caddyhttp: Optimize logs using zap's WithLazy()

[AlliBalliBaba]

After #6541 I realized that using zap's WithLazy() for the error logger was still much more efficient even if we are cloning the request. Originally I wanted to try using this originalRequest, which is made for logging but just gives a shallow copy without request headers. Instead, using request.Clone() gives a deep copy that just omits the request body.

This makes a simple request about 7% more efficient:

:80 {
    respond "Hello World"
}

wrk -t 4 -c 200 -d 60 http://localhost (20 cores Docker Ubuntu WSL)
With() ~205.000 RPS
WithLazy() ~219.000 RPS

Flamegraph With()

Flamegraph WithLazy() and Copy()

Note that this will make the case where an error is logged about 1% slower since we are doing an additional request.Clone()

[francislavoie]

We could possible do the same in reverseproxy.go and fastcgi.go?

[mholt]

That's very interesting -- thanks for submitting this PR!

I second Francis' question, but I think this LGTM either way.

[AlliBalliBaba]

I can look into reverseproxy.go and fastcgi.go if there's similar potential for improvement. Probably something for another PR

[mholt]

Okay, doing it in increments sounds reasonable.

I'll merge this, but before I do, why does cloning the request make it faster? That seems counterintuitive. (I wonder if we should comment next to that line, either a link to this PR or a brief explanation)

[AlliBalliBaba]

Cloning the request is faster than json encoding the request and cloning the logger. But yeah it seems counterintuitive, I'll add a comment.

[mholt]

Thank you! Looking forward to seeing if this works well in any other places.