How to adjust the keep alive timeout for http1?

[droundy]

I'm observing in my axum server that connections are dropped 1 minute after I get a spike of requests. I already set /proc/sys/net/ipv4/tcp_keepalive_time to 300 seconds from its former value of 60 seconds, but I'm still seeing connections dropped after just 60 seconds.

Is there a way to set this time in hyper and ideally via axum, but right now I'm sort of feeling desperate, as I'm not even sure how to tell whether the timeout is being set in my rust code versus in my system or somehow in the load balancer.

[seanmonstar]

If you mean for your server, hyper has a header read timeout value: https://docs.rs/hyper/latest/hyper/server/conn/http1/struct.Builder.html#method.header_read_timeout. However, that default is 30 seconds.

At 60 seconds, it's more likely that is terminated above, either by something like nginx, an ALB, or even the client's own pool idle timeout.

Note that setting tcp_keepalive_time to a higher number does not do what you think. That value is not how long to keep TCP alive; at the TCP level, it wants to stay alive forever. That option is the amount of time that a connection must be idle before TCP will send extra empty packets to signal along the route that the machine is still alive. So by increasing the time, you will be making it wait longer to send anything, and intermediaries may decide to close the connection as dead, but it also reduces wasted packets if they are not needed.