer(and more "explicit") but this is proper pythonic style and generally you should use it in python(maybe not the numerical one except when getting len of a collection)(unless you specifically depend on different code for different falsey values).
Explicit is better than implicit.
if var is not None:
...This has now proven to be false, and python is starting to add optional type annotations. It's not practical to introduce static typing more quickly - it would break too many existing programs.
The latter seems a lot less common and surprising.
You could argue that the python is badly designed in this case and that the two should be equivalent, but that's a different argument.
if var:
...
However, it's harmful in one of its most common use cases: default values.
For example, let's say you're working with a function that takes an optional argument similar to:
def foo(x, values=None):
if values is None:
values = []
Let's say a user mistakenly passes "values=0" instead of "values=[0]".
If you had done:
if not values:
values = []
You can make the (very reasonable) argument that this is all the fault of dynamic typing, and if python was just a staticly typed language, the compiler would catch all of this, but that's beside the point.
Be aware of what you're testing if you choose "if var:" instead of "if var is not None:"
I didn't argue, and certainly didn't mean to imply, that it is always better, but in the sort of case we're considering in this example, i.e. we expect a valid time value or None, It really is madness to imply a time value of midnight is "empty".
(For certain types, they do have the same behavior. In those cases, the subjective question is fine. But times are not a case where the behavior is the same.)
From PEP-8: http://legacy.python.org/dev/peps/pep-0008/
The real problem with your argument is that the poorly-styled code of people who haven't learned better yet is in production. This isn't a theoretical exercise. This is a tool used in industry, and "teaching people to style their code better" is not a valid excuse for costing industrial users of the language money.
Further, Python markets itself as a beginner-friendly language. Your attitude is the exact opposite of beginner-friendly.
Industry must bend before truth, not bend truth over itself...
Uhm maybe I'm confused but isn't that the whole design philosophy behind python? It makes bad indentation a syntax error!
No, its syntax for determining block structure is indentation. The designers didn't say "Hey lets complain about bad indentation" they said "Hey, people should be indenting their code properly anyway, so why not make that the syntax?"
Also, they didn't introduce it to an existing project, it was there from the beginning (AFAIK).
If they wanted to make people write better code they would throw an error on coercion from time to bool so that everyone would have to fix the code.
I can't imagine a single application where you would need to preserve the value of hashed format strings. But specifying their value so that you can rely on them just seems to be a bad idea.
The case you might have to make to upgrade the version of the tool you are using has to take into account the risks, and something that is maintaining a high level of backwards compatibility has a much lower level of risk.
I generally eventually have to re-write to use a totally different tool for either of the middle two cases, because they just can't keep up. I would much rather make changes to my use of them based on deprecations than be forced to ditch them entirely.
(This is more a general statement - obviously refusing to make this datetime change isn't realistically going to push anybody away from Python.)
So? "The only valid type for a conditional is boolean" still works if types only apply to values. Under that principle, anything but True or False value encountered in evaluating the condition of an if statement ought to throw a TypeError, not be evaluated for truthiness. If you are expecting something else, call an explicit, use-case-appropriate function to get the right in-context truth value.
(Note, I'm not saying Python should do this, I'm explaining how the logic applies to Python without any contradiction to the "variables don't have types, values do" principle.)
> IMO every value except boolean false (and, if you insist, nil/null) should be truthy in a dynamic language.
I think that's better than what Python does (Ruby does that by default, with nil included as false, though it is IIRC possible-but-extremely-strongly-discouraged to override the default truthiness of objects so you could have classes with falsey values), and I lean toward preferring that approach, but the idea that a dynamic language would do well to just allow True and False as the only valid (non-error-producing) values for an "if" statement is not, IMO, without some merit.
Having a simple universal truthiness rule seems preferable and more in the spirit of a dynamic language.
It really depends on what language you're using. This is more or less true in Python, Ruby, and Javascript. In Lisp generic methods you can specify the type for the input variables and be guaranteed that if you are inside the method the parameters are of that specific type. For optimization reasons you can also declare to the compiler that variables are a specific type. This is really important when doing numeric computations in a tight loop. I've had 50% - 80% performance improvement by declaring types (amounting to hours of run time). I think Groovy and Clojure also allow this but I don't know much about those two.
