Bug: graceful shutdowns delayed due to duplicate sleeps.

[jaw-sh]

Describe the bug

Graceful shutdowns take roughly double the expected time per runtime. When graceful_shutdown_timeout_seconds is set to a value higher than the default 5s, this is very noticeable. The code calls thread::sleep(shutdown_timeout) runs after rt.shutdown_timeout(shutdown_timeout), doubling the length of this window, for no obvious reason.

Pingora info

Pingora version: 0.7.0 (current main, but the issue has existed since graceful_shutdown_timeout_seconds was made configurable in PR #99 )
Rust version: cargo 1.86.0
Operating system version: Arch Linux 6.18, also reproduced on Debian 12

Steps to reproduce

1. Configure a Pingora server with graceful_shutdown_timeout_seconds = 30 and grace_period_seconds = 5
2. Start the server with --daemon behind systemd (Type=forking)
3. Send SIGTERM (systemctl stop)
4. Observe that shutdown takes ~65 seconds (5 + 30 + 30) instead of the expected ~35 seconds (5 + 30)

The relevant code is in pingora-core/src/server/mod.rs:

let join = thread::spawn(move || {
    rt.shutdown_timeout(shutdown_timeout);  // blocks up to shutdown_timeout
    thread::sleep(shutdown_timeout)         // then sleeps for shutdown_timeout again
});

Runtime::shutdown_timeout() already blocks the calling thread for up to the full timeout duration.
The thread::sleep afterward just adds dead time.

Expected results

Shutdown should complete in grace_period_seconds + graceful_shutdown_timeout_seconds. With my config that's 35 seconds.

Observed results

Shutdown takes grace_period_seconds + graceful_shutdown_timeout_seconds × 2. With my config that's 65 seconds, which exceeds my systemd TimeoutStopSec and gets the process SIGKILL'd every time.

Additional context

I think this was fine in the original 0.1.0 code where the shutdown timeout was hardcoded to 5 seconds, but when PR #99 made it configurable, the sleep started using the configurable value too, which turns it into a noticeable problem for anyone setting a longer timeout.

I'm currently patching it out locally by just removing the thread::sleep line. Haven't had any issues with that, runtimes shut down cleanly without the extra wait. I am doing daily production changes with my code so if I do not run into any problems with this change I will PR it.

[daviscloudflare]

Looks like a dupe of:

• On shutdown Pingora always sleep for graceful_shutdown_timeout_seconds #727

[jaw-sh]

It is. Are you interested in a PR? It's just one line.