mapConcurrent() with InterruptedException

[Jige Yu]

Checking in on what you've found out, Viktor.

>From where we left off, I understand that you were looking at alternatives
instead of silent truncation?

Have you reached any conclusion?

We touched on disallowing interruption during mapConcurrent(). I still have
concerns with disabling cancellation, because it basically undoes this API
note from the javadoc
<https://cr.openjdk.org/~alanb/sc-20240503/java.base/java/util/stream/Gatherers.html#mapConcurrent(int,java.util.function.Function)>
:

API Note:
In progress tasks will be attempted to be cancelled, on a best-effort
basis, in situations where the downstream no longer wants to receive any
more elements.

In reality, people will use mapConcurrent() to fan out rpcs. Sometimes
these rpcs are just a single blocking call; yet sometimes they may
themselves be a Structured Concurrency scope, with 2 or 3 rpcs that
constitute a single logical operation. Under two conditions, cancellation
is imho important semantic:

   1. The downstream code uses filter().findFirst(), and when it sees an
   element, it will return and no longer needs the other pending rpcs to
   complete. If cancellation is disabled, these unnecessary rpcs will waste
   system resources.
   2. One of the rpc throws and the Stream pipeline needs to propagate the
   exception. Again, if the other rpcs cannot be cancelled, we'll have many
   zombie rpcs.

Zombie rpcs may or may not be a deal breaker, depending on the specific use
case. But for a JDK library, losing cancellation would have a negative
impact on usability.

My 2c,


On Fri, Jan 3, 2025 at 9:18 AM Viktor Klang <viktor.klang at oracle.com> wrote:

> Hi Ben,
>
> Thanks for raising these questions—getting feedback is crucial in the
> Preview stage of features.
>
> I wrote a reply to the Reddit thread so I'll just summarize here:
>
> It is important to note that *mapConcurrent()* is not a part of the
> Structured Concurrency JEPs, so it is not designed to join SC scopes.
>
> I'm currently experimenting with ignoring-but-restoring interrupts on the
> "calling thread" for *mapConcurrent()*, as well as capping
> work-in-progress to *maxConcurrency* (not only capping the concurrency
> but also the amount of completed-but-yet-to-be-pushed work). Both of these
> adjustments should increase predictability of behavior in the face of
> blocking operations with variable delays.
>
> Another adjustment I'm looking at right now is to harden/improve the
> cleanup to wait for concurrent tasks to acknowledge cancellation, so that
> once the finisher is done executing the VTs are known to have terminated.
>
> As for not preserving the encounter order, that would be a completely
> different thing, and I'd encourage you to experiment with that if that
> functionality would be interesting for your use-case(s).
>
> Again, thanks for your feedback!
>
> 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:* Friday, 3 January 2025 17:53
> *To:* core-libs-dev at openjdk.org <core-libs-dev at openjdk.org>
> *Subject:* mapConcurrent() with InterruptedException
>
> Hi Java Experts,
>
> I sent this email incorrectly to loom-dev@ and was told on Reddit that
> core-libs-dev is the right list.
>
> The question is about the behavior of mapConcurrent() when the thread is
> interrupted.
>
> Currently mapConcurrent()'s finisher phase will re-interrupt the thread,
> then stop at whatever element that has already been processed and return.
>
> This strikes me as a surprising behavior, because for example if I'm
> running:
>
>    Stream.of(1, 2, 3)
>         .gather(mapConcurrent(i -> i * 2))
>         .toList()
>
> and the thread is being interrupted, the result could be any of [2], [2,
> 4] or [2, 4, 6].
>
> Since thread interruption is cooperative, there is no guarantee that the
> thread being interrupted will just abort. It's quite possible that it'll
> keep going and then will use for example [2] as the result of doubling the
> list of [1, 2, 3], which is imho incorrect.
>
> In the Reddit
> <https://www.reddit.com/r/java/comments/1hr8xyu/observations_of_gatherersmapconcurrent/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button>
> thread, someone argued that interruption rarely happens so it's more of a
> theoretical issue. But interruption can easily happen in Structured
> Concurrency or in mapConcurrent() itself if any subtask has failed in order
> to cancel/interrupt the other ongoing tasks.
>
> There had been discussion about alternative strategies:
>
>    1. Don't respond to interruption and just keep running to completion.
>    2. Re-interrupt thread and wrap the InterruptedException in a standard
>    unchecked exception (StructuredConcurrencyInterruptedException?).
>
>
> I have concerns with option 1 because it disables cancellation propagation
> when mapConcurrent() itself is used in a subtask of a parent
> mapConcurrent() or in a StructuredConcurrencyScope.
>
> Both equivalent Future-composition async code, or C++'s fiber trees
> support cancellation propagation and imho it's a critical feature or else
> it's possible that a zombie thread is still sending RPCs long after the
> main thread has exited (failed, or falled back to some default action).
>
> My arguments for option 2:
>
>    1. InterruptedException is more error prone than traditional checked
>    exceptions for *users* to catch and handle. They can forget to
>    re-interrupt the thread. It's so confusing that even seasoned programmers
>    may not know they are *supposed to* re-interrupt the thread.
>    2. With Stream API using functional interfaces like Supplier,
>    Function, the option of just tacking on "throws IE" isn't available to many
>    users.
>    3. With Virtual Threads, it will be more acceptable, or even become
>    common to do blocking calls from a stream operation (including but
>    exclusive to mapConcurrent()). So the chance users are forced to deal with
>    IE will become substantially higher.
>    4. Other APIs such as the Structured Concurrency API have already
>    started wrapping system checked exceptions like ExecutionException,
>    TimeoutException in unchecked exceptions ( join()
>    <https://download.java.net/java/early_access/loom/docs/api/java.base/java/util/concurrent/StructuredTaskScope.html#join()> for
>    example).
>    5. Imho, exceptions that we'd rather users not catch and handle but
>    instead should mostly just propagate up as is, should be unchecked.
>
> There is also a side discussion
> <https://www.reddit.com/r/java/comments/1hr8xyu/comment/m4z4f8c/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button>
> about whether mapConcurrent() is better off preserving input order or push
> to downstream as soon as an element is computed. I'd love to discuss that
> topic too but maybe it's better to start a separate thread?
>
> Thank you and cheers!
>
> Ben Yu
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/core-libs-dev/attachments/20250126/55d7264a/attachment.htm>