[#3821] Fix FluxBuffer to request 1 when buffer is not modified

[Sage-Pierce]

If a Set is used as the destination in Flux.buffer, the stream will not hang if/when there are duplicates in a given buffer

Previously, FluxBuffer was not taking the result of adding to the buffer into account. If adding to the buffer does not result in modifying it, an extra request(1) should be issued.

Fixes #3821

[chemicL]

Hey, @Sage-Pierce !

I apologize it took some time to respond. I reworked the bufferTimeout operator with fair backpressure recently, that's the reason for it.

Considering all existing buffer* implementations I believe their behaviour should be consistent across the offering. I wonder how this requirement would influence other implementations as it feels a bit risky in the face of concurrency. Can you share your thoughts? After looking at the codebase it seems the authors didn't consider this scenario and assumed there's no filtering happening and once an item is added, the collection's size increases. I wonder whether including this check and handling is the way to go or perhaps limiting the possible Collection types is better while combining with another operator to remove the duplicates would be a safer bet.

Let me know, thanks!

[Sage-Pierce]

Hi @chemicL 👋

No worries about the delay 😄 I didn't immediately look into the other buffer* permutations, as I precisely wanted to see what the feedback on this change was before undertaking that. I would agree that consistency across all the permutations is desirable. The weird bit is that all the other permutations use some form of sizing and/or either some asynchronous trigger or a predicate. It's not clear to me that this would actually be an issue in all those other permutations, and after some initial poking around, it doesn't seem to be.

it feels a bit risky in the face of concurrency

I don't think that there's any more risk with this change regarding concurrency, since onNext is serialized, and the interaction with the buffer is already synchronized where necessary.

I wonder whether including this check and handling is the way to go or perhaps limiting the possible Collection types is better while combining with another operator to remove the duplicates would be a safer bet.

I don't think that there is currently another way to implement "give me n distinct items at a time", though I initially thought there could be with something like window, but I couldn't/can't figure it out

[chemicL]

I don't think that there is currently another way to implement "give me n distinct items at a time", though I initially thought there could be with something like window, but I couldn't/can't figure it out

I came up with this:

public static void main(String[] args) {
	Sinks.Many<Integer> makeACut = Sinks.many().unicast().onBackpressureBuffer();
	Flux.just(1, 1, 2, 3, 4, 4, 4, 4, 5, 6, 6, 7)
			.window(makeACut.asFlux())
			.concatMap(w -> w.reduceWith(HashSet::new, (set, i) -> {
				set.add(i);
				if (set.size() == 2) {
					makeACut.tryEmitNext(0);
				}
				return set;
			}))
			.filter(set -> !set.isEmpty())
			.doOnNext(System.out::println)
			.blockLast();
}

Just as a conversation starter :) I do imagine this doesn't look as nice and the performance would be incomparable.

I'll try to digest the rest of the comments and review the other implementations next. For now, can you also prepare a few sample {input, output} sets so that we know what the end goal is? I mean a sequence 1, 2, 1, 3 would yield [1, 2] and [1, 3] for n == 2, but would yield [1, 2, 3] for n == 3. Is that desired? Can you share some real world scenarios that come to mind that this would benefit? I tend to first try to understand the need and then try to work towards a solution that matches the expectations. This potential mismatch regarding expected supplied aggregator types is puzzling and it would be neat if we could comprehensively address this.

[Sage-Pierce]

Ah nice, that was abstractly what I had in my head, but I couldn't come up with that Sink feedback loop that you used 😄

can you also prepare a few sample {input, output} sets so that we know what the end goal is? I mean a sequence 1, 2, 1, 3 would yield [1, 2] and [1, 3] for n == 2, but would yield [1, 2, 3] for n == 3. Is that desired?

The test I wrote for this changeset covers the basic expectation I think, and it looks like you already understand my intent quite well. I'll just format those and a few more below:

Given flux.buffer(2, LinkedHashSet::new):

• [1, 2] -> [1, 2]
• [1, 1, 2] -> [1, 2]
• [1, 2, 1] -> [1, 2], [1]
• [1, 2, 1, 3] -> [1, 2], [1, 3]
• [1, 1, 2, 3] -> [1, 2], [3]
• [2, 1, 1, 3] -> [2, 1], [1, 3]

Can you share some real world scenarios that come to mind that this would benefit?

In my use case, I am iterating over time-bucketed data elements (from a database) and executing an I/O-bound process on them (a service call). That service call is maximally efficient when passed n distinct elements. The "buckets" are large, usually much larger than n, so the optimization I'm driving toward is iterating over n distinct elements at a time from each "bucket" and executing the I/O bound process for every n elements.

[chemicL]

Thanks.