There's also the maintenance issue though. After a few years of maintaining a fairly large Lisp project I've become pretty convinced that the only way to stay sane, at least for me, is to treat it as a statically typed language. I have asserts and check-type macros all over the place. If variables may have multiple types they can be checked against '(or type1 type2) but things do become more complicated then.
First, True and False were just the integers 1 and 0. Then they were given special values that numerically evaluated to 1 and 0, but __str__ now emitted 'True' or 'False'. They were still global variables, though, so they could be reassigned ('True, False = False, True') and cost a dictionary lookup to use. Finally, with Python 3, they (along with None) were made keywords so that such shenanigans could be stopped.
I started on the side of the original person but was persuaded round by the responder.
"if x:" means something completely different than "if x is not None", adn that needs to be understood. The fact that they might evaluate to the same result a lot of the time is luck, and is how people have got into this mess. Changing the boolean evaluation in this case is actually making the mess even worse.
I don't see how that's possible. Assuming that using "if x:" is wrong (for some value of wrong) here, this issue only affects people who are using it wrong anyway. So all fixing it does is make the consequences of doing bad stuff less painful. I can't see how that would hurt things. Even if you think "if x:" should never be used on non-booleans, leaving land mines in there to hurt newbie (or lazy) developers is not a good way to enforce that convention. It just creates pain.
Or more simply: whether having boolean coercion for conditionals is a good idea is orthogonal to how such a coercion should work. It sounds like Python should deprecate truthiness altogether. But it hasn't, so in the meantime truthiness should work in a reasonable way, and falsy midnights are not reasonable.
Pointing out that it's documented is unhelpful (the documentation is just the wrongness restated in a different language), as is "That test does not mean what you are intending it to mean" (Tautological. The OP was suggesting changing what the test means).
Any object can be tested for truth value, for use in an if or while
condition or as operand of the Boolean operations below.
The following values are considered false:
None
False
zero of any numeric type, for example, 0, 0L, 0.0, 0j.
any empty sequence, for example, '', (), [].
any empty mapping, for example, {}.
instances of user-defined classes, if the class defines a __nonzero__()
or __len__() method, when that method returns the integer zero or bool value False. [1]
All other values are considered true — so objects of many types are always true.
But it just goes entirely against the spirit of that documentation to do so. I would be fairly shocked if there are many other examples of exceptions to this rule in the standard library. That's just not how it is supposed to work. I'm primarily a python developer, and I have that list of False things very deeply internalized. They are False, other things are True. I'm sure most others devs have as well.
It is right at the top of the page on the documentation of the standard types.
I'm sure there could other classes where the truthiness had some obvious physical meaning, and objects could be either True or False in a meaninful way. But midnight is not one of those.
The issue I take with this behavior, or at least the justification for closing discussion of whether or not it's applicable (though I do agree it's not a "bug" per se, simply on the merit that it's documented--even if the documentation is unclear), is this notion that it would break code where this behavior is relied upon. First, you're assuming that someone understands the behavior clearly enough to exploit it (while there's obviously a non-zero population who aren't aware enough to avoid it). Second, you're assuming that they're not inclined to realize what an outrageously stupid idea it is to rely on midnight == False. If someone who's attempting to use truthiness as a means of determining if a time object (say, from a database) is None or falsey isn't aware of this behavior enough to not get bitten by it, what makes us think that the percentage of people who would be exploiting this behavior is somewhat higher? It's insanity. More so when each of the examples presented in the discussion for how it might be used are outrageous edge cases.
Now, would it be possible to workaround this by using datetimes, which as far as I can tell cannot be made falsey, then extract the time component later when needed after validating that the datetime is indeed not None? I can't think of any real world circumstances where you're not going to need to be aware of the date, timezone, and therefore DST when extracting times except for profiling or one-off quicky applications that don't need TZ awareness.
in Boolean contexts, a time object is considered to be true if and only if, after converting it to minutes and subtracting utcoffset() (or 0 if that’s None), the result is non-zero.
My initial reaction was to think "How the heck would you deduce midnight from this?" and I think arguing that the behavior is fully documented from this line alone is a bit of a stretch. It's true that it is documented, but not in a manner that is immediately obvious. Worse, there are 87 instances of "None" on that page, which makes searching for a specific issue somewhat daunting.
The irony (in terms of an unexpected side effect) is that by having the discussion they did, even if nothing comes of it in terms of fixes or changes, future developers bit by this behavior will be able to readily find it via a search for midnight, None values from time objects, etc.
