RFR: 8072727 - add variation of Stream.iterate() that's finite
Stuart Marks
stuart.marks at oracle.com
Wed Feb 24 01:29:14 UTC 2016
Finally getting back to this.
So, now that this has been clarified to model a for-loop, this seems fine. It
looks like Paul is sponsoring this for you. Great, let's move ahead with it.
As the author of this particular RFE, though, I should say that this isn't what
I had in mind when I wrote the RFE. :-)
There are two main issues I'm concerned about:
1) modeling loops that aren't for-loops; and
2) modeling loops where the termination condition is known to the stream source
but isn't part of the stream value.
If covering these requires a different API and RFE, then that's fine. I just
want to make sure these cases aren't dropped.
For case (1) I think the Collatz example is a good one, since you basically had
to cheat in order to get it to work with this API, either by appending 1, or by
introducing a sentinel value. It's not always possible to do this.
The kind of loops I'm interested in look like this:
do {
emitValue();
} while (condition);
or
while (true) {
emitValue();
if (condition)
break;
stuff();
}
Clearly it's possible to rearrange these to fit into the 3-arg iterate() method,
but it's unnatural to do so. For example, to implement the craps dice game, the
obvious Java approach is to maintain the game state in an object or even local
variables that are mutated as the game progresses. To use the 3-arg iterate()
method, you wrote a State object (a value), which knows how to generate its
successor State, and from which the int values are extracted. This is a very
nice functional approach, but I expect that many people will have data sources
that rely on mutation or external resources. It doesn't seem reasonable to
require rewriting the source logic in order to produce a stream from it. And in
general it might be arbitrarily difficult.
It's probably easier to wrap a Spliterator around the source logic and simply
call it from tryAdvance(). In fact, this is what I'd like to see in a
stream-source API, to make it easier to create such stream sources without
having to subclass AbstractSpliterator.
For (2) it's critical that the stream source have control over production of
values. Placing takeWhile() / takeWhileInclusive() downstream doesn't work,
since the stream machinery (particularly in parallel) may aggressively pull
elements from the source before they're presented to takeWhile().
For example, consider parsing int values from a Scanner into an IntStream, until
there are no more, leaving the scanner at a known state. Suppose there were an
iterate(Supplier<OptionalInt>) overload. Then something like this would work:
Scanner sc = new Scanner("1 2 3 4 foo");
IntStream.iterate(
() -> sc.hasNextInt() ? OptionalInt.of(sc.nextInt())
: OptionalInt.empty())
.forEach(System.out::println);
assert sc.hasNext();
System.out.println(sc.next()); // prints "foo"
Boxing things into optionals is a bit cumbersome, and I'm also concerned about
the overhead. But it illustrates the point that the termination condition is
separate from the values that go into the stream. In particular there's no
"seed" value to which a termination condition can be applied. (And "iterate"
might not be the best name for this method, either.)
Oh well, I think this is turning into another RFE.
s'marks
On 2/16/16 9:48 PM, Tagir F. Valeev wrote:
> Hello, Stuart!
>
> Thank you for your comments.
>
> SM> I'd suggest focusing on the API first before worrying about how to track the
> SM> stream state with booleans, etc. Is the API convenient to use, and how well does
> SM> it support the use cases we envision for it?
>
> As Brian already noted, the most benefit of such signature is the
> resemblance to the good old for loop. Also it's good, because the
> lambdas don't need to maintain external mutable state in this case
> (the state is encapsulated in the current element). Most of your
> proposed examples, however, need to do it as they don't receive the
> existing state. Also I see no reason to create a method which is
> compatible with iterator::hasNext/iterator::next or even
> spliterator::tryAdvance. If you already have a spliterator, you can
> create a stream using StreamSupport.stream(spliterator, false). If you
> have an iterator, you can convert it to spliterator using
> Spliterators.spliterator[UnknownSize]. Well, this probably looks ugly,
> but more flexible and signals that some low-level stuff is performed.
>
> Supplier<Stream<T>> is definitely bad in terms of performance.
