Rootconf Mini 2024

TCP achieves reliable communication over an unreliable network (IP). It does so by abstracting the underlying network and observing only the sender and the receiver. This is a cornerstone for the internet’s success and is called the “end-to-end principle”.

As a corollary, it is then possible to impress global behavior by making local measurements and choices regarding reliability - and that is exactly what TCP’s congestion control does - no central repository of metrics, no distributed tracing or structured logging - simple local measurements and decisions made on one machine - converging eventually to make the global inter-network congestion free.

Congestion can happen in webservers too - think about too many requests from the end-user or a degraded downstream dependency causing all requests to pile up, eventually making the service non-responsive to any new request.

So how can we leverage the end-to-end principle in avoiding congestion in our web servers? In this talk I take inspiration from TCP’s congestion control and explain how that can help us build more resilient web services.

There are two important questions I will answer: How can we detect congestion? What can we do to recover from it once we detect it? I will answer both questions using local measurements in web applications that scale globally to a fleet of web servers, just like TCP’s congestion control.