[1] http://docs.python.org/3.4/library/datetime.html#time-object...
Edit: Left off the citation. Sorry. ;)
Which doesn't mean I disagree with changing the behaviour, but I doubt Linus would use that reason for that.
I never saw Linus complaining about changing undocummented and not used behaviour.
Midnight UTC, timezone EST: True
Midnight UTC, timezone CET: False
Midnight UTC, timezone CST: True
Midnight UTC, timezone WIT: False
Midnight UTC, timezone PST: True
If there is a use case for this behavior, I sure can't see it. Unless you like bugs appearing when you change timezones.
Language interpreters can be annoying, but not to the same degree. It is quite easy to run multiple Python versions on one system. It's also relatively easy to find and fix the broken Python code in a case like this -- it's always sitting there on your drive, open for inspection. And it's not black magic -- anyone with a little programming experience, even if it's not in Python, can easily understand the issue and the fix, and make the appropriate change.
Nuance matters, and unlike a boolean, principles are not always binary choices.
I will give you an example that bit my project hard, caused hours upon hours of wasted time for bugfixes, broke our software, etc. and yet was a good thing: PostgreSQL 8.3's removal of most implicit type casts. Prior to PostgreSQL 8.3, the following query would be valid and evaluate to false:
select '2014-01-01' = 2014-01-01;
SELECT '2014-01-01'::text = (2014 - 1 - 1)::text;
SELECT '2014-01-01'::text = '2012'::text
Now, Linux exists in an ecosystem of mostly source-compatible operating systems. People write portable software that they expect to run on FreeBSD, Linux, and so forth with minimal tweaking. I am not sure "we don't break user space ever" would be a compelling argument against fixing non-POSIX-compliant but existing and documented behavior. But this scenario (and frankly the stupidity in the midnight type handling in Python) is what pre-review is supposed to help prevent in the first place.
I think it's more likely just a bad habit that some people may have picked up when using PHP or JavaScript, and mistakenly continued using after moving to Python.
PEP 8 (http://legacy.python.org/dev/peps/pep-0008/#programming-reco...) is quite clear about this:
"Comparisons to singletons like None should always be done with is or is not, never the equality operators.
Also, beware of writing if x when you really mean if x is not None -- e.g. when testing whether a variable or argument that defaults to None was set to some other value. The other value might have a type (such as a container) that could be false in a boolean context!"
PEP 8 is more representative of the actual Python community's views than the mistakes of some programmers who are more accustomed to certain languages that are not Python.
I feel zero value of non abelian group should not mean
False in bool context.
() + () == ()
"" + "" == ""
0 + 0 == 0
timedelta() + timedelta() == timedelta()
time() + time() => TypeError
[] + [] == []
[1,2] + [3] != [3] + [1,2]edit: last Tuesday, 5:00pm + Christmas two years ago at noon = ???
if x:
<destructure x, recurse on subparts>
else:
return <something>
>>> {} + {}
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
TypeError: unsupported operand type(s) for +: 'dict' and 'dict'
>>> bool({})
False
Those who think it this argument is pursuasive, should bool({}) raise an exception? Or should {1:"A"}+{1:"B"} not raise an exception and instead return ... what?
Or do I misunderstand something?
{1: "A"} + {}
What about
{1: "A"} + {} + {1: "B", 2: "Z"} + {1: "C"} + {}
Guido chose the current behavior because there is no clear-cut winner. "Refuse the temptation to guess."
If {1:"A"}+{} == {1:"A"} then the type of each value is unchanged from the input list.
If {1:"A"}+{1:"B"} == {1:["A","B"]} then the type of each value (or at least each shared value) is a 2-element list.
If both are true, then no one can write sane code which knows how deal with the sum of two dictionaries.
That is, the insane code has to have checks for if one of the dictionaries is empty, and if so, do one thing, otherwise do another thing. In which case, why is this definition useful?
Therefore, either {} is the zero value, or dictionary summation produces lists for the resulting values, but not both.
{1: "A"} + {} == {1: "A"}
{1: "A"} + {} + {1: "B", 2: "Z"} + {1: "C"} + {}
{1: ("A" + "B" + "C"), 2: "Z"}
That's how it usually works with functions.
