[External] : Re: Should mapConcurrent() respect time order instead of input order?

Mon Jun 2 16:54:08 UTC 2025

Hi Viktor,

Thanks for your reply and for sharing your experience regarding user
preferences. I appreciate that perspective.

You're right, if an unordered version of mapConcurrent proves to be widely
beneficial and is implemented and adopted by the community, it could
certainly make a strong case for future inclusion in the JDK.

I wanted to clarify a nuance regarding user preference that I might not
have articulated clearly before. If the question is simply "ordered or
unordered?", in isolation, I can see why many, myself included, might lean
towards "ordered" as a general preference.

However, the decision becomes more complex when the associated trade-offs
are considered. If the question were phrased more like, "Do you prefer an
ordered mapConcurrent by default, even if it entails potential performance
overhead and limitations for certain use cases like race() operations,
versus an unordered version that offers higher throughput and broader
applicability in such scenarios?" my (and perhaps others') answer might
differ. The perceived cost versus benefit of ordering changes significantly
when these factors are explicit.

My initial suggestion stemmed from the belief that the performance and
flexibility gains of an unordered approach for I/O-bound tasks would, in
many practical situations, outweigh the convenience of default ordering,
especially since ordering can be reintroduced relatively easily, and
explicitly, when needed.

Thanks again for the discussion.

Best regards,

On Mon, Jun 2, 2025 at 8:51 AM Viktor Klang <viktor.klang at oracle.com> wrote:

