Back when I was an undergraduate, as part of a class called “Computer Systems Engineering”, we read numerous classic papers of systems design. I enjoyed and learned a great deal from many of these papers, but one that paper that has stuck with me in particular was Saltzer et al’s “End-to-End Arguments in Systems Design”. The paper is a very general tract on systems design – it does explore several examples of concrete systems or applications, but it ultimately expounds upon the end-to-end principle as a perspective or design heuristic that can apply to virtually any system design.
I recommend reading the paper – it’s short and very readable – but in brief, it argues that many functions of a system are better addressed at the “ends” of the system, instead of at each lower-level interface boundary. As a concrete example, the paper argues that a file transfer system is better served by assuring correctness end-to-end by way of strong checksums and retries as needed, than by insisting on perfect lossless transfer out of each of its underlying network layers.
The longer I work professionally in software and design and work with systems, the more I find myself reflecting on the end-to-end argument and its nuance and implications for sytems design. I want to here commit to writing two different perspectives I find particularly interesting.
To understand either perspective, we’ll need to first reflect on what we mean by a “system” and “system design.” Merriam-Webster defines a “system” as “a regularly interacting or interdependent group of items forming a unified whole”. The essential character of a system, then, is that it is a unit that decomposes into some number of sub-components, which can be understood or considered substantially in isolation of each other. Systems design, then, is the process of performing this decomposition: of deciding how to break a desired system into individual units that can be built up into a functioning whole.
The Depressing Take: You Can’t Compose Correctness
One property we might hope for in systems design is that we can, in some sense, arrive at a correct system merely by properly composing correct subsystems. This hope is sometimes phrased in terms of desiring “LEGO blocks” for software design: If we can just design the right, robust, fundamental building blocks of software design, we’ll be able to snap them together in arbitrary ways and design complex systems at low cost and effort.
The end-to-end argument challenges this optimism directly. No matter the sophistication of the underlying building blocks, it argues, we’ll always have to define and enforce the essential correctness properties of our system at the topmost end-to-end layer of design. We can’t trivially derive correctness from the correctness of our subsystems: we must always consider it as an end-to-end property.
The end-to-end argument thus also speaks to a lack of scale-invariance in abstraction and systems design. Systems that own the “ends” are subject to the end-to-end principle and responsible for enforcing correctness. Conversely, systems that exist only as subsystems or components of a larger whole may not have such responsibility, and can delegate certain guarantees to the end-to-end system outside of their scope.
This scale-dependence, therefore, implies a certain lack of “fundamentally correct” abstractions in the world. If we want to design a component for a certain class of functionality, the appropriate design constraints and guarantees depend on whether we’re building an end-to-end system or merely a component of a larger one. And, if we’re designing a component, which guarantees we must provide depends on which properties the top-level system is prepared to enforce itself.
In this regard, I think this take on the end-to-end argument is a close peer to the Law of Leaky Abstractions, which states that all sufficiently-complex abstractions are to some degree leaky. Both principles express a certain inevitability for “out-of-band” interactions between higher levels of the abstraction stack and the implementation details of the underlying system.
This is the perspective that struck me most strongly when I first read the paper in college. I found it fairly disheartening at the time; As a young enthusiastic aspiring systems designer, I wanted to believe in the existence of some set of idealized platonic abstractions for fundamental capabilities, which, if we could just extract them from the raw ether of software stuff, would seamlessly and forever raise the level of abstraction at which software designers work. The end-to-end principle is a statement that the world is messy, and that while there may exist “good-enough” abstractions to be nearly-universal, we’ll forever be carrying complexity with us to higher and higher levels of the systems design stack.
The Optimistic Take: The TCB Perspective
This second perspective is one that I’ve appreciated more and more as I’ve done actual systems design, and I think is closer to the wisdom the paper’s authors hoped to impart.
Correctness is hard. Anyone who’s worked for long in software is well familiar with the reality that it’s dramatically easier to write software that works most of the time than software that works all of the time. This truism is only more true in large systems, where we must interact with external components that themselves may have uncertain reliability properties.
The end-to-end argument encourages us to accept and embrace this reality. Instead of demanding absolute correctness from every part of our system, we can choose some essential correctness properties (e.g.: messages are copied unmodified from point A to point B; every transaction appears exactly once in our ledger) and to locate those properties within a subset of our system (at the “ends”). Once we’ve done this design step, we can demand correctness from these end-to-end components, and treat the rest of the system as more of an optimization problem.
I call this the “TCB” perspective because I liken it to the notion of a “trusted computing base” in operating systems theory. We seek to reduce the portions of the system for which bugs directly threaten the integrity of the entire system; having done so, we can then develop the rest of the system safe in the confidence that errors may manifest as performance problems or even as detectable faults in the overall system, but that it should be impossible to silently cause catastrophic failure.
Systems design is intricate, and full of tradeoffs, tools, tricks, and heuristics for how to assemble complex systems out of components. I really love the end-to-end argument, and this paper about it, because I think it’s one of our best “top-down” tools for systems design. Many techniques and principles focus on “building up” – taking components, composing them, architecting up from the implementation primitives available to us. But at the end of the day, our goal is to solve a concrete problem with an overall system, and the end-to-end argument is a powerful tool to relate this overall goal to the decomposition of the system into components and modules.
This perspective, of working top-down and relating end-to-end function to component design, can be frustrating, as it means we may need to visit every systems design problem substantially anew; components that look similar between two problems may have importantly different functional requirements depending on the ultimate goal. At the same time, however, it gives us a powerful tool and structure for doing this decomposition and thinking about systems design.