>>> from collections import Counter
>>> Counter({1: "a"}) + Counter({1: "b"})
Counter({1: 'ab'})
Lance Ellinghouse:
why can't I add dictionaries like I can other objects?
...
if a and b both had the same keys, then I think b should override
the keys in a.
Guido responded in http://ftp.ntua.gr/mirror/python/search/hypermail/python-199... saying that he could think of no reason to prefer replacement over keep behavior.
In my own early Python code, I sometimes used + when I wanted an update semantic, so I can confirm that your expectations aren't universal.
And that's what this comes down to - is bool(time(0, 0))==False the generally expected behavior (which is a higher threshold than reasonable behavior)? What are the likely errors? Do the advantages outweigh the disadvantages?
That's why an argument about the zero element is only somewhat relevant - most programmers aren't mathematicians, and practicality beats purity.
>>> a = Counter()
>>> a[1] = a
>>> b = Counter()
>>> b[1] = b
>>> a + b
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "...../collections.py", line 536, in __add__
newcount = self[elem] + other[elem]
File "...../collections.py", line 534, in __add__
result = Counter()
RuntimeError: maximum recursion depth exceeded
That said, this is a good example of the expressivity of this, which mathematical definitions often implicitly disallow.
For example, NUL terminated strings lead to a lot of user errors, some of which lead to security holes. Other string representations don't have that flaw, though come with a different cost concern. Is a buffer overflow "user error"? Some would say it is. But the language design makes those errors more dangerous.
Or in Python, there's no technical reason to have a ":" at the end of the line before an indented block. Instead, it's there because user studies show that people learning ABC (which influenced a lot of early Python) made fewer mistakes if the ":" was present than if it wasn't.
Are indentation mistakes user error? Certainly users play a role. But again, the design does as well.
So saying that something is a "user error" with no further analysis absolves the designer of any responsibility, and I disagree with that idea.
There are at least 4 different ways to handle dict+dict in Python. At least three have come up in this thread as the proposed correct solution. Even if there is a mathematically clean solution, if only 10% of the people expect it to work that way, then why should Python introduce something which is so error prone? No support for dict+dict is 100% error prone, of course, but trivially identified in testing. While Counter+Counter-like behavior has subtle consequences that will trip people up.
(As another example of the subtleties, consider an inverted index mapping word to a list of document ids:
collection_a = {"a": [0, 3, 4], "the": [0, 2, 3]}
collection_b = {"a": [5], "an": [6, 7]}
collection_c = {"the": [10], "not": [10]}
merged_collection = a+b
merged_collection += collection_c
"In the face of ambiguity, resist the temptation to guess."
You could try to define the addition of keys like you suggest, but you need to force a value type which defines a + operation into the same type. Which is what Counter() does, below (appending strings).
Python is dynamically typed, so you can't maintain these expectations. Therefore, summation and bool({}) should be an error, or merely "true" (ie, yes, it exists, and is not None) -- as per the original thread and resolution.
In any case, it's a bit of a distraction. "Most programmers" aren't all that good at programming, or judging what makes a good language. How important then is it that I weigh your projection of their ideas of right or wrong?
In Python it absolutely, positively, without a doubt is not a bad thing to have a function which accepts multiple data types. Some trivial examples are len() and iter(). Since you don't like the core language design in Python, I think it's safe to argue that your sensibilities are going to have many objections to any other part of the language.
> In Python it absolutely, positively, without a doubt is not a bad thing to have a function which accepts multiple data types. Some trivial examples are len() and iter().
I'm not arguing that this shouldn't be the case. My argument is specifically about bool(). If len() or iter() worked as strangely with basic data types as bool(), then I would probably extend my argument to cover them as well. For any data type I'm aware of that works with len() or iter(), the logic of what is returned is pretty clear and obvious, but that's not the case (in my opinion) with bool(). If len(-5) returned 5 (i.e. absolute value or "length from zero") then I would argue that it's a bad idea.
