Most of the books I read about systems are really about computers. Donella Meadows' "Thinking in Systems: A Primer" isn't, and that's exactly why it stuck. It's a slim thing, written for people worried about fisheries and economies and thermostats, and it handed me a vocabulary I didn't realise I'd been missing for fifteen years of running software.
The core idea is stocks and flows. A stock is an amount of something, a queue depth, a connection pool, water in a bath. Flows fill and drain it. Once you start seeing your production system that way, a lot of confusing behaviour stops being mysterious. The reason a service oscillates instead of settling isn't gremlins. It's a feedback loop with a delay in it, where the correction arrives after the thing it was correcting has already moved on.
That word, delay, is the one that rearranged my head. Meadows is brilliant on how delays in a feedback loop cause overshoot and oscillation, and the moment I read it I recognised an autoscaler I'd been fighting. It scaled up hard, the load had already passed by the time the new instances were warm, so it scaled down hard, and round it went. I'd been tuning thresholds. The problem was the delay, and no threshold fixes a delay.
She also has a chapter on leverage points, the places where a small push changes the whole system's behaviour, and how the obvious levers are usually the weak ones. It's the most underlined thing I've read this year. Short, humane, and it'll change how you read a graph that won't sit still.