Join GitHub today
GitHub is home to over 40 million developers working together to host and review code, manage projects, and build software together.
Sign uptask: Introduce a new pattern for task-local storage #2126
Conversation
This PR introduces a new pattern for task-local storage. It allows for storage
and retrieval of data in an asynchronous context. It does so using a new pattern
based on past experience.
A quick example:
```rust
tokio::task_local! {
static FOO: u32;
}
FOO.scope(1, async move {
some_async_fn().await;
assert_eq!(FOO.get(), 1);
}).await;
```
## Background of task-local storage
The goal for task-local storage is to be able to provide some ambiant context in
an asynchronous context. One primary use case is for distributed tracing style
systems where a request identifier is made available during the context of a
request / response exchange. In a synchronous context, thread-local storage
would be used for this. However, with asynchronous Rust, logic is run in a
"task", which is decoupled from an underlying thread. A task may run on many
threads and many tasks may be multiplexed on a single thread. This hints at the
need for task-local storage.
### Early attempt
Futures 0.1 included a [task-local storage][01] strategy. This was based around
using the "runtime task" (more on this later) as the scope. When a task was
spawned with `tokio::spawn`, a task-local map would be created and assigned
with that task. Any task-local value that was stored would be stored in this
map. Whenever the runtime polled the task, it would set the task context
enabling access to find the value.
There are two main problems with this strategy which ultimetly lead to the
removal of runtime task-local storage:
1) In asynchronous Rust, a "task" is not a clear-cut thing.
2) The implementation did not leverage the significant optimizations that the
compiler provides for thread-local storage.
### What is a "task"?
With synchronous Rust, a "thread" is a clear concept: the construct you get with
`thread::spawn`. With asynchronous Rust, there is no strict definition of a
"task". A task is most commonly the construct you get when calling
`tokio::spawn`. The construct obtained with `tokio::spawn` will be referred to
as the "runtime task". However, it is also possible to multiplex asynchronous
logic within the context of a runtime task. APIs such as
[`task::LocalSet`][local-set] , [`FuturesUnordered`][futures-unordered],
[`select!`][select], and [`join!`][join] provide the ability to embed a mini
scheduler within a single runtime task.
Revisiting the primary use case, setting a request identifier for the duration
of a request response exchange, here is a scenario in which using the "runtime
task" as the scope for task-local storage would fail:
```rust
task_local!(static REQUEST_ID: Cell<u64> = Cell::new(0));
let request1 = get_request().await;
let request2 = get_request().await;
let (response1, response2) = join!{
async {
REQUEST_ID.with(|cell| cell.set(request1.identifier()));
process(request1)
},
async {
REQUEST_ID.with(|cell| cell.set(request2.identifier()));
process(request2)
},
};
```
`join!` multiplexes the execution of both branches on the same runtime task.
Given this, if `REQUEST_ID` is scoped by the runtime task, the request ID would
leak across the request / response exchange processing.
This is not a theoretical problem, but was hit repeatedly in practice. For
example, Hyper's HTTP/2.0 implementation multiplexes many request / response
exchanges on the same runtime task.
### Compiler thread-local optimizations
A second smaller problem with the original task-local storage strategy is that
it required re-implementing "thread-local storage" like constructs but without
being able to get the compiler to help optimize. A discussion of how the
compiler optimizes thread-local storage is out of scope for this PR description,
but suffice to say a task-local storage implementation should be able to
leverage thread-locals as much as possible.
## A new task-local strategy
Introduced in this PR is a new strategy for dealing with task-local storage.
Instead of using the runtime task as the thread-local scope, the proposed
task-local API allows the user to define any arbitrary scope. This solves the
problem of binding task-locals to the runtime task:
```rust
tokio::task_local!(static FOO: u32);
FOO.scope(1, async move {
some_async_fn().await;
assert_eq!(FOO.get(), 1);
}).await;
```
The `scope` function establishes a task-local scope for the `FOO` variable. It
takes a value to initialize `FOO` with and an async block. The `FOO` task-local
is then available for the duration of the provided block. `scope` returns a new
future that must then be awaited on.
`tokio::task_local` will define a new thread-local. The future returned from
`scope` will set this thread-local at the start of `poll` and unset it at the
end of `poll`. `FOO.get` is a simple thread-local access with no special logic.
