Range API

[Olexandr Rotan]

Text copied from PR draft since no one really goes through 2-month old
drafts ;)

For anyone wondering what's up with this PR, I have decided to take a
little break from coding itself and focus on the more fundamental part of
ranges. Most of the time I was thinking about the range lifecycle, how the
range that was acquired from the chain of operations is fundamentally
different from the plain range created using Range.of and stuff like that.
Also I have been thinking about how exactly should custom comparators be
consoled with range operations, which is obviously the point of most
friction regarding comparators for ranges.

1. Lifecycle

Generally, I see 4 use cases for ranges. I will put them in "popularity"
order, speaking from my experience
1. Receive boolean value (does contain, does intersect etc.)
2. iterate over (It may become more popular as ranges are present in jdk,
but as Brian Goetz once mentioned in mail, no one is really hurt by the
necessity to write for loop instead of foreach over range, so I don't
consider stuff like for(var i : Range.of(0, 10) a valuable use case).
3. Provide as input (also known as slicing in python and things alike in
other languages)
4. Receiving range as result of method (not really a terminal operation,
but case is huge enough to acknowledge it separately)

I have been through multiple iterations of use-case investigation. At some
point, I have even thought about dropping range operation completely. What
stopped me, though, is one of the tasks I have received at work recently,
which completely changed my mind. The one particular use case that I find
huge enough to be worth support in API is scheduling with only one task per
moment, i.e. scheduling lessons for teachers. for example, accounting for
business hours and stuff like that. I believe this is one of the most
common things business logic programmers deal with when talking about
datetime ranges.

>From this point, I have been looking at this topic from a somewhat
different perspective. I believe that most of the time (dominating majority
of times) results of range operations would be used to acquire boolean
value. Passing it as input would be a mess to handle as range can be
interrupting/continuous. Iterating over ranges like that may lead to
confusing behaviour, for example, consider:

[1;4] U [5;10] with step 3. Should it be like 1,4,7,5,8 or 1,4,5,7,8 or
maybe 1,4,5,8 ? This is just too confusing to consider.

On the other hand, continuous, as-is ranges (ones that are created directly
with factory method), would be a much better candidate for iteration, for
many cases, from safe iteration over substring indexes to just plain for
loop replacement (people for some reason want this a little bit too much).

So, summarizing everything above: range operation is a valid request,
although support for everything available to "plain" ranges would be
inappropriate for more complex ranges, specifically unions. Unions are
mostly used for boolean operations, while plain ranges can have a broader
number of use cases. For iteration over unions, the best option I see is to
provide a method that returns a List of their components so people can
by-hand decide which behaviour exactly do they want.

2. Range operations and comparators rivalry

I think everyone could in 5 seconds come up with at least one case when
these two conflict. I was thinking a lot about this. Prototyping some sorts
of operation contexts (something like java.math.MathContext), creating
classes that were responsible for range operations that required providing
comparator for each action (this one I brushed off for multiple reasons: 3
parameters for each method sounds horrific, many implementation
intricacies, and generally not intuitive way to operate on rages
considering 95% of use cases will be bound to datetime and numerics and
things backed by them), and other stuff. Surprisingly, where I found
similar problems is in the collection framework, specifically in sets. The
problems here were really much more alike to range then I have expected
(basically the same except for the fact that set has only one 2 possible
relations to element: contains or not, while ranges have 5 possible
relations to other ranges (one can be before, after, contained, or
intersect at start or end with other) which obviously complicates things).
Collections framework just state behaviour in operations between sets that
use different equality metrics as well, not undefined, but confusing for
anyone who doesn't know implementation details. I am planning to prototype
something with the same kind of behaviour for ranges, but I still haven't
decided how 2 ranges will communicate with each other (analogy for how sets
communicate using contains).

