Improve cache related trace logging.

[sipsma]

This adds trace logs to the solver's cache manager to assist debugging
cache misses+hits.

It also replace the LazyStackTrace struct with a function that returns a
stack trace string only if trace level is enabled, an empty string
otherwise. This makes it easier to use with logrus hooks that pass the
fields around to different goroutines without needing to try to "unlazy"
the fields first.

[jedevc]

I'm not too familiar with logrus internals, but I'm curious as to why this doesn't work in hooks?

It appears that the hooks get called before formatting the logs, within the same call to .Log. So surely to get the stack trace, you'd just need to call .String within the hook (though if the hook does something like spawn a goroutine, I guess that confuses it)?

If we can, I'd rather not add additional indentation for checks for if logrus.GetLevel() == logrus.TraceLevel: instead of doing that, we could preserve the spirit of the original by doing something like:

func lazyStackTrace() string {
	if logrus.GetLevel() == logrus.TraceLevel {
		return string(debug.Stack())
	}
	return ""
}

[crazy-max]

I think we could use a hook like https://github.com/docker/buildx/blob/50fbdd86f928f9ed9c858e401783021264462077/commands/root.go#L47-L53

See https://github.com/docker/buildx/blob/50fbdd86f928f9ed9c858e401783021264462077/util/logutil/filter.go#L10-L38

[sipsma]

So surely to get the stack trace, you'd just need to call .String within the hook (though if the hook does something like spawn a goroutine, I guess that confuses it)?

Yeah my phrasing in the commit message wasn't super accurate, it's not hooks in and of themselves that are the problem but more that the hook has to be sure the field gets "unlazied" before passing it off to a separate goroutine despite all the values being of type any. That's possible but also just sort of a huge gotcha (which is my fault since I added LazyStackTrace, just trying to correct my own mistakes here 😅).

E.g. the relevant line in Dagger is here. I think we'd need to either iterate over all the fields, doing a type assertion and calling .String() one-by-one, or do a full json marshal/unmarshal cycle every time.

Either way, I think your suggestion of func lazyStackTrace() string is much better, I'll update to do that.