disable connection linger timeout by default

[fwbrasil]

Problem

Tapir sets the SO_LINGER socket option by default. This config holds the connection for some time after it's closed and Netty uses a global thread pool (GlobalEventExecutor) with a single ephemeral thread to delay the connection close. Netty doesn't set this config by default and I couldn't identify a mechanism to make it use another thread pool.

I noticed this issue when benchmarking with a high rate of short-lived connections, which ended up generating a backlog in this thread pool and memory pressure. The GlobalEventExecutor also keeps constantly tearing down its thread and creating new ones, which introduces additional overhead.

Solution

Don't set the configuration by default.

Notes

• There can be cases where the config makes sense but I'd say it's important that it is done by an expert considering the characteristics of specific systems.

[adamw]

I'm trying to understand more about the setting, and it seems that by default SO_LINGER should be set to a non-zero value. What they recommend on SO is having the client initiate connection close - which, I guess, happens in HTTP (unless there's some error server-side)? But then I don't understand why your benchmark would behave the way it does. Anybody knows more about sockets & linger? ;)

[kciesielski]

@adamw I read the same thread and this response seems interesting. It suggests that there's a distinction between using netty as a client and as a server, and in the latter case disabling linger seems ok?

[adamw]

@kciesielski Ah makes sense, thanks! :)