> >My perspective is that strict adherence to input order for
> mapConcurrent() might not be the most common or beneficial default
> behavior for users.
>
> If there is indeed a *majority* who would benefit from an unordered
> version of mapConcurrent (my experience is that the majority prefer
> ordered) then, since it is possible to implement such a Gatherer outside of
> the JDK, this is something which will be constructed, widely used, and
> someone will then propose to add something similar to the JDK.
>
> >While re-implementing the gatherer is a possibility, the existing
> implementation is non-trivial, and creating a custom, robust alternative
> represents a significant undertaking.
>
> The existing version needs to maintain order, which adds to the complexity
> of the implementation. Implementing an unordered version would likely look
> different.
> I'd definitely encourage taking the opportunity to attempt to implement it.
>
> Cheers,
> √
>
>
> *Viktor Klang*
> Software Architect, Java Platform Group
> Oracle
>
> ------------------------------
> *From:* Jige Yu <yujige at gmail.com>
> *Sent:* Monday, 2 June 2025 17:05
> *To:* Viktor Klang <viktor.klang at oracle.com>
> *Cc:* core-libs-dev at openjdk.org <core-libs-dev at openjdk.org>
> *Subject:* Re: [External] : Re: Should mapConcurrent() respect time order
> instead of input order?
>
>
> Thank you for your response and for considering my feedback on the
> mapConcurrent() gatherer. I understand and respect that the final
> decision rests with the JDK maintainers.
>
> I would like to offer a couple of further points for consideration. My
> perspective is that strict adherence to input order for mapConcurrent() might
> not be the most common or beneficial default behavior for users. I'd be
> very interested to see any research or data that suggests otherwise, as
> that would certainly inform my understanding.
>
> From my experience, a more common need is for higher throughput in
> I/O-intensive operations. The ability to support use cases like race()—where
> the first successfully completed operation determines the outcome—also
> seems like a valuable capability that is currently infeasible due to the
> ordering constraint.
>
> As I see it, if a developer specifically requires the input order to be
> preserved, this can be achieved with relative ease by applying a subsequent
> sorting operation. For instance:
>
> .gather(mapConcurrent(...))
> .sorted(Comparator.comparing(Result::getInputSequenceId))
>
> The primary challenge in these scenarios is typically the efficient
> fan-out and execution of concurrent tasks, not the subsequent sorting of
> results.
>
> Conversely, as you've noted, there isn't a straightforward way to modify
> the current default ordered behavior to achieve the higher throughput or
> race() semantics that an unordered approach would naturally provide.
>
> While re-implementing the gatherer is a possibility, the existing
> implementation is non-trivial, and creating a custom, robust alternative
> represents a significant undertaking. My hope was that an unordered option
> could be a valuable addition to the standard library, benefiting a wider
> range of developers.
>
> Thank you again for your time and consideration.
>
>
> On Mon, Jun 2, 2025 at 7:48 AM Viktor Klang <viktor.klang at oracle.com>
> wrote:
>
> >Even if it by default preserves input order, when I explicitly called
> stream.unordered(), could mapConcurrent() respect that and in return
> achieve higher throughput with support for race?
>
> The Gatherer doesn't know whether the Stream is unordered or ordered. The
> operation should be semantically equivalent anyway.
>
> Cheers,
> √
>
>
> *Viktor Klang*
> Software Architect, Java Platform Group
> Oracle
> ------------------------------
> *From:* Jige Yu <yujige at gmail.com>
> *Sent:* Monday, 2 June 2025 16:29
> *To:* Viktor Klang <viktor.klang at oracle.com>; core-libs-dev at openjdk.org <
> core-libs-dev at openjdk.org>
> *Subject:* [External] : Re: Should mapConcurrent() respect time order
> instead of input order?
>
> Sorry. Forgot to copy to the mailing list.
>
> On Mon, Jun 2, 2025 at 7:27 AM Jige Yu <yujige at gmail.com> wrote:
>
> Thanks Viktor!
>
> I was thinking from my own experience that I wouldn't have automatically
> assumed that a concurrent fanout library would by default preserve input
> order.
>
> And I think wanting high throughput with real-life utilities like race
> would be more commonly useful.
>
> But I could be wrong.
>
> Regardless, mapConcurrent() can do both, no?
>
> Even if it by default preserves input order, when I explicitly called
> stream.unordered(), could mapConcurrent() respect that and in return
> achieve higher throughput with support for race?
>
>
>
> On Mon, Jun 2, 2025 at 2:33 AM Viktor Klang <viktor.klang at oracle.com>
> wrote:
>
> Hi!
>
> In a similar vein to the built-in Collectors,
> the built-in Gatherers provide solutions to common stream-related
> problems, but also, they also serve as "inspiration" for developers for
> what is possible to implement using Gatherers.
>
> If someone, for performance reasons, and with a use-case which does not
> require encounter-order, want to take advantage of that combination of
> circumstances, it is definitely possible to implement your own Gatherer
> which has that behavior.
>
> Cheers,
> √
>
>
> *Viktor Klang*
> Software Architect, Java Platform Group
> Oracle
> ------------------------------
> *From:* core-libs-dev <core-libs-dev-retn at openjdk.org> on behalf of Jige
> Yu <yujige at gmail.com>
> *Sent:* Sunday, 1 June 2025 21:08
> *To:* core-libs-dev at openjdk.org <core-libs-dev at openjdk.org>
> *Subject:* Should mapConcurrent() respect time order instead of input
> order?
>
> It seems like for most people, input order isn't that important for
> concurrent work, and concurrent results being in non-deterministic order is
> often expected.
>
> If mapConcurrent() just respect output encounter order:
>
> It'll be able to achieve *higher throughput* if an early task is slow,
> For example, with concurrency=2, and if the first task takes 10 minutes to
> run, mapConcurrent() would only be able to process 2 tasks within the first
> 10 minutes; whereas with encounter order, the first task being slow doesn't
> block the 3rd - 100th elements from being processed and output.
>
> mapConcurrent() can be used to implement useful concurrent semantics, for
> example to *support race* semantics. Imagine if I need to send request to
> 10 candidate backends and take whichever that succeeds first, I'd be able
> to do:
>
> backends.stream()
>     .gather(mapConcurrent(
>         backend -> {
>           try {
>             return backend.fetchOrder();
>            } catch (RpcException e) {
>              return null; // failed to fetch but not fatal
>            }
>         })
>         .filter(Objects::notNull)
>         .findFirst(); // first success then cancel the rest
>
> Cheers,
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/core-libs-dev/attachments/20250602/b3e1519c/attachment-0001.htm>