That's basically it for now, i will continue with investigation, maybe work
will also give some food for though as with schedule, and when I feel like
I have some common ground for all of the things I would want to see in the
API, I will proceed with some sort of more mature API description

On Thu, Sep 26, 2024 at 11:00 PM Olexandr Rotan <rotanolexandr842 at gmail.com>
wrote:

> but I also don't sense people beating down the doors for that (even if the
>> language had range literals, like `0..<100`).
>
>
> True, that what I was thinking also: "is iteration over numeric range is
> so important to challenge general versatility of API?", but
>
> for (int j=index; j<index+target.length(); j++)
>
>
> is the use case that I haven't even thought of. After some research, I
> found a family of methods across jdk that could benefit from ranges. Their
> names match X::indexOf(X) or X::indexOfSubX(X) or similar patterns. I did
> some research and found it surprising that the only young language, which
> all now have first-class support for ranges, that takes advantage of range
> return types is Swift and its string::range(of: substring) (maybe there is
> similar methods for sublists in Swift, I think it is not that significant).
> Others either return an optional int or int itself.
>
> Much more languages, though, have adapted ranges as input parameters for
> slicing and other range-specifying operations, not lastly due to issues like
>
> Similarly, I see errors in API usage because sometimes we specify range by
>> (start, end) and sometimes by (start, length)
>
>
> Though, there are some differences with the vast majority of languages and
> this API currently, mostly with the fact that almost all languages do not
> support any range algebra out-of-the-box (closest I have seen were
> itertools in Rust), and, subsequently, there is nowhere to emerge for unios
> from, unlike current implementation. Adapting return types for being
> three-state algebraic data types (i.e. result of passing empty range,
> result of passing uninterrupted and result of passing union) is not that
> hard, but just another work to do in case similar APIs would be adopted in
> jdk.
>
> My research about iterability of ranges in other languages showed that
> there is, unfortunately, no magic pill invented. Basically, there are two
> approaches to making range iterable:
>
> 1) Implement iterator
> 2) Implement increment/decrement
>
> The concrete way might vary from traits in Rust (and something like this
> in Haskell I suppose) to operator overloading in c++, but concepts are the
> same. Most languages also provide out-of-the-box support for integer
> ranges.
>
> And that is what I concluded on, at least for now. I think that API should
> provide support for any custom iteration strategy using something like
> Collector/Gatherer wrappers over a functional interface that will transform
> increment/decrement function into iterator. Besides that, I think that it
> makes sense to create separate Range.OfInts(int, int) or just
> Range.of(int,int) overload, that will return thin extension of
> Range<Integer> (noone convinces me to create new specialized class), that
> implements iterable using simple int increments, which would both benefit
> the performance and address most of the use cases for "obviously iterable"
> ranges. I also think there is no need for open iterable ranges, even if
> start-closed. I think it is better to have some method to return Stream
> than implement Iterable, since this might lead to unexpected problems.
>
> Best regards
>
> On Thu, Sep 26, 2024 at 4:30 PM Brian Goetz <brian.goetz at oracle.com>
> wrote:
>
>> Sorry for the not-good news, but I'm not too surprised.  Computational
>> domains like "32 bit integers" seem like they should have a lot in common
>> with algebraic structures like groups and rings, but when you start poking
>> at them, the compromises we make to fit them into hardware registers start
>> to bite.  (And let's not get started on floating point...)  Lots of
>> research into numeric towers in various languages, or capturing fundamental
>> properties in type classes like Haskell's `Eq` and `Ord`, offers plenty of
>> compromise to go with its promise.
>>
>> I think a big part of what you are running into is that you've started
>> with a _concept_ (a deceptively simple one, at that), rather than
>> _requirements_.  And it is the open-endedness of this concept (discrete vs
>> continuous, bounded vs half-open, including endpoints or not, etc) that
>> resists abstraction.  Plus, without clear requirements, you will be subject