> Creating a new stream is not nearly free. To illustrate this I wrote a
> simple benchmark which compares .flatMapToInt(OptionalInt::stream)
> with Java8 way .filter(OptionalInt::isPresent).mapToInt(OptionalInt::getAsInt)
>
> Here's source code and full output:
> http://cr.openjdk.java.net/~tvaleev/jmh/optionalstream/
>
> Benchmark (n) Mode Cnt Score Error Units
> OptionalTest.testOptionalFilterGet 10 avgt 30 0,171 ± 0,011 us/op
> OptionalTest.testOptionalFilterGet 1000 avgt 30 6,295 ± 0,046 us/op
> OptionalTest.testOptionalFilterGet 1000000 avgt 30 12597,706 ± 69,214 us/op
> OptionalTest.testOptionalStream 10 avgt 30 0,330 ± 0,002 us/op
> OptionalTest.testOptionalStream 1000 avgt 30 27,552 ± 0,577 us/op
> OptionalTest.testOptionalStream 1000000 avgt 30 30837,240 ± 812,420 us/op
>
> Involving intermediate streams makes the thing at least twice slower.
> Surely this delay could become negligible in some scenarios, but I
> think it's inacceptable to enforce users to create new source with a
> bunch of streams. At least primitive specializations will become
> meaningless in this case: boxing would eat much less time compared to
> stream creation.
>
> As for elements drawn from the queue, it's much better to use existing
> takeWhile method:
>
> queue.stream().takeWhile(x -> x.equals(sentinel));
>
> True, such approach will not include the sentinel element to the
> result, and there's no easy way to do it with current API. Probably
> additional method (takeWhileInclusive?) could be considered to solve
> such problems. Still, I think, drawing from the queue is not the
> problem which should be solved with new iterate() method.
>
> As for Collatz conjecture, it's quite easy to iterate without trailing
> one:
>
> IntStream.iterate(start, val -> val != 1,
> val -> val % 2 == 0 ? val / 2 : val * 3 + 1)
>
> If having one is desired, then it would be easier just to append one
> to the stream (even if Collatz conjecture is false we will have an
> infinite stream, so appended one will never appear):
>
> IntStream.concat(
> IntStream.iterate(start, val -> val != 1,
> val -> val % 2 == 0 ? val / 2 : val * 3 + 1),
> IntStream.of(1))
>
> A side note: having IntStream.append(int... numbers) would be really
> nice:
>
> IntStream.iterate(start, val -> val != 1,
> val -> val % 2 == 0 ? val / 2 : val * 3 + 1).append(1)
>
> Another approach would be to introduce a special stop value (for
> example, -1):
>
> IntStream.iterate(start, val -> val != -1,
> val -> val == 1 ? -1 : val % 2 == 0 ? val / 2 : val * 3 + 1)
>
> This stream produces Collatz series, including the trailing one.
>
> As for Craps, I never heard about such game. If I understood the rules
> correctly, it's good to represent the state as separate object and
> define state transition via its method. Something like this should
> work:
>
> Random r = new Random();
> IntSupplier dice = () -> r.nextInt(6)+r.nextInt(6)+2;
> class State {
> int roll, point;
>
> State(int roll, int point) {
> this.roll = roll;
> this.point = point;
> }
>
> State() {
> this(dice.getAsInt(), 0);
> }
>
> boolean isStopRound() {
> return roll == 7 || (point == 0 && (roll > 10 || roll < 4)) || (point != 0 && roll == point);
> }
>
> State next() {
> return isStopRound() ? null : new State(dice.getAsInt(), point == 0 ? roll : point);
> }
> }
> Stream.iterate(new State(), Objects::nonNull, State::next)
> .mapToInt(state -> state.roll)
> .forEach(System.out::println);
>
> With best regards,
> Tagir Valeev.
>
>
> SM> In particular, I can imagine a number of cases where it would be very helpful to
> SM> be able to support an empty stream, or where the computation to produce the
> SM> first element is the same as the computation to produce subsequent elements.
> SM> Requiring a value for the first stream element is at odds with that.
>
> SM> Here are some ideas for use cases to try out:
>
> SM> - a series of dice rolls representing a round of craps [1]
> SM> - elements drawn from a queue until the queue is empty or until
> SM> a sentinel is reached
> SM> - a sequence of numbers that (probably) terminates but whose length
> SM> isn't necessarily known in advance (e.g. Collatz sequence [2])
>
> SM> [1] https://en.wikipedia.org/wiki/Craps
>
> SM> [2] https://en.wikipedia.org/wiki/Collatz_conjecture
>
> SM> Note that in some cases the sentinel value that terminates the stream should be
> SM> part of the stream, and in other cases it's not.
>
> SM> I'm sure you can find more uses cases by perusing Stack Overflow. :-)
>
> SM> I'm a bit skeptical of the use of "iterate" for producing a finite stream. There
> SM> are the usual issues with overloading, but there's also potential confusion as
> SM> some forms of iterate() are infinite and others finite. I'll suggest the name
> SM> "produce" instead, but there are surely better terms.
