stream: coordinating Tokio 1.0 with the availability of Stream in std

[carllerche]

Currently, Tokio uses the Stream trait provided by futures-core. There is an ongoing effort to define a Stream trait within std. Tokio is currently aiming to release 1.0 by the end of the year. Given the Rust release cadence, Stream may not be available in std on the stable channel by the end of the year.

If Stream is not available on the Rust stable channel by the end of the year, we have two options.

1. Delay Tokio 1.0 until Stream becomes available in std.
2. Remove Stream usage from Tokio until 6 months after it becomes available in std (due to the proposed MSRV policy).

Stream usage

Currently, Stream is implemented in Tokio by:

• mpsc::Receiver
• broadcast::Receiver
• signal::Signal

Tokio also provides a tokio::stream module containing a number of useful stream utilities. A list can be found here.

Removing the Stream dependency

If Stream is removed from the tokio crate, it can be placed temporarily in the tokio-util crate. This crate is not intended to reach 1.0 by the end of year. Stream usage within Tokio is already isolated via the stream feature flag. While useful, it is not a critical dependency.

Removing Stream from tokio would impact the discoverability of the utilities. This could be mitigated by including an empty tokio::stream module with documentation explaining where to find the utilities.

Adding Stream post 1.0

Adding Stream after an initial 1.0 release would take at least 6 months given the currently proposed MSRV policy (#2718). According to the policy, 1.x releases are permitted to increase the MSRV, but must support all Rust versions released in the past 6 months.

Next steps

I will leave this open-ended for now as we collect comments. We can then decide on the direction we want to take.

[sfackler]

I don't know what the current thought is, but there was an idea that the futures crate could start transparently reexporting std's Stream when being built against a suitable version to avoid forcing a big migration.

[taiki-e]

See also rust-lang/futures-rs#2207

[carllerche]

@taiki-e thanks for that. My understanding is that it would still technically be a breaking change. My gut is that I would rather start a 1.0 stability guarantee on strong footing.

[davidbarsky]

For exhaustiveness, I think there that there's a third option: a six month MSRV policy can take effect well after a 1.0 is released. I'm not convinced that a Tokio 1.0 release needs to be coupled to a six month MSRV policy at release. This is something that can be added retroactively.

[taiki-e]

@carllerche Yeah, that's technically breaking change. AFAIK, theoretically, if the signature of the std Stream trait doesn't change after futures_core v1 released, and the build script works, and the user doesn't accidentally rely on the unstable Stream trait (that requires#![feature]), no breakage will occur. But, in any case, when defining Stream in futures_core, the risk cannot be zero.

FYI: I've created a demo of how that's implemented: https://github.com/taiki-e/futures-compat-experiment

[taiki-e]

Perhaps another idea for this would be to use a (disabled by default) feature flag that includes the version number: like stream03, stream1.
(Once std Stream is stable, make the stream1 feature to no-op and always implement the stream from futures_core. If the implementation of futures_core's Stream is different from std's Stream, we can implement std's Stream directly as they don't conflict.)

[Keruspe]

Related: rust-lang/rust#79023

[carllerche]

As of this comment, Stream trait RFC is not yet merged. I expect it is mostly ready to go, but I am not exactly clear on the expected timeline.

If Stream lands in stable sometime in January 2021, I think we could consider delaying the 1.0 release, but probably not more than that. However, I would expect Stream to not be available in std until later.

Assuming that Stream lands in stable February 2021 or later, I propose:

• Extract the contents of tokio::stream into the tokio-stream crate.
• Leave an empty tokio::stream module documenting that Stream utilities exist in tokio-stream and will be merged back into tokio once the Stream trait lands in std.
• Remove Stream impls for misc types and document how to use async-stream to get your own stream implementation. These types already have inherent fns to consume and do not depend on their Stream implementations.
  • tokio::fs::ReadDir
  • tokio::io::Lines
  • tokio::net::{TcpListener, UnixListener}
  • tokio::signal::{CtrlC, CtrlBreak}
  • tokio::sync::broadcast
  • tokio::time::Interval

Once Stream lands in std, we add tokio::stream back. However, we must do this in a way that preserves the Tokio's MSRV guarantee of ~6 months. To do this, we can add a build script that detects the Rust version. If using a version of Rust that includes Stream, then re-export that Stream and use it. Otherwise, define a Stream trait that matches the version in std and use that one. This is a similar strategy as the one proposed for futures-rs but does not introduce Stream until after Stream is in the std stable channel. This prevents potential breaking changes with mismatched Stream versions