Misc #17637
openEndless ranges with `nil` boundary weird behavior
Description
Basically it's about this https://andycroll.com/ruby/watch-out-for-nils-in-ranges/
Since Ruby 2.6 we have this weird syntax (0..nil) which is really really bug prone
e.g. we have dynamic upper boundary like
lower = 0
upper = some_method(arg1, arg2)
(lower..upper).each do { |s| some_method2(s) }
We rarely do nil
checks in Ruby so it's really easy to have Infinity loop in the end.
Previous Argument error was more intuitive since it throws exception instead of silently looping forever.
- some additional strange behavior:
(0..nil).count
=> Infinity
(0..Float::INFINITY).count
=> hangs, I guess same infinity loop
Having explicit parameter Float::INFINITY
(as in previous versions) looks more like a proper design instead of allowing nil
as a valid parameter.
You may think of it as I would like to have a range from 0 to nothing, what is it actually ?
And I guess the answer is Nothing.
Fixing (0..Float::INFINITY).count
this case it also important I believe.
Tested on ruby 2.7.1p83
Updated by mame (Yusuke Endoh) almost 4 years ago
- Related to Bug #14845: Endless Range with nil added
Updated by zverok (Victor Shepelev) almost 4 years ago
I believe that using nil
as a signifier of the "open end" is a compromise due to Ruby's ranges polymorphism. You can have range from a string, from time, from date, from any custom comparable class, how you'd signify the "open end" in this case? In some statically typed language it probably could've been some Infinity<Type>
, but in Ruby... IDK, maybe the alternative would be some "generic" Infinity
constant/special value (as incompatible with any type as nil
is, but having different semantics), but it would be too large a change.
Updated by marcandre (Marc-Andre Lafortune) almost 4 years ago
It's not clear what you are proposing. If it is to restore previous behavior, this won't be acceptable because of compatibility.
Note: you should be using size
(lazy), not count
(typically exhaustive):
(0..nil).size # => Infinity
(0..Float::INFINITY).size # => Infinity
Updated by mame (Yusuke Endoh) almost 4 years ago
Hi, I proposed and implemented a endless range.
This is a trade-off between early failure and usability/consistency.
While the feature is indeed error-prone in some cases, it is more consistent and useful.
It is possible to allow only (1..)
and deny (1..nil)
.
In fact, (1..nil)
used to raise an exception for a short period of development phase of Ruby 2.6.
https://github.com/ruby/ruby/commit/48de2ea5f9b9067779acb0f7f76e5f879f2b42c0
But, to create a conditionally endless range, we need to write max ? (1..max) : (1..)
or Range.new(1, max)
if (1..nil)
is prohibited.
The current behavior allows to just write (1..max)
. Thus, it was reverted.
It is very difficult to change the behavior from now because of the compatibility issue.
But as I recall, this is the third time for me to see this issue reported.
(The first is #14845. I couldn't find the second but I think someone said it in GitHub comments or else.)
If it is a major source of bugs, and if a conditionally endless range is very rare, I'm personally open for the change.
Updated by gud (gud gud) almost 4 years ago
What I would like to have in a programming language is standard library explicitly designed.
So my example from the description with range 0 to Nothing speaks for itself.
0..Float::INFINITY
is pretty intuitive but 0..nil
looks more like a bug (I understand this is personal view)
But, to create a conditionally endless range, we need to write max ? (1..max) : (1..) or Range.new(1, max) if (1..nil) is prohibited.
The current behavior allows to just write (1..max). Thus, it was reverted.
I see, but on the opposite hand to handle this nil-case you need something like:
lower = 0
upper = some_method(arg1, arg2)
raise ArgumentError unless upper
(lower..upper).each do { |s| some_method2(s) } # or unless upper
which isn't handy also.
This is all about trade-offs and I understand that, but having (1..)
syntax only is way less implicit and is a good way to leave this functionality if it's really needed, but make a code base slightly cleaner and more intuitive.
