I read Donella Meadows' Thinking in Systems a few years late and a few years too lightly the first time. It sat on the shelf looking earnest. I came back to it after an incident that nobody could explain, where every component was healthy and the whole thing still fell over, and suddenly the book had teeth.
The idea that stuck is stocks and flows. A queue isn't a number on a dashboard, it's a stock, and what matters is the rate going in versus the rate coming out. If inflow beats outflow, the queue grows no matter how healthy each worker looks. I'd been staring at the workers. The book made me stare at the loops instead.
The other one is delays. Feedback that arrives late makes systems oscillate, and we build late feedback everywhere: a five-minute metrics scrape, an autoscaler that reacts to load from three minutes ago, a deploy whose effect you don't see until traffic peaks. Half the "weird" behaviour I've chased was just a delay in a loop, overshooting and correcting.
It's not a technical book and it's better for it. It gave me vocabulary for things I'd felt but couldn't name, and now I can't un-see them in a postmortem. Worth the reread.