[xebulon:~/tmp] dalke% cat tmp.cc
#include <iostream>
template<typename T>
int f(T s) {
return s ? 2 : 0;
}
main() {
std::cout << "int 0 " << f(0) << std::endl;
std::cout << "int 8 " << f(8) << std::endl;
std::cout << "float 0.0 " << f(0.0) << std::endl;
std::cout << "float -1.0 " << f(-1.0) << std::endl;
}
[xebulon:~/tmp] dalke% g++ tmp.cc
[xebulon:~/tmp] dalke% ./a.out
int 0 0
int 8 2
float 0.0 0
float -1.0 2
C++ containers don't have a bool (or at least vector<> doesn't). For one, until C++11 there was no "explicit operator bool", and a simple "operator bool" was too permissive because of implicit type conversion. C++11 introduced a more contextual conversion to bool.
More and more components have explicit bool support. For example, http://www.boost.org/doc/libs/1_55_0/libs/smart_ptr/shared_p... ?
Notes: This conversion operator allows shared_ptr objects to be
used in boolean contexts, like if(p && p->valid()) {}.
[The conversion to bool is not merely syntactic sugar. It allows shared_ptrs
to be declared in conditions when using dynamic_pointer_cast
or weak_ptr::lock.]
I do not track C++ well enough to go into any more depth than this, especially as it concerns the history and future.
I like to say "if x: ..." and not "if len(x) == 0:" in order to check if a dictionary is empty.
Following Scheme, I can understand that something like "if empty(x)" might be more explicit. But I have a decent amount of code where I do something like:
def process(a, rename=None):
if rename:
a = [rename.get(x, x) for x in a]
... do things with a ...
Without bool, I could write it as:
if rename is not None:
if rename is not None and not empty(rename):
BTW, I don't use "def process(a, rename={}):" because that is one of Python anti-patterns: the default values are constant over the life of the function, so
def something(a, b={}):
b[a] = a
return b
$ python
>>> import time, datetime
>>> () + ()
()
>>> "" + ""
''
>>> 0 + 0
0
>>> datetime.timedelta() + datetime.timedelta()
datetime.timedelta(0)
>>> time.time() + time.time()
27848327051.323207
time.time() returns the seconds elapsed since the epoch.
The code examples seem to be referring to the existence of an identity element, which is necessary for all groups.
The truth is, time doesn't have a sensible addition/combination operator, never mind an identity and inverse, so it isn't even a group.
This is venturing off topic, but if you really wanted to, you could define addition for times in terms of timedeltas. It makes sense to subtract times and get a timedelta, so you could define addition as the subtraction with signed times and signed timedeltas. So, since `time(today) - time(yesterday) = timedelta(1 day)`, we could say that `time(today) + time(yesterday) = -timedelta(1 day)` and `time(today) - -time(yesterday) = timedelta(1 day)`. It's a bit strange to define the notion of a signed time object, but it could work.
a + 0:00 = a
15:00 + 13:00 = 02:00
12:00 + 12:00 = 00:00
I'll assume the times you gave were in the UK's timezone. Here they are rewritten to use Germany's timezone.
a + 1:00 = a
16:00 + 14:00 = 03:00
13:00 + 13:00 = 01:00
Time deltas? Then you don't have the timezone or epoch problem. Works fine.
But time stamps? You can't add two time stamps and get a sensible answer.
Time stamps form a torsor, not a group: http://math.ucr.edu/home/baez/torsors.html
(And 0 == False, "" != False, but both 0 and "" are considered false? C'mon Python, that's borderline JavaScript territory.)
That shortens code like:
if not x:
do_something()
It is a matter of taste, so I rather like it.
Most often than not in my code that wouldn't break functionality. In the case of dates it is broken, so same code pattern there would be broken.
At this point a lot more functionality would break if 0 become non-False or threw a TypeError exception.
I do hope the ticket in question is reconsidered though. Midnight is obviously a "valid and populous instance" that shouldn't evaluate False.
Agreed, "0 as False" sounds like "let's take a leaf from C and its bad type system". I generally rely on falsiness as well, except when dealing with number, and the inconsistency is annoying. I also agree that midnight is not inherently more false than noon or any other date. But I guess that's a risk with falsiness, just like it is with other magic methods: sometimes you take the magic too far.
>>> bool() == int()
TrueWell I am a python programmer that has been using it for almost 10 years, 7 full time, and I say it is nice. So yes, I would call it as a matter of taste.
Also, 1 == True, "x" != True, but both 1 and "x" are considered true. I don't quite understand why you find this so surprising.
