Hi Viktor, thanks for your clarifications.
I agree that from a performance point of view, there isn't all that much to be gained. I thought more about parallelizing distinctBy and windowSliding, Perhaps one can squeeze out a modest gain, but I am not excited by the potential. AFAIK, the "greedy/short-circuiting" decision point doesn't have a major impact on performance either. Or am I wrong there? In my mind, given that performance is not worth more than maybe a tweak, this amplifies my first issue with the surface API. I started out thinking that almost nobody is going to write a gatherer, so why worry? But I found myself writing a couple in the last few days. And I wonder whether the current API is at "peak complexity". If I use the factory methods, I have to make a choice between of/ofGreedy and ofSequential/of. And if I don't use the factory methods, I have to mess with a marker interface or a method yielding a magic default. Is there some virtuous collapse? First off, I think factory methods should be the favored approach. And "of" should be the safe choice. That's why I would rename ofSequential into of, and introduce ofParallel for optimization. Like with of/ofGreedy. I have seen some gatherer implementations that don't use the factory methods, even though they could have. Is this flexibility useful? The details are fussy, with the marker interface and the magic default combiner. That's where I thought the characteristics approach is a better API. It has precedence, and it is unfussy. I realize it's not a big deal, and I was going to let it slide. Until I heard Brian's Devoxx talk where he mentioned "peak complexity", and I felt, that's, in a small way, what is present here. Cheers, Cay -- (Moving this to core-libs-dev) Cay, Regarding 1, Characteristics was a part of the Gatherers-contract for a very long time, alas it didn't end up worth its cost. There's a longer discussion on the topic here: https://mail.openjdk.org/pipermail/core-libs-dev/2024-January/118138.html (and I'm sure that there were more, but that's the one which comes to mind) Regarding 2, I did have a prototype which had a Downstream in the Combiner, but created a new dimension of complexity which made it even harder for people to move from sequential to parallelizable. The door isn't closed on it, but I remain unconvinced it's worth the surface area for performance reasons. As a bit of a side-note, it's worth knowing that in the reference stream implementation, it is not unusual that parallel-capable stages are executed as "islands" which means that short-circuiting signals cannot travel across those islands. Since parallel-capable Gatherers can be fused to execute in the same "island" if we get to a place where "all" intermediate operations are parallel-capable Gatherers, there'd be a single end-to-end "island" and hence the ability to propagate the short-circuiting would be preserved in all modes of execution. Also worth knowing that a `gather(…)` immediately followed by a `collect(…)` can also be fused to run together. Cheers, √ Viktor Klang Software Architect, Java Platform Group Oracle ________________________________ From: jdk-dev <jdk-dev-retn at openjdk.org> on behalf of Cay Horstmann <cay.horstmann at gmail.com> Sent: Friday, 4 October 2024 19:58 To: jdk-dev at openjdk.org <jdk-dev at openjdk.org> Subject: Re: New candidate JEP: 485: Stream Gatherers Hi, I have some belated questions about the design choices in this API that I could not find addressed in the JEP. 1. Why aren't characteristics used to express "greediness/short-circuiting" or "sequentialness/parallelizability"? I understand that for the former I use ofGreedy/of, or implement Gatherers.Integrator.Greedy/Gatherers.Integrator. And for the latter, I use ofSequential/of, or, if I implement the Gatherer interface, have the combiner return defaultCombiner() or not. But it seems a bit complex and less familiar than the characteristics mechanism that exists for spliterators, streams, and collectors. The original design document (https://cr.openjdk.org/~vklang/Gatherers.html) used characteristics, so I wonder what motivated the change. 2. Why wasn't the combiner() designed to allow pushing of elements to the end of the first range's sink? Then distinctBy could be parallelized without buffering the elements. More generally, with some state fiddling one can then handle the elements around range splits. As it is, I don't see how to parallelize such computations other than to buffer all elements. I looked at the project at https://github.com/jhspetersson/packrat that implements a number of gatherers. Only one uses a combiner, to join buffers until their contents can be pushed in the finisher. Cheers, Cay -- Cay S. Horstmann | https://horstmann.com