>> to an endless barrage of "what about my pet use case" (e.g., "what about
>> the numbers zero to ten, advancing by two").  Meanwhile, domain-specific
>> libraries such as java.time will invent their own domain-specific answers,
>> like Interval.
>>
>> Rather than starting from the algebraic properties, perhaps start from
>> the other end: what are the use cases where the lack of a range abstraction
>> is problematic.  I get that
>>
>>     for (int i=0; i<100; i++) { ... }
>>
>> is uglier and less abstract than
>>
>>     for (int i : Range.of(0, 100)) { ... }
>>
>> but I also don't sense people beating down the doors for that (even if
>> the language had range literals, like `0..<100`).
>>
>> Where I do see people having trouble is that many range computations are
>> error prone.  For example, `String::indexOf` returns the starting index of
>> a match; if you want to actually iterate over the characters of such a
>> match, you have to do something like
>>
>>     for (int j=index; j<index+target.length(); j++)
>>
>> and you are at risk for fencepost errors when recreating the range.
>> Whereas an indexOf method (under a more suitable name) that returned a
>> range, would be more amenable to downstream processing.  Similarly, I see
>> errors in API usage because sometimes we specify range by (start, end) and
>> sometimes by (start, length), and since both are ints, we get no type
>> checking when you pass the wrong kinds of ints to such a method.
>>
>> But, the mere existence of a Range type would do little to help String,
>> Arrays, and other range-happy APIs, because we would have to update them to
>> include new overloads that dispense and consume ranges.  So that's a big
>> project.
>>
>> Still, I think investigating use cases involving libraries that work
>> intensively with ranges like this would likely yield useful information for
>> what a Range type would want to provide.
>>
>> HTH,
>> -Brian
>>
>>
>>
>>
>>
>>
>> On 9/26/2024 9:07 AM, Olexandr Rotan wrote:
>>
>> Researching the of deriving some iterable representations from ranges,
>> and I am not here with the good news.
>>
>> Unlike range algebra and boolean operations, which generalize
>> extremely well, iterability of ranges... Well, it's safe to say it
>> doesn't generalize at all. Analyzing key features people expect iterable
>> ranges to have, I ended up concluding there are basically two groups / two
>> use cases for them. First is plain and simple, arguably the most popular
>> one: iterating over a range of integer numbers, i.e. `for (i : Range.of(1,
>> 10))`. Another use case is for more complex iterations over ranges of
>> reference types, most commonly dates/time.
>>
>> There are two groups of values by their nature: discrete and continuous.
>> Most of the types belong to the second group, as there is no direct
>> increment AND decrement for them (we will omit hardware limitations for
>> simplicity), such as floating point values. What is the increment of 1,3?
>> 1.31 or 1.30000000001, or maybe something even more unreadable? On the
>> other hand, the increment of LocalDate in context of range iteration that
>> represents today is rather obvious - it is tomorrow.
>>
>> There is a pretty limited number of discrete types in jdk. Dates, whole
>> numbers and basically that's it. The discrete types that are not present in
>> jdk can be really various. For example, users can define a comparable
>> type "F1Team" and compare them based on their position in the last race.
>> There, increment would most likely be the next team in rating. There are
>> many domain-specific cases like this.
>>
>> This is where the problem comes from. If the user would always have to
>> pass a comparator to create a range, it would be consistent to make the
>> user define increment/decrement as well. But we don't want users to pass a
>> comparator if the type is already comparable. Similarly, we don't want
>> users to define increment/decrement if there is already one in the
>> language! I think defining increments for dates (say LocalDate.plusDays(1))
>> would be acceptable, even defining increments for floats in context of
>> ranges might be acceptable, but making people define increments for
>> integers is, in my opinion, completely not. Besides performance impact,
>> this is a terrible user experience.
>>
>> There are a few solutions to this:
>> 1) Define ton of overrides for factory methods and specialized types for
>> this (uhh, sounds awful)
>> 2) Introduce new interface, say Discrete<T>, that defines T increment()
>> (and possible T decrement()) methods. From now on, there are 2 branches:
>> 2.1) Leave things as is, allow users to define incrementation logic for