The most problematic thing here, IMO, is that bool is a subtype of int in Python.
if cond1:
return -1
elif cond2:
return -2
elif cond3:
return -3
if cond1:
raise cond1Exception
elif
raise cond2Exception
As for uses: because empty lists, zero and empty strings are false, you can do things like:
if validation_errors: if special_instructions: if unusued_widgets:
Of course those are just saved keystrokes, and can lead to bugs, but so is being able to throw a list into a condition.
I think the history of Lisp is instructive. In traditional Lisp all the way through Common Lisp, the empty list, also known as NIL, is false, and everything else is true. This is a relatively benign form of truthiness, but even so, the type-punning it invites tends to make programs less clear (even though they're a little shorter).
Since Scheme was a clean sheet design, Steele and Sussman had the opportunity to change this, and they did: Scheme has explicit '#t' and '#f' literals, and no implicit conversions from other types to boolean. I think their reasoning was sound.
Basically from UTC east to the IDL, it behaves as you describe it. Anything west, it evaluates to true as long as it is set and there is a timezone offset. This is because once you are at a negative offset, you subtract and get a higher number so it will never evaluate to false with a negative offset.
Basically this current behavior is useless for anything I can imagine. You can't use it to check for midnight UTC. You can't use it to check for whether it is set. You can probably use to check for midnight with no timezone but add a timezone and now your handling is messy.
This is why contracts need to be sanely written from a code contract point ahead of time. This was apparently a developer describing how his code worked and later calling it a contract, which is a very bad thing :-D
The rule, in its entirety, is this:
- Booleans are falsy when false.
- Numbers are falsy when zero.
- Containers are falsy when empty.
- None is always falsy.
- No other type of value is ever falsy.
I can think of two ways we might possibly want to alter the rule.
The first is to expand the idea of number to include arbitrary groups (or monoids?), with the identity element being falsy. So, for example, a matrix with all entries zero might be falsy. Or a 3-D transformation might be falsy if it does not move anything.
The second is one I have encountered in C++. There, an I/O stream is falsy if it is in an error state. This makes error checking easy; there is one less member-function name to remember. We might expand this idea to include things like Python's urllib, or any object that wraps a connection or stream of some kind.
EDIT: OTOH, there is the Haskell philosophy, where the only thing that can be evaluated in a Boolean context is a Bool, so the only falsy thing is False.
EDIT 2: The comment by clarkevans (quoting a message from INADA Naoki) already partially addressed the above group idea: "I feel zero value of non abelian group should not mean False in bool context."
x == 0
isempty(x)
x == nothing
Oh the other hand, the cost of truthiness is pretty significant. Each person reading or writing code in a language with truthiness must remember all the arbitrary truthiness rules that particular language uses – and they're quite different for each of Python, Perl, Ruby, C, JavaScript, etc. As this issue demonstrates, whenever you open the door a little, even in a generally sane language like Python, at some point some weird decisions get made and you find yourself with some strange corner cases like midnight being falsey. If you wanted to know if a date was midnight, wouldn't it be easy enough to just explicitly test that?
E.g. you can't say following in C#
if(number) //number is of type integer.
if(o1) //o1 is a reference type variable.
if("string")
if(number = 10) // number = 10 is an assignment and produces 10.
You must write:
if(number == 0)
if(o1 != null)
if("string".Length > 0)
if(number == 10)
This results in some weird results from an intuitive perspective, but is very principled and elegant in other ways.
My one objection was that I don't know how None fits in.
That's like saying "north" is an additive identity on a compass. It's not; "+0° clockwise" is an additive identity.
midnight is not, nor has it ever been, the "zero-time", it's simply a point whose typical representation contains some 0s. It is no more Falsey than the origin (0, 0) in the cartesian coordinate system is Falsey.
This is pretty well-known, I thought.
if datetime.time(0, 0, 0):
print "foo"
else:
print "bar"
There's nothing about midnight that makes sense for it to behave as Falase.
"goto fail" is a well-known error handling mechanism in open source
software, widely reputed for its robusteness:
http://opensource.apple.com/source/Security/Security-55471/libsecurity_ssl/lib/sslKeyExchange.c
https://www.gitorious.org/gnutls/gnutls/source/6aa26f78150ccbdf0aec1878a41c17c41d358a3b:lib/x509/verify.c
I believe Python needs to add support for this superior paradigm.
It would involve a new keyword "fail" and some means of goto'ing to it.