This strategy solves both problems. Task-locals can be scoped at any level and
can leverage thread-local compiler optimizations.
Going back to the previous example:
```rust
task_local! {
static REQUEST_ID: u64;
}
let request1 = get_request().await;
let request2 = get_request().await;
let (response1, response2) = join!{
async {
let identifier = request1.identifier();
REQUEST_ID.scope(identifier, async {
process(request1).await
}).await
},
async {
let identifier = request2.identifier();
REQUEST_ID.scope(identifier, async {
process(request2).await
}).await
},
};
```
There is no longer a problem with request identifiers leaking.
## Disadvantages
The primary disadvantage of this strategy is that the "set and forget" pattern
with thread-locals is not possible.
```rust
thread_local! {
static FOO: Cell<usize> = Cell::new(0);
}
thread::spawn(|| {
FOO.with(|cell| cell.set(123));
do_work();
});
```
In this example, `FOO` is set at the start of the thread and automatically
cleared when the thread terminates. While this is nice in some cases, it only
really logically makes sense because the scope of a "thread" is clear (the
thread).
A similar pattern can be done with the proposed stratgy but would require an
explicit setting of the scope at the root of `tokio::spawn`. Additionally, one
should only do this if the runtime task is the appropriate scope for the
specific task-local variable.
Another disadvantage is that this new method does not support lazy initialization
but requires an explicit `LocalKey::scope` call to set the task-local value. In
this case since task-local's are different from thread-locals it is fine.
[01]: https://docs.rs/futures/0.1.29/futures/task/struct.LocalKey.html
[local-set]: #
[futures-unordered]: https://docs.rs/futures/0.3.1/futures/stream/struct.FuturesUnordered.html
[select]: https://docs.rs/futures/0.3.1/futures/macro.select.html
[join]: https://docs.rs/futures/0.3.1/futures/macro.join.html
This comment has been minimized.
This comment has been minimized.
|
Quick question: while in a "scope", if I |
This comment has been minimized.
This comment has been minimized.
|
@seanmonstar That is not directly possible to do since that would require us to know dynamically all the scopes set on the stack. The real solution is as a user to know what task locals you might want to propagate and set them explicitly on spawn. tokio::task_local! {
static FOO: u32;
}
FOO.scope(1, async move {
tokio::spawn(FOO.scope(FOO.get(), async move {
assert_eq!(FOO.get(), 1);
});
}).await;Maybe, we can add an API that can do the Also, this is non-trivial when it comes to types that don't implement I think this is ok to keep it this way since this is task locals and this behavior is the same with thread locals. |
This comment has been minimized.
This comment has been minimized.
|
It might make more sense to factor in "inheritable" locals w/ "structured concurrency" (#1879). Thoughts @Matthias247? |
This comment has been minimized.
This comment has been minimized.
pandaman64
commented
Jan 17, 2020
|
I think the thread-local approach is definitely better for Tokio considering safety now, but hopefully std's |
| ($(#[$attr:meta])* $vis:vis $name:ident, $t:ty) => { | ||
| static $name: $crate::task::LocalKey<$t> = { | ||
| std::thread_local! { | ||
| static __KEY: std::cell::RefCell<Option<$t>> = std::cell::RefCell::new(None); |
This comment has been minimized.
This comment has been minimized.
Matthias247
Jan 17, 2020
I'm not super familiar with how thread-locals are implemented. Do they when defined use memory on every thread - even if that one doesn't make use of them? Or is it lazily allocated? If it requires memory everywhere then adding lots of task-locals will obviously make it worse. But on the other hand I don't expect people to use too many task-locals with this kind declaration mechanism.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
Matthias247
commented
Jan 17, 2020
These challenges seem to come up in multiple places: tasks locals, tracing (which is related), running concurrently with a Or whether we rather try to keep the task concept clean as "an individual work strand directly scheduled on the executor". Which could e.g. be done by discouraging to use So that was just a random thought - since it actually does not matter a lot for the sake of this PR :-)
I think Oh, and on the question whether task-locals should be supported at all: I'm rather a fan of explicit parameter passing than implicit. It's easier to understand and easier to test. And actually for each parameter someone could argue that it's more convenient not to pass it - and we end up with all global variables in the end |
|
Looks great to me |
LucioFranco commentedJan 16, 2020
•
edited
Introduce a new pattern for task-local storage
This PR introduces a new pattern for task-local storage. It allows for storage
and retrieval of data in an asynchronous context. It does so using a new pattern
based on past experience. This API is similar to the one from
stdexcept forchanges that were needed to be made due to the differences required by async
code.