Updated by Dan0042 (Daniel DeLorme) almost 4 years ago
mame (Yusuke Endoh) wrote in #note-4:
It is possible to allow only
(1..)
and deny(1..nil)
.
+1 for that. I was surprised that (1..nil)
is allowed. If you have (1..x)
and x is nil, it seems more likely to me that it was unexpectedly nil rather than intended as conditionally endless range.
Maybe something more explicit could be allowed, like (1..:endless)
, then (1..x||:endless)
would be possible without ambiguity.
Updated by Eregon (Benoit Daloze) over 3 years ago
If we change anything here, it should probably be done with beginless Ranges too for consistency.
And then we'd have 4 cases instead of 1 like in @mame's reply.
Also (..)
doesn't parse, one needs (nil..)
.
I think it's not worth breaking compatibility,
especially considering that third-party libraries most likely already rely on Range#{begin/end} == nil => beginless/endless.
How often does one actually get bugs based on this?
I would guess it's pretty rare, most Ranges are on numeric values or Strings, and if you get a nil out of arithmetic, there is definitely something quite fishy in the code.
It seems a good idea to validate upper
in the example if it is potentially very high or not numeric, and the same would apply if upper is Float::INFINITY or 1.0/0.0.
Updated by Dan0042 (Daniel DeLorme) over 3 years ago
What I was suggesting was to use a symbol when creating an endless range, but the end
value would still be nil.
(1..).end #=>nil
(1..nil).end #error
(1..:endless).end #=>nil
It seems pretty easy to accidentally create an endless range, like (1..values.max)
if values is empty. Until 2.6 this kind of validation was built-in so this would be restoring broken functionality; imho this is a bug fix. Although the fact there has been few bug reports in the last 2 years means this is not likely a large problem.
I'm very surprised that (nil..nil)
is even valid.
Updated by gud (gud gud) over 3 years ago
Also (..) doesn't parse, one needs (nil..).
Wow, I think it's absolutely OK that (..)
doesn't parse. Like is it Ruby or https://uk.wikipedia.org/wiki/Brainfuck
And (nil..)
what is that ? Like really if I saw that code 5 years ago without debugging I would say it shouldn't parse as well, my guess would be this is a bug.
My main point is: we already had (0..Float::INFINITY)
in previous versions having that we had proper ArgumentError
on nil
s. So I am not sure why we had to expand the std library for new syntax which bring a lot of confusing e.g.:
a = 1
b = 1
a == b => true
(a...b).size => 0
a = 'a'
b = 'a'
a == b => true
(a...b).size => 0
And now "special case"
a = nil
b = nil
a == b => true
(a...b).size => Infinity
And (1..nil)
iterates for ever which I can read like nil === Float::INFINITY
but actually this returns false
and nil.to_i
returns 0
, but (0..nil.to_i).size
=> 1
If this is not confusing for newcomers then what is ?
I think it's not worth breaking compatibility,
especially considering that third-party libraries most likely already rely on Range#{begin/end} == nil => beginless/endless.
I am trying to think of a good use case for that stuff and I can't find an answer.
If you need an infinite loop
you can always use well known while true
or even loop do
in Ruby. And those are more common and intuitive way to do such a thing.
You can't iterate over nil..nil
because really, what is that ?
Sometimes it's good to revert the stuff even if it's already been using (like pipe operator), still I think the usage of it is not that common.
How often does one actually get bugs based on this?
Not so often but it happens and when it happens it creates a memory leak.
I had a begin/rescue ArugmentError
around my code and had a feeling I was OK and then this happened. BOOM.
I also do have a feeling that regular usage of range e.g. (lower..upper)
is really common in Ruby code, so making this syntax dangerous for a better usability of conditional endless range
which I guess is really rare... feels wrong.
There is a lot of text and please don't get it personal but sometimes I guess this "syntax sugar" stuff drives Ruby in a wrong direction. Aliases/3-4 ways to do the same thing and so on.