Into (original) (raw)

Remi Forax forax at univ-mlv.fr
Sat Dec 22 09:55:56 PST 2012

On 12/22/2012 06:46 PM, Brian Goetz wrote:

I was not proposing "let's replace into with toList". I already proposed in a previous message "let's replace into() with something whose semantics are more like reduce/forEach rather than the current weird semantics". The recent developments in reduce give us a consistent framework for doing that.

You want to replace into() with a kind of reducers and toList is a way to hide the implementation of that kind of reducers so ...

Into is bad. It seemed clever at the time we first thought of it (turning the problem around on the target collection), but if you look at its semantics and its implementation, its complicated, it's neither-here-nor-there, its unlike any other op, and the obvious usages will likely parallelize terribly. Something that is more reduce-like (or forEach like with a concurrent collection, if we had one) will be much easier to understand AND will perform better.

I understand why into() is bad when you call parallel() on the stream but it's also verify useful if you don't call parallel on a stream because at the boundary of methods people can use collections and are able to choose the implementation.

So the question is not "should we replace into with toList." We have to have a discussion about what we replace into() with. This query was about whether sequential() still has a use other than as a (bad) implementation crutch for into(). And it seems that there is none, which is nice -- one less weird operation, replaced with one more version of the reduce swiss-army-knife. If there is a toList() it will (obviously) be sugar on top of something more general.

and here you agree with me that toList() on a parallel stream is a way to abstract the implementation of some magic concurrent reducer.

Rémi

On 12/22/2012 12:29 PM, Remi Forax wrote: On 12/22/2012 06:16 PM, Brian Goetz wrote: Right. You want to do the upstream stuff in parallel, then you want to do the downstream stuff (a) serially, (b) in the current thread, and probably (c) in encounter order.

So, assume for sake of discussion that we have some form of .toList(), whether as a "native" operation or some sort of reduce/combine/tabulate. Then you can say: parallelStream()...toList().forEach(...) and the list-building will happen in parallel and then forEach can help sequentially. Given that, is there any reason left for sequential()? You need also a toSet() and yes, in that case you don't need sequential() and it's better because it's more explicit. But there is still a lot of cases where you are sequential and you want to control the destination collection, that's why we need into(). I don't think it's a good idea to oppose toList() and into() both have their purposes. Rémi

On 12/22/2012 11:55 AM, Joe Bowbeer wrote: The main use case is sequential().forrEach(), which inserts any ol' for-loop into a computation. On Dec 21, 2012 1:48 PM, "Brian Goetz" <brian.goetz at oracle.com_ _<mailto:brian.goetz at oracle.com>> wrote: If we get rid of into(), and replace it with an explicit reduction [1], then we may be able to get rid of sequential() too. The primary use case for sequential() was for using non-thread-safe Collections as an into() target. The convoluted into() turns the problem around on the target, calling target.addAll(stream) on it, and the typical implementation of into() is like the one in Collection: default void addAll(Stream<? extends E> stream) { if (stream.isParallel()) stream = stream.sequential(); stream.forEach(this::add); } or more compactly default void addAll(Stream<? extends E> stream) { _stream.sequential().forEach(this::add); } since sequential() is now a no-op on sequential streams. The code looks pretty, but the implementation is not; sequential() is a barrier, meaning you have to stop and collect all the elements into a temporary tree, and then dump them into the target. But it is not obvious that it is a barrier, so people will be surprised. (And on infinite streams, it's bad.) What used to be implicit in sequential() can now be made explicit with: if (stream.isParallel()) stream = _stream...whateverWeCallOrderPreservingI_nto().stream() That offers similar semantics and at least as good performance, while also being more transparent and requiring one fewer weird stream operation. (I don't yet see the way to getting rid of .parallel(), but we can possibly move it out of Stream and into a static method Streams.parallel(stream), at some loss of discoverability. But we can discuss.) [1] Actually, we have to replace it with two explicit reductions, or more precisely, a reduction and a for-eaching. One is the pure reduction case that involves merging, and is suitable for non-thread-safe collections (and required if order preservation is desired); the other is the concurrent case, where we bombard a concurrent collection with puts and hope it manages to sort them out. The two are semantically very different; one is a reduce and the other is a forEach, and so they should have different manifestations in the code. Though there are really no concurrent Collections right now anyway (though you could fake a concurrent Set with a concurrent Map.)

On 12/21/2012 12:50 PM, Brian Goetz wrote: I'm starting to dislike "into". First, it's the only stream method which retains mutable state from the user. That's not great. Second, the parallel story is bad. People are going to write list.parallel(e -> e+1).into(new ArrayList<>()); which will do a whole lot of trivial computation in parallel, wait on the barrier implicit in sequential(), and then do an O(n) serial thing. Third, the semantics are weird; we do this clever trick where collections have to decide whether to do insertion in serial or parallel. But as we all learned from Spinal Tap, there's a fine line between clever and stupid. Instead, we could treat this like a mutable reduce, where leaves are reduced to a List, and lists are merged as we go up the tree. Even with dumb merging is still going to be much faster than what we've got now; no barrier, no buffer the whole thing and copy, and the worst serial step is O(n/2) instead of O(n). So probably 3x better just by improving the serial fractions. But with a smarter combination step, we can do better still. If we have a "concatenated list view" operation (List concat(List a, List b)), which returns a read-only, conc-tree representation), then the big serial stage goes away. And, of course, building atop reduce makes the whole thing simpler; there are fewer ops that have their own distinct semantics, and the semantics of into() is about as weird as you get. Now that the tabulators framework gets users comfortable with the explicit choice between functional and concurrent aggregation for tabulation, it is a much shorter hop to get there. So let's build on that and find some sort of way to surface mutable and concurrent versions of "into". (Currently we have no good concurrent list-shaped collections, but hopefully that changes.) Something like: _Stream.tabulate(collector(ArrayList::new)) _Stream.tabulate(_concurrentCollector(ConcurrentFooList::new)) Maybe with some rename of tabulate. I think there's a small reorganization of naming lurking here (involving tabulate, Tabulator, ConcurrentTabulator, MutableReducer, reduce) that recasts into() either as an explicit functional or concurrent tabulation. And one more tricky+slow special-purpose op bites the dust, in favor of something that builds on our two favorite primitives, fold (order-preserving) and forEach (not order-preserving.)

More information about the lambda-libs-spec-experts mailing list