A quick example:
Background of task-local storage
The goal for task-local storage is to be able to provide some ambiant context in
an asynchronous context. One primary use case is for distributed tracing style
systems where a request identifier is made available during the context of a
request / response exchange. In a synchronous context, thread-local storage
would be used for this. However, with asynchronous Rust, logic is run in a
"task", which is decoupled from an underlying thread. A task may run on many
threads and many tasks may be multiplexed on a single thread. This hints at the
need for task-local storage.
Early attempt
Futures 0.1 included a task-local storage strategy. This was based around
using the "runtime task" (more on this later) as the scope. When a task was
spawned with
tokio::spawn, a task-local map would be created and assignedwith that task. Any task-local value that was stored would be stored in this
map. Whenever the runtime polled the task, it would set the task context
enabling access to find the value.
There are two main problems with this strategy which ultimately lead to the
removal of runtime task-local storage:
compiler provides for thread-local storage.
What is a "task"?
With synchronous Rust, a "thread" is a clear concept: the construct you get with
thread::spawn. With asynchronous Rust, there is no strict definition of a"task". A task is most commonly the construct you get when calling
tokio::spawn. The construct obtained withtokio::spawnwill be referred toas the "runtime task". However, it is also possible to multiplex asynchronous
logic within the context of a runtime task. APIs such as
task::LocalSet,FuturesUnordered,select!, andjoin!provide the ability to embed a minischeduler within a single runtime task.
Revisiting the primary use case, setting a request identifier for the duration
of a request response exchange, here is a scenario in which using the "runtime
task" as the scope for task-local storage would fail:
join!multiplexes the execution of both branches on the same runtime task.Given this, if
REQUEST_IDis scoped by the runtime task, the request ID wouldleak across the request / response exchange processing.
This is not a theoretical problem, but was hit repeatedly in practice. For
example, Hyper's HTTP/2.0 implementation multiplexes many request / response
exchanges on the same runtime task.
Compiler thread-local optimizations
A second smaller problem with the original task-local storage strategy is that
it required re-implementing "thread-local storage" like constructs but without
being able to get the compiler to help optimize. A discussion of how the
compiler optimizes thread-local storage is out of scope for this PR description,
but suffice to say a task-local storage implementation should be able to
leverage thread-locals as much as possible.
A new task-local strategy
Introduced in this PR is a new strategy for dealing with task-local storage.
Instead of using the runtime task as the thread-local scope, the proposed
task-local API allows the user to define any arbitrary scope. This solves the
problem of binding task-locals to the runtime task:
The
scopefunction establishes a task-local scope for theFOOvariable. Ittakes a value to initialize
FOOwith and an async block. TheFOOtask-localis then available for the duration of the provided block.
scopereturns a newfuture that must then be awaited on.
tokio::task_localwill define a new thread-local. The future returned fromscopewill set this thread-local at the start ofpolland unset it at theend of
poll.FOO.getis a simple thread-local access with no special logic.This strategy solves both problems. Task-locals can be scoped at any level and
can leverage thread-local compiler optimizations.
Going back to the previous example:
There is no longer a problem with request identifiers leaking.
Disadvantages
The primary disadvantage of this strategy is that the "set and forget" pattern
with thread-locals is not possible.
In this example,
FOOis set at the start of the thread and automaticallycleared when the thread terminates. While this is nice in some cases, it only
really logically makes sense because the scope of a "thread" is clear (the
thread).
A similar pattern can be done with the proposed stratgy but would require an
explicit setting of the scope at the root of
tokio::spawn. Additionally, oneshould only do this if the runtime task is the appropriate scope for the
specific task-local variable.
Another disadvantage is that this new method does not support lazy initialization
but requires an explicit
LocalKey::scopecall to set the task-local value. Inthis case since task-local's are different from thread-locals it is fine.
Overall, I think this is a much better improvement over what we originally had
in futures 0.1. This version is also much easier to work with and reason about!