2015-04-23 0:32 GMT+02:00 Vicente J. Botet Escriba <vicente.botet@wanadoo.fr>:
Le 22/04/15 14:28, Konrad Zemek a écrit :
Hi,
When creating a continuation future with future<>::then(), and the policy is boost::launch::async, the created future's destructor blocks, like a future created by boost::async, and probably for the same reasons (as outlined by the standardization committee in n3679).
The same behavior is coded into continuation futures' destructor when using an executor. Why is it needed here? The situation with executor-run continuations is a little different, as even when the future is destroyed, the lifetime of running job is still bound to the executor.
An Executor can also create a thread for each task, and the Executor destructor don't need to wait until all the tasks have finished (even if some can do it). I don't see how to make the difference without making the Executor concept more complex.
You're right, I erroneously assumed that executors have to join on destruction much like a future returned from async. Re-reading the proposal I see that join behavior is specified for a concrete executor, not the general executor concept.
Note however that, I have in my todo list an implementation that doesn't block at all. Just that everything need more time than we have. This would mean a breaking change, and all of us know how annoying is to introduce breaking changes.
My use case: I've coded a network communication stack that returns to the caller a future<NetworkResponse>. I've previously used std::async(std::launch::deferred, ...) to transform the future's content before returning it to the caller (bytes -> protobuf -> internal object), and I consider such manipulation of a future value to be a very powerful feature. I've used std::launch::deferred to reduce the number of running threads, but the downside is that the client can't wait for the future value with a timeout. On the other side of the spectrum is std::launch::async, which would run a new thread - per pending communication - that would do little more than block. boost::future<>::then is a fantastic fit for my use case, as I can use my boost::asio::io_service as an executor and let my communication stack's threads do the work without having them block on future.get() to first retrieve the result. The caller can then call future<>::wait_for() to wait for the reply with a custom timeout. This being network communication, though, a reply message may never arrive - the corresponding promise will eventually be destroyed by the stack, but I can't have users block on the future's destructor until that happens, after they have already decided that the reply is no longer worth waiting for.
I would expect that the program must take care of the lost of communication and do a set_exception on the promise before the promise is destroyed. However, the call to wait_for seems to be a good hint that the user knows what is doing. I could consider that any call to a timed wait function disable blocking. An alternative could be to have a way to request to don't block. Please let me know if one of these options would take in account your use case.
Either of these options would work for me, as I could emulate a "don't block" request with a call to future<>::wait_for(0). I'd prefer the latter to the former though, if only because it fits my needs better and it would make it more explicit that future's behavior is modified.
For the time being (and very temporarily) , if you don't want anymore this blocking future, you can move it to a list of detached futures.
Please advise if there's a workaround for this behavior that doesn't involve me distributing a custom version of Boost.Thread with my binaries? :)
You can send any PR you consider improves the behavior of the library.
Of course; I just prefer to discuss it first to find out if others consider such a change to be an improvement. :) Konrad