Is there a limit of requests per connection?

[atimin]

Hey all,

I'm using axum as an HTTP front-end for my database. Very often, it has to receive thousands of requests from a client. After some number of requests, the HTTP server drops the connection, and I get a ServerDisconnect error. I use the Python aiohttp library, and people suggest a workaround:

https://stackoverflow.com/a/67641667/1077167

For me, it also works to send "Connection: close" every 1000th request to renew the current connection.

I've tried to find something about the limit for HTTP requests in Hyper and Axum documentation but didn't find any mention. I also tried to change the limit by using Keep-Alive: timeout=Y max=X header, but it looks like it doesn't change the behavior.

Could someone explain how Axum or Hyper manages this limit?

Thanks!

[davidpdrsn]

I'm not aware of such a limit. Feels like it might be the client's fault. Are you able to reproduce it with something like reqwest?

You can also try asking in #hyper in the tokio discord. axum doesn't itself deal with connections directly, that's all delegated to hyper.

[atimin]

Thanks for the response. I've written a Rust program by using reqwest to repeat the issue. You're right, it is something in my Python client , because the Rust program sends thousands requests without any problems.

However, if the server returns an HTTP error for a POST request, reqwest crashes with a Broken pipe even if I use the "Expect: 100-continue" header. I' wrote about it here #2016. I'm getting more confident that this case doesn't handle properly on the server side. I'll follow to your advice and write about it in the tokio community.

[atimin]

Sorry, I wasn't precise. It is not Broken pipe in the Rust client, it is IncompleteMessage:

Error: reqwest::Error { kind: Request, url: Url { scheme: "https", cannot_be_a_base: false, username: "", password: None, host: Some(Domain("play.reduct.store")), port: None, path: "/api/v1/b/datasets/imdb", query: Some("q=236"), fragment: None }, source: hyper::Error(IncompleteMessage) }