Reputation: 549
I have a program which generates data slowly (we can say it's computationally intensive, like computing digits of pi). It produces a lot of data; each response can be 1GiB, will not fit in memory, and must be generated on demand. I'm using hyper to write a web service to generate the content when requested.
Let's skip the boilerplate (service_fn
, Server::bind
).
The API which generates the data slowly might be something like
use std::io;
impl SlowData {
fn new(initial: &str) -> SlowData {
unimplemented!()
}
fn next_block(&self) -> io::Result<&[u8]> {
unimplemented!()
}
}
type ResponseFuture = Box<Future<Item = Response, Error = GenericError> + Send>;
fn run(req: Request) -> ResponseFuture {
// spawn a thread and:
// initialize the generator
// SlowData::new(&req.uri().path());
// spawn a thread and call slow.next_block() until len()==0
// each byte which comes from next_block should go to the client
// as part of the Body
}
Note that SlowData::new
is also computationally intensive.
Optimally, we'd minimize the copies and send that &[u8]
directly to hyper without having to copy it into a Vec
or something.
How do I fulfill a hyper Request's body from a side thread?
Upvotes: 5
Views: 3056
Reputation: 430270
Spin up a thread in a thread pool and send chunks of data across a channel. The channel implements Stream
and a hyper Body
can be constructed from a Stream
using wrap_stream
:
use futures::{channel::mpsc, executor::ThreadPool, task::SpawnExt, SinkExt, Stream}; // 0.3.1, features = ["thread-pool"]
use hyper::{
service::{make_service_fn, service_fn},
Body, Response, Server,
}; // 0.13.1
use std::{convert::Infallible, io, thread, time::Duration};
use tokio; // 0.2.6, features = ["macros"]
struct SlowData;
impl SlowData {
fn new(_initial: &str) -> SlowData {
thread::sleep(Duration::from_secs(1));
Self
}
fn next_block(&self) -> io::Result<&[u8]> {
thread::sleep(Duration::from_secs(1));
Ok(b"data")
}
}
fn stream(pool: ThreadPool) -> impl Stream<Item = io::Result<Vec<u8>>> {
let (mut tx, rx) = mpsc::channel(10);
pool.spawn(async move {
let sd = SlowData::new("dummy");
for _ in 0..3 {
let block = sd.next_block().map(|b| b.to_vec());
tx.send(block).await.expect("Unable to send block");
}
})
.expect("Unable to spawn thread");
rx
}
#[tokio::main]
async fn main() {
// Construct our SocketAddr to listen on...
let addr = ([127, 0, 0, 1], 3000).into();
// Create a threadpool (cloning is cheap)...
let pool = ThreadPool::new().unwrap();
// Handle each connection...
let make_service = make_service_fn(|_socket| {
let pool = pool.clone();
async {
// Handle each request...
let svc_fn = service_fn(move |_request| {
let pool = pool.clone();
async {
let data = stream(pool);
let resp = Response::new(Body::wrap_stream(data));
Result::<_, Infallible>::Ok(resp)
}
});
Result::<_, Infallible>::Ok(svc_fn)
}
});
// Bind and serve...
let server = Server::bind(&addr).serve(make_service);
// Finally, run the server
if let Err(e) = server.await {
eprintln!("server error: {}", e);
}
}
When creating a thread, there's no way to avoid copying the slice to a Vec
.
See also:
Upvotes: 6