For bufferTimeout:

Replace

.<Set<Integer>>buffer(2, HashSet::new)

with

.<Set<Integer>>bufferTimeout(2, Duration.ofDays(30), HashSet::new)

and you can observe the same outcome.

Out of the others, bufferWhen also works with a Supplier of Collection, but there's no specific size so this wouldn't be an issue for that variant.

[Sage-Pierce]

For the bufferTimeout behavior, that issue is slightly different than the issue being fixed here. In bufferTimeout, the amount of buffered items is tracked independently of the buffer size. This leads to FluxBufferTimeout not having the same issue of hanging as FluxBuffer. However, it does make the size-triggered emission slightly different:

Given flux.bufferTimeout(2, Duration.ofSeconds(1), LinkedHashSet::new), the current behavior is (differences marked with *):

• [1, 2] -> [1, 2]
• [1, 1, 2] -> [1], [2]*
• [1, 2, 1] -> [1, 2], [1]
• [1, 2, 1, 3] -> [1, 2], [1, 3]
• [1, 1, 2, 3] -> [1], [2, 3]*
• [2, 1, 1, 3] -> [2, 1], [1, 3]

Due to that significant difference in how FluxBufferTimeout keeps track of the number of buffered elements, I'm not convinced this particular behavior can be made consistent with FluxBuffer without significant refactoring of either operator.

Given that there isn't actually a "hanging" issue with bufferTimeout and the usage of a Set buffer supplier, and given that FluxBuffer's upstream request count is decoupled from the number of elements actually contained in emitted buffers, I'd be inclined to treat this behavioral difference as a separate issue, and one which may not be a practical issue at all, since bufferTimeout users inherently have to be prepared to handle buffers with sizes less than the max buffer size.

Thoughts?

[chemicL]

Thanks for following up. I agree that bufferTimeout can be treated differently, thanks for sharing that view. Let's consider what would be a consistent approach to this that wouldn't surprise users of the API once we make changes around supporting Set. From the Javadoc perspective of bufferTimeout:

(...) buffers that will be emitted by the returned {@link Flux} each time the buffer reaches a maximum size OR the maxTime {@link Duration} elapses (...)

In my view, the current behaviour when presented with a Collection that can return false upon add() doesn't deliver what the Javadoc's promises. For one, this PR sounds justifiable from that perspective and thank you for starting the conversation :)

I think in order to merge something we'd need to cover all buffer* cases that accept a Supplier<C> where C is a Collection for the upstream type.

As this currently doesn't work correctly nor consistently we should make an effort to bring more clarity in the docs and tests.

For bufferTimeout:

1. I'd suggest adding a test which validates the Set against an upstream with N consecutive duplicate values, where N is higher than the prefetch size. If there is no issue here, let's proceed with next steps, otherwise we should seek on explaining the failures (just like some operators state that an accepted Publisher needs to be finite -> we are sometimes not at liberty to limit implementations by type but only by the specification).
2. Please document in the Javadoc that for Collections that are capable of returning false from add() the buffers can be smaller than the expected buffer size.

For buffer:

1. Please also address the other Subscriber implementations. You touched BufferExactSubscriber, but we also have BufferSkipSubscriber and BufferOverlappingSubscriber.
2. We need more test cases.

I understand this requires more work so please let me know if you're still keen to contribute. I'd just like us to have a consistent UX across similar operators and that requires a holistic approach. I'll be away for a week but if you make any progress, please do commit and I'll review the changes when I'm back.

Thanks again @Sage-Pierce and I look forward to where this discussion leads us :)

[Sage-Pierce]

@chemicL I don't mind taking a stab at all of that 😄 May take some time, but may have an updated review next week.

[Sage-Pierce]

@chemicL I believe I have addressed your feedback, and I look forward to your re-review when you return. I will be on vacation for the first half of July, so it may take me a bit to follow up on further feedback.

[chemicL]

Thanks for the follow-up. Please have a look at my comments. For the skip != maxSize case I have different expectations for the results in the test cases after reading the javadoc. These should:

• a) Either comply with the existing specification (new buffer after skip items were emitted by the source).
• b) Or adjust the spec to the special case.

With a) it would be necessary to emit the current buffers despite their size being smaller than max but according to considered emitted items so far that fall into the window observed by a particular buffer... And maxSize == skip should not be a special case. So this would lead to an inefficiency in just emitting smaller buffers and that's not your goal.

I think we're left with b) and I guess that in such a case the overlapping buffers would need to be smaller in size once they fall out of scope. For the disjoint case (skip >= maxSize) we can just pretend that the discarded items were never emitted.

Additionally, please note the "Discard support" section in the relevant Javadocs -> they should also explain what's happening here.

Let me know your thoughts and thanks for the effort so far.

[chemicL]

Suggested change

	* less than the specified max size.
	* less than the specified max size. The element will be discarded in such a case.

[chemicL]

Same suggestion as above.

