-
Notifications
You must be signed in to change notification settings - Fork 468
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add receiver.disconnect() for atomic graceful shutdown #959
Conversation
@taiki-e hey, i know you're busy... but, could you take a look? I'm very eager to finish this pr off. |
@taiki-e i see your tagline says |
@taiki-e hi, I'm still interested in pushing this pr into merging. also, now that rust-lang/rust#108164 is merged downstream (rustc), how about cherry-picking the fix into this repo as well? |
@taiki-e hi, another ping. could you spend a little of time to decide whether this api addition could ever be accepted or not at least? thanks in advance. :) |
0f05c78
to
2ef7c29
Compare
Thanks for the PR. I feel this would need investigations of possible designs, their advantages and disadvantages, and how they should interact with new concepts such as reconnection (#750), etc. |
really thanks for replying! however, i don't think there's many possible designs, though? Implementing it was straightforward in terms of adjusting the code base. it's semantically just like the
like the above reasoning, i think they can just co-exist. i.e. you can freely |
Glad to see this happening. I have an application with channels between multiple processes, and clean shutdown without deadlocking requires way too much complexity. (Use case: a metaverse client, which works much like a game client. When the client disconnects from a server, most of the system needs to shut down, but it's not a panic condition. Disconnection leads back to the login screen. There are channels between various threads, sometimes in both directions. When the only way to shut down a connection is to drop the Sender, the workaround is to encapsulate a Sender in Arc<Mutex<Option>>. Then a disconnect can be forced by taking the sender from the option and deleting it. This is clunky. Also, that mutex can result in deadlocks on bounded channels, if the sender blocks.) |
@John-Nagle thanks for sharing your use case. glad that i hear another independent demand for this. Considering past similar issues as well, now that i think there's some demand for this functionality certainly. Also, recently I came up another related concern: dropping hugely-buffered channels. If the channel contains very large number of not- .... so, i think we need ideally the new api would look like this: impl Sender {
fn disconnect(self) -> Option<DisconnectedSender>;
}
impl Receiver {
fn disconnect(self) -> Option<DisconnectedReceiver>;
}
struct DisconnectedSender(...)
struct DisconnectedReceiver(...)
impl DisconnectedSender {
// for the use case where there's `fn finish(self)` on T
// (i.e. non-trivial object cleaning up code outside `::drop()`)
fn into_iter(self) -> impl Iterator;
// ... or no inherent methods?
}
impl DisconnectedReceiver {
fn into_iter(self) -> impl Iterator;
} |
now that #1040 had been released, I'm focusing on this.. After another long thinking session, I think this is the latest api, which is simple and solves everything: impl Sender {
fn disconnect(&self);
fn connect(&self);
}
impl Receiver {
fn disconnect(&self);
fn connect(&self);
} I'll try to create a pr for this. stay tuned. :) |
Closing this in favor of #1114 |
Firstly, congrats for adaptation by
std::sync::mpsc
!Hi, yes... Me again. :) I'm the guy who submitted this stale issue: #852. (i haven't forget it... just priority thing...)
However, I'm eager for this completely unrelated pr to get across the line this time really :)
In short, there's no way to shutdown channels from the receiver-side with guarantee of no loss of messages.
That's because the only way to shutdown channels explicitly is
drop()
-ing the last alivereceiver
and i can't.try_recv()
, now that it's gone. ;)So, there's still a chance sender has managed to sneak more, no matter hard I'm draining with
loop
-edtry_recv()
beforedrop()
at the receiver side in advance. (i.e. tocttou)In other words, I'd like to clarify the semantics of channel shutdown and want to play with many short-lived ones.
This desire is arising from the use-case where I want to commit tasks from many worker threads to the committer thread and the committer is time-sliced and the worker needs to know which current time slice it succeed to commit or not as extremely fast as possible.
So, I'd like to map the time slice to each crossbeam_channels 1-to-1. At the end of any time slice, the workers-side and the committer-side should agree with identical set of committed tasks. Also, it's okay to take some time to switch to new crossbeam channel after first commit error due to expired time slice, which the committer determines and initiates the rollover via
.disconnect()
.So that there's no need for round-trip with
futex
s, which is the overhead for common pattern of passing a sender to return the commit result back. the.disconnect()
approach will just require a few of just atomic CASes.I think there's no performant alternative way, other than trying to persuade 70M-all-time-downloaded crate's maintainer's mind. :)
This pr is still a draft but I'd like to work on more if this sounds sensible (writing tests/supporting all flavors).
Thanks.