>
> SM> One thing to think about is where the state of the producer is stored. Is it
> SM> expected to be in an argument that's passed to each invocation of the functional
> SM> argument, or is it expected to be captured? I don't think there's an answer in
> SM> isolation; examining use cases would probably shed some light here.
>
> SM> Here are a few API ideas (wildcards elided):
>
> SM> --
>
> SM> <T> Stream<T> iterate(T seed, Predicate<T> predicate, UnaryOperator<T> f)
>
> SM> The API from your proposal, for comparison purposes.
>
> SM> --
>
> SM> <T> Stream<T> produce(Supplier<Optional<T>>)
>
> SM> Produces elements until empty Optional is returned. This box/unboxes every
> SM> element, maybe(?) alleviated by Valhalla.
>
> SM> --
>
> SM> <T> Stream<T> produce(BooleanSupplier, Supplier<T>)
>
> SM> Calls the BooleanSupplier; if true the next stream element is what's returned by
> SM> calling the Supplier. If BooleanSupplier returns false, end of stream. If you
> SM> have an iterator already, this enables
>
> SM> produce(iterator::hasNext, iterator::next)
>
> SM> But if you don't have an iterator already, coming up with the functions to
> SM> satisfy the iterator-style protocol is sometimes painful.
>
> SM> --
>
> SM> <T> Stream<T> produce(Predicate<Consumer<T>> advancer)
>
> SM> This has an odd signature, but the function is like Spliterator.tryAdvance(). It
> SM> must either call the consumer once and return true, or return false without
> SM> calling the consumer.
>
> SM> --
>
> SM> <T> Stream<T> produce(Consumer<Consumer<T>> advancer)
>
> SM> A variation of the above, without a boolean return. The advancer calls the
> SM> consumer one or more times to add elements to the stream. End of stream occurs
> SM> when the advancer doesn't call the consumer.
>
> SM> --
>
> SM> <T> Stream<T> produce(Supplier<Stream<T>>)
>
> SM> A variation of Supplier<Optional<T>> where the supplier returns a stream
> SM> containing zero or more elements. The stream terminates if the supplier returns
> SM> an empty stream. There "boxing" overhead here, but we don't seem to be bothered
> SM> by this with flatMap().
>
> SM> --
>
> SM> s'marks
>
>
> SM> On 2/14/16 6:53 AM, Tagir F. Valeev wrote:
>>> Hello!
>>>
>>> I wanted to work on foldLeft, but Brian asked me to take this issue
>>> instead. So here's webrev:
>>> http://cr.openjdk.java.net/~tvaleev/webrev/8072727/r1/
>>>
>>> I don't like iterator-based Stream source implementations, so I made
>>> them AbstractSpliterator-based. I also implemented manually
>>> forEachRemaining as, I believe, this improves the performance in
>>> non-short-circuiting cases.
>>>
>>> I also decided to keep two flags (started and finished) to track the
>>> state. Currently existing implementation of infinite iterate() does
>>> not use started flag, but instead reads one element ahead for
>>> primitive streams. This seems wrong to me and may even lead to
>>> unexpected exceptions (*). I could get rid of "started" flag for
>>> Stream.iterate() using Streams.NONE, but this would make object
>>> implementation different from primitive implementations. It would also
>>> be possible to keep single three-state variable (byte or int,
>>> NOT_STARTED, STARTED, FINISHED), but I doubt that this would improve
>>> the performance or footprint. Having two flags looks more readable to
>>> me.
>>>
>>> Currently existing two-arg iterate methods can now be expressed as a
>>> partial case of the new method:
>>>
>>> public static<T> Stream<T> iterate(final T seed, final UnaryOperator<T> f) {
>>> return iterate(seed, x -> true, f);
>>> }
>>> (same for primitive streams). I may do this if you think it's
>>> reasonable.
>>>
>>> I created new test class and added new iterate sources to existing
>>> data providers.
>>>
>>> Please review and sponsor!
>>>
>>> With best regards,
>>> Tagir Valeev.
>>>
>>> (*) Consider the following code:
>>>
>>> int[] data = {1,2,3,4,-1};
>>> IntStream.iterate(0, x -> data[x])
>>> .takeWhile(x -> x >= 0)
>>> .forEach(System.out::println);
>>>
>>> Currently this unexpectedly throws an AIOOBE, because
>>> IntStream.iterate unnecessarily tries to read one element ahead.
>>>
>
More information about the core-libs-dev
mailing list