>> their types, but don't touch integers and other built-ins.I see this option
>> as extremely inconsistent and not solving the main issue, which is
>> iterability of integers.
>> 2.2) Retrofit (scary) existing types to implement this interface. This
>> should not have any compatibility nor security implications, but still
>> sneaking into java.lang every time we need some new API to be more
>> user-friendly is obviously not a way to go. This basically comes down to a
>> question about how deep we want to integrate ranges into language, and is
>> range  generalization even worth the invasion into the core of
>> language (imo yes).
>> 3) Leave things as they are, just let users derive iterables using
>> something like range.asIterableWithStep(IncremetStartegy increment). I
>> think this would make an API too narrow as no one will use it for routine
>> tasks the same way people do in Rust, Kotlin and other languages.
>>
>> I would love to hear community opinion on this matter. Which option is
>> the most preferable, maybe some compromise between a few of them, or maybe
>> there is a better way to go that I didn't mention here?
>>
>> Best regards
>>
>> On Tue, Sep 24, 2024 at 5:11 PM Alan Snyder <javalists at cbfiddle.com>
>> wrote:
>>
>>> I have another example: I have a datatype that represents a region of an
>>> audio track, for example, one tune in a medley of tunes. I allow the region
>>> to
>>> specify both a start and end time, but the end time is optional (and
>>> mostly not used). When the end time is not specified, the region ends at
>>> the start of the next region, or at
>>> the end of the track if there is no next region. The latter case is
>>> useful because the exact track length may not be known. The optionality of
>>> the end time
>>> is not represented in the type system.
>>>
>>> Having said that, I’m not sure that a general abstract interface would
>>> be useful for this example.
>>>
>>> On Sep 24, 2024, at 2:13 AM, Olexandr Rotan <rotanolexandr842 at gmail.com>
>>> wrote:
>>>
>>> As part of the redesigning  process , I am researching whether or not
>>> there are use cases that require asserting that the range is exactly
>>> half-bounded. This is important because I plan to switch to
>>> BoundedAtEnd/BoundedAtStart sealed interfaces instead of flags and runtime
>>> checks: Here is what I gathered for now.
>>>
>>>
>>>    - *Date/Time Handling (Historical or Forecast Data)*: When dealing
>>>    with events that started at a specific time but have no known end (e.g.,
>>>    open-ended employment contracts or ongoing subscriptions)
>>>    - *Stream Processing (Real-time Event Streams)*: In real-time
>>>    systems, you might process data that has a start time but no defined end,
>>>    such as monitoring a live video feed or logging system. The range is
>>>    bounded at the start and unbounded at the end as more data will
>>>    continuously arrive.
>>>    - *Data Pagination (Fetch Until Condition)*: When implementing
>>>    pagination, sometimes you might want to fetch items starting from a
>>>    specific index up to an unbounded limit (e.g., fetching all items after a
>>>    certain point until memory runs out or a condition is met).
>>>    - *Auditing and Monitoring*: In systems where audit trails or
>>>    logging data should capture all events after a certain point (bounded
>>>    start) with no foreseeable end (unbounded end), such as monitoring changes
>>>    to records in a database starting from a fixed timestamp.
>>>    - *Scientific or Statistical Ranges*: When modeling physical systems
>>>    or statistical ranges, you might want to capture measurements that begin at
>>>    a known threshold but theoretically have no upper or lower bound. For
>>>    example, recording temperature data starting at absolute zero and
>>>    increasing without any known upper limit.
>>>    - *Inventory or Resource Allocation*: Resource allocation policies,
>>>    such as those for virtual machines, may be based on known minimum
>>>    allocation thresholds but have flexible or unbounded resource caps,
>>>    depending on availability.
>>>
>>>    I am writing to ask whether anyone who worked with such systems
>>>    could confirm/deny that those are real use cases. If so, would it be
>>>    satisfying enough to assert one-way unboundness with instanceof checks,
>>>    i.e. range instanceof UnboundedEndRange && !(range instanceof
>>>    UnboundedStartRange). Would appreciate any feedback.
>>>
>>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/core-libs-dev/attachments/20241121/e42785b5/attachment-0001.htm>