When a backend service struggles to keep up with traffic, the default industry response is to put a queue in front of it. Under sustained overload, unbounded queues do not prevent failure; they simply delay it, converting a throughput problem into a catastrophic latency spike before inevitably crashing the server or making it unresponsive.
In this talk, we analyze system overload through the lens of Queueing Theory and Systems Thinking. We will examine why buffering is the wrong default structural intervention when the inflow of requests exceeds the outflow of processing capacity.
We will explore how to design systems that structurally reject the "queue to fix overload" anti-pattern. I will how to implement strictly bounded message channels, drop-on-full semantics, and synchronous backpressure signaling in native systems languages. Using patterns from Project Tina (an open-source thread-per-core framework in Odin), we will show how pushing flow-control decisions back to the caller allows a system to predictably shed load, maintain bounded latencies, and stay upright under massive concurrency.