Don't block the runtime
Let's circle back to yield points.
Unlike threads, Rust tasks cannot be preempted.
tokio
cannot, on its own, decide to pause a task and run another one in its place.
The control goes back to the executor exclusively when the task yields—i.e.
when Future::poll
returns Poll::Pending
or, in the case of async fn
, when
you .await
a future.
This exposes the runtime to a risk: if a task never yields, the runtime will never be able to run another task. This is called blocking the runtime.
What is blocking?
How long is too long? How much time can a task spend without yielding before it becomes a problem?
It depends on the runtime, the application, the number of in-flight tasks, and many other factors. But, as a general rule of thumb, try to spend less than 100 microseconds between yield points.
Consequences
Blocking the runtime can lead to:
- Deadlocks: if the task that's not yielding is waiting for another task to complete, and that task is waiting for the first one to yield, you have a deadlock. No progress can be made, unless the runtime is able to schedule the other task on a different thread.
- Starvation: other tasks might not be able to run, or might run after a long delay, which can lead to poor performances (e.g. high tail latencies).
Blocking is not always obvious
Some types of operations should generally be avoided in async code, like:
- Synchronous I/O. You can't predict how long it will take, and it's likely to be longer than 100 microseconds.
- Expensive CPU-bound computations.
The latter category is not always obvious though. For example, sorting a vector with a few elements is not a problem; that evaluation changes if the vector has billions of entries.
How to avoid blocking
OK, so how do you avoid blocking the runtime assuming you must perform an operation
that qualifies or risks qualifying as blocking?
You need to move the work to a different thread. You don't want to use the so-called
runtime threads, the ones used by tokio
to run tasks.
tokio
provides a dedicated threadpool for this purpose, called the blocking pool.
You can spawn a synchronous operation on the blocking pool using the
tokio::task::spawn_blocking
function. spawn_blocking
returns a future that resolves
to the result of the operation when it completes.
#![allow(unused)] fn main() { use tokio::task; fn expensive_computation() -> u64 { // [...] } async fn run() { let handle = task::spawn_blocking(expensive_computation); // Do other stuff in the meantime let result = handle.await.unwrap(); } }
The blocking pool is long-lived. spawn_blocking
should be faster
than creating a new thread directly via std::thread::spawn
because the cost of thread initialization is amortized over multiple calls.
Further reading
- Check out Alice Ryhl's blog post on the topic.
Exercise
The exercise for this section is located in 08_futures/05_blocking