I suggest "raise to fail":
if (some_error):
raise to fail
fail:
<error handling code>
Unless there are many objections, this fantastic idea might be submitted
in a (short) PEP somewhere around the beginning of next month.
There is some obvious overlap with the rejected "goto PEP" (PEP 3163)
and the Python 2.3 goto module. However, the superiority of goto fail as
error generation and error handling paradigm has since then been
thoroughly proven.
(There is absolutely no certificate store checking for certs by default, nor is there any hostname checking, or any of the myriad of other checks one might expect a reasonable TLS implementation to perform. Use the requests module.)
Midnight is a value, not a special value. There is no reason why it or any other valid time should be falsey on a daily cycle.
However, usually I am happy that '0 is false' in PostgreSQL raises a type exception.
Python is stable[0] and places a high degree of importance on backwards compatibility.
This behaviour is well documented (and called out for particular note). This reinforces that it is (a) official and (b) not a bug because it is the documented behaviour.
On the other hand Ruby (and most Ruby libraries) seem both less concerned with backwards compatibility, have less thorough documentation[1] but are more willing to change and improve.
There isn't a right and a wrong between these approaches although for most things I think I would prefer something between the two. I think I generally prefer Python in terms of syntax (Ruby is a bit too flexible with too many ways to do things for my taste) but I do wonder if Python will be left a little behind.
[0] Python 2/3 transition is a single big deliberate change.
[1] I have an open Rails issue that I don't know if is a bug or not because there isn't documentation that is sufficient to compare the behaviour with so it is a case of what feels right/wrong: https://github.com/rails/rails/issues/6659
1) Is there a (native or custom) date type in Python? Is it an object?
2) Midnight when? Today? This date last year? Sure there's a "zero value" for dates - it's the epoch for whichever platform or library you're using.
3) Why in would anyone call it a "date" if it's really a time?
Maybe I'm getting off into the philosophical decisions of the reptile wranglers, but this particular debate sounds a lot like someone made a decision long ago that had ramifications further than expected and now the justification is engrained, things are built on it, and no one's willing to make the 'correction.'
I can understand not wanting to make many inferences, but that sort of answers your questions.
It was a bad idea.
Everyone is talking higher echelons of consideration, but what effect is there on generated byte code or in fitting within the virtual machine's tight pants?
My first thought was this:
http://docs.oracle.com/cd/B19306_01/server.102/b14200/sql_el...
"in Boolean contexts, a time object is considered to be true if and only if, after converting it to minutes and subtracting utcoffset() (or 0 if that’s None), the result is non-zero"
Changing at this point would possibly break code that relied on documented library behavior. That's not a responsible thing to do.
Simple is better than complex.
BTW, do you know the original author of the Zen of Python has posted to this very dicussion (while also being the orginal author of said module):
https://mail.python.org/pipermail/python-ideas/2014-March/02... https://mail.python.org/pipermail/python-ideas/2014-March/02...
The "if timeval:" case may contain fewer characters, but it's less explicit. Being less explicit opens it up to greater ambiguity. Ambiguity is a form of complexity. Complexity is the opposite of simplicity.
The explicit "is not None" check may require more typing, but it's far more explicit and exact. That means it's much less ambiguous, and thus less complex, and thus exhibits greater simplicity.
1. Before applying a numerical value to a Boolean test, ask whether it can ever be zero when that's not the intent of the test.
2. Create a new rule that forbids testing numerical values as though they're Booleans, and break nearly every program in existence.
Hmm ... wait ... I'm thinking it over.
This is the obvious intuition, which doesn't match the current implementation.
> Why would anyone evaluate dates in a boolean context?
I'd guess, usually to tell whether a variable has a date or None. For example, maybe you have some kind of GUI with a form where the user is supposed to fill in a time, and some variable somewhere in the code is set to None [1] when the time field is blank [2], and a time object representing the time entered by the user otherwise.
Maybe you want to require the user to fill in all fields in the form before advancing to the next step in your program's workflow. If you like falsiness, you might write:
if not form.time_field:
show_warning("You must enter a time to proceed")
continue_current_phase()
else:
proceed_to_next_phase()
if form.time_field is not None:
[1] None is the Python equivalent of other programming languages' null (or NULL).
[2] Or unparseable
The current behavior is insane - just fix it! No need for days of discussion on the mailing list or three-point non regression plans.