[chemicL]

Same suggestion as above.

[chemicL]

And the same suggestion as above.

[chemicL]

The return here doesn't comply with the promise from the Javadoc:

Buffers can be created with gaps, as a new buffer will be created every time {@code skip} values have been emitted by the source

In this case what would happen is that the item that's not considered for the buffer would not be accounted in the skip tracking, which doesn't refer to the buffer size, but to the emitted items (line 324 would need to be executed in any case). Being compliant would lead to a few more problems, because now we have a non-emitted buffer but the next step could create a new buffer... A few things to consider, suddenly this case would need to handle overlapping buffers, etc.

I suggest documenting this nuance for Collections that can return false upon addition, that if items get discarded, the accounting is not applied. I think that would be ok, as currently the behaviour is already broken for this specific case (stalls happen due to no further requests taking place).

[chemicL]

I imagine it can be the case. In a recent buffer the item is not a duplicate, but in the older buffers it can appear to be a duplicate. Therefore, an item can be discarded only provided that it is a duplicate in all currently tracked buffers. That also necessitates that the old buffers get emitted despite being smaller in size, otherwise there is a risk of leaking memory. Consider the following:

maxSize = 3
skip = 1
input: 1|2|1|2|1|2

step1: 1
  buf1 = 1     // ok
step2: 2
  buf1 = 1|2   // ok
  buf2 = 2     // ok
step3: 1
  buf1 = 1|2   // fail(1)
  buf2 = 2|1   // ok
  buf3 = 1     // ok
step4: 2
  buf1 = 1|2   // fail(2)
  buf2 = 2|1   // fail(2)
  buf3 = 1|2   // ok
  buf4 = 2     // ok
and so on.

[chemicL]

Ah and regarding the build failure -> make sure to run ./gradlew spotlessApply before committing.

[chemicL]

After giving it some though, I wonder whether it makes sense to introduce a new operator for this particular purpose, like

public final <C extends Collection<? super T>> Flux<C> accumulate(Supplier<C> bufferSupplier, Predicate<C> emitPredicate)

which would emit the buffer once the condition is met and then start a new buffer.

Usage:

flux.accumulate(() -> new HashSet(), set -> set.size() == N);

And for the existing nuance with Collections that can return false, just document that it's not supported for buffer operators.

WDYT?

[chemicL]

The more I think about it the more doubts I have :)
As a generic operator it feels that any attempt with Sets is spawning a set of problems. Considering the example of
1|2|1|2|1|2... stream, any strong choice of the size parameter would never emit anything until the source completes. That sounds like a risk for more complaints than the gain of addressing the particular need that you have. Your examples make some implicit assumptions about the source stream, like the global approximation of monotonicity over time with possible local duplicates that you wish to eliminate with "best-effort" approach. However, as a generic set of operators, we can't make such assumptions.

I consulted RxJava's codebase, which shares the implementation of the buffer operator with reactor. There is also no handling of the return value of C.add() in the implementation. In Akka, similar operators, called grouped*, just work with List and don't run into such an issue.

To add to that, the suggestion I made with using window sounds feasible in the face of your use case, where you actually intend to send the aggregations of unique items over the network. In such a case, the overhead of using window instead of buffer should be diminished. Considering that particular use case, I imagine a case where a situation might happen where a time-related trigger can also be necessary. E.g. in case you have aggregated N-1 items in the buffer and now there is a stall in the source emissions for e.g. 1s. It would probably be reasonable to send the incomplete buffer after a smaller delay according to the SLA of your business component.

Consider the following idea for your use case (with artificial delays introduced):

	int maxSize = 3;
	Duration timeBoundary = Duration.ofMillis(100);
	Duration itemDelay = Duration.ofMillis(60);

	Sinks.Many<Object> makeACut = Sinks.many()
	                               .unicast()
	                               .onBackpressureBuffer();
	Flux.just(1, 2, 1, 2, 1, 2)
			.delayElements(itemDelay)
			.window(makeACut.asFlux())
			.concatMap(items -> items.distinct()
			                         .<Integer>zipWith(Flux.fromStream(IntStream.iterate(1, i -> i + 1).boxed())
			                                               .timeout(timeBoundary)
			                                               .doOnError(e -> makeACut.tryEmitNext("TIME_SLICE"))
			                                               .onErrorComplete())
			                         .map(t -> {
				                         if (t.getT2() == maxSize) {
					                         makeACut.tryEmitNext("SIZE_SLICE");
				                         }
										 return t.getT1();
			                         })
			                         .collectList()
			)
		.doOnNext(buf -> System.out.println(Instant.now() + ": " + buf))
        .blockLast();

[chemicL]

I feel I've re-invented the bufferTimeout operator in the above comment 😅 I'm looking forward to your comments @Sage-Pierce .