RyanFrantz

Systems Thinking in Practice

I recently finished reading Donella H. Meadows' Thinking in Systems. [1]

I really enjoyed this book. While I thought that I approached my work as an engineer holistically, Meadows' introduction to systems thinking filled in many gaps.

After an examination of what systems are (at the basic level: stocks, flows, and feedback loops), how they operate, and how small systems can be comprised into larger systems, she enumerates a set of leverage points that can be used to increase desirable behavior and decrease undesirable outcomes within systems.

Meadows is very candid in stating that the presented list is neither authoritative nor comprehensive. Nonetheless, she makes a strong case for them (and their order of impact). Whenever I learn something new, I find I can internalize the content better if I can find a way to apply it in some tangible fashion.

So it is that I found myself reflecting on Meadow's leverage points and trying to find corollaries within my areas of expertise. What follows is Meadows' list of leverage points outfitted with those my ideas of where or how they can be applied to software development and web operations.

Leverage Points

Meadows presented her leverage points in ascending order, from least impactful to most. I will do the same.

NOTE: In no way is the list of ideas below either authoritative or comprehensive.

12. Numbers

These are values that can be easily changed and typically produce some immediate and short-term effect in a system. The sorts of knobs that came to mind were feature flags, Apache worker count, node weights in a load balancer.

11. Buffers

In Meadows' book, buffers refer to "the size of stocks relative to their flows[.]" Honestly, of all the leverage points, this one is the most immediately understood and applied. We use some form of memory/storage to contain objects that are frequently requested in tools such as memcache and Varnish. Our network interfaces buffer traffic so that it can be processed in more efficient batches rather than one-by-one. An interesting lever we have, with respect to caches, is the TTL. We can persist objects in a cache for as long as we need. That, in tandem with the available storage within a cache, controls how our buffers behave.

10. Stock-and-Flow Structures

Here, we are concerned with the existence and availability of the physical components that comprise our systems. Outside of computing, Meadows' statement that "changing [them] is rarely quick or simple" is very likely true (i.e. house plumbing, highway infrastructure). In web operations, the high accessibiliy of computing resources in the form of cloud services make spinning up web infrastructure near instanteneous.

9. Delays

"Delays in feedback loops are critical determinants of system behavior." When I consider what we have and use as feedback loops to operate our systems, I think of the metric collection and display tools we use; the monitoring tools we use to gauge system and service health. I see possible delays in the chosen polling frequency of these tools and in the confidence we have that they are doing their jobs, helping us build mental models of what is happening.

Top-down management came to mind as well. Such a structure doesn't move as fast, in terms of making decisions, as does a group of engineers at the edges. Dekker would refer to the latter group of practioners as being on the "sharp end" where they have more immediate impact. Part of that impact is governed by the fact that they will quickly receive and act on the consequences of the decisions they make.

I also imagined a delay in the time it takes new technology to be introduced, to the time it is adopted in an organization, to the point at which that organization operates the technology with some degree of (agreed upon) confidence.

8. Balancing Feedback Loops

The role of the balancing feedback loop is to aid the system in self-correcting. Self-correction can mean many different things, given the context in which such a loop exists. Says Meadows, "[t]he strength of a balancing feedback loop is important relative to the impact it is designed to correct." This is where we build (sometimes discover) resilience in our systems.

For example, rather than present a user with an HTTP 404 response for a failed request, can we instead provide a degraded experience, directing them somewhere else in the meantime? If traffic volume is exceeding typical demand, can we allocate additional capacity, even temporarily?

7. Reinforcing Feedback Loops

These feedback loops are self-reinforcing, meaning that the more they succeed, the better able they are to continue succeeding. The "snowball effect" comes to mind.

From a technical standpoint, I considered negative feedback loops, such as when one or more web requests contribute to resource starvation. In a finite system, if requests take a long time to complete, increasing the number of requests means the system as a whole takes longer to respond. As more requests come in and are not satisfied in a timely manner, they too contribute to resource starvation. IP ports can be tied up; database connection limits can be exhausted.

An example of a positive reinforcing loop would be one where the marketing team starts a campaign for a product or website. As more people enjoy their experience and talk to each other about it, others join in, talking to still more people until the demand for a product or service is growing exponentially.

6. Information Flows

Meadows describes this leverage point as being about access to information. That is, who does and does not have that access and what that means for the operation of the system. We don't know what we don't know so I it's safe to say that our general awareness of available information has just as much impact.

For example, one can collect many metrics and curate many dashboards to surface the behavior of a system. But it is always possible that there is some thing that is not being watched, that contributes to system failure. All that can be done is to perform a review, learn from the things that surprised us, and update our metrics collection/monitoring.

I recall a presentation several years ago from an engineer at Google that described a scenario where a new feature was released and it passed all the written tests for the code. The feature made a change to images on a page so that they were transparent. The tests confirmed that an image was present but not that it was visible! It took humans to analyze the issue to find the problem. As a result, they wrote a tool to perform a "visual diff" before and after code deploys.

In a similar vein, listening to users in forums is just as important as the data visible in a dashboard. Oftern times, understanding how problems are manifesting for users is useful in diagosing a problem.

5. Rules - Incentives, Punishments, Constraints

Truly, Meadows' frames this lever succintly: "The rules of the system define its scope, its boundaries, its degrees of freedom."

How do we deliver features to our site? Are they large, infrequent chunks? Or small, frequent, easier-to-reason-about changes?

How do we address failures in a system? Do we find someone to blame for making a mistake? Or do we try to understand the context of an event and learn how it is that the system allowed the event to take place?

What contrains us? Physical hardware? Network bandwidth? Service-level agreements? Contractual obligations?

4. Self-organization

Self-organization allows a system to evolve, to change with its environment. This is immediately applicable in the realm of software development and web operations. We can move very quickly, in most cases, responding to changes in demand, creating new products and services faster than anything in the physical world. We can adapt and, importantly, repair, very quickly.

Meadows calls self-organization the "strongest form of system resilience."

3. Goals

Goals are more important than any other leverage point that has thus been defined below this point. A system's purpose for being, the direction it intends to move in governs everything else, including how information flows, what the rules, incentives, and punishments are, and how (even if) the system self-organizes.

What product or service do you want to build? What market segment/demand does it satisfy? Why build this thing rather than do some other type of work?

Does it matter that code/servers can be delivered fast and easy? Or is slow and steady more important?

Does management do everything it can to make it easier for engineers to do their work? Or does management need to be consulted for approval for everything?

In short, you will get out of the system what you put into designing and adhering to its goals.

2. Paradigms

Meadows' describes paradigms as "shared idea[s] in the minds of society[.]" These ideas and beliefs are part of what shapes the goals we set for our systems. I can't help but think about open source software and human factors.

"Paradigms are the sources of systems."

Open Source Software

If you believe in the value of open source software, as I do, it is because you believe in (at least) the following:

Human Factors

The New View of human factors challenges the tradionally held (i.e. Old View) notion of bad/incompetent actors that misbehave within a system, causing failure. Rather than assume we can know everything, given the luxury of being outside an event, and that people simply can't work within the parameters of a well-designed system, the New View of human factors takes the approach that we need to try and understand an operator's context during an event, their biases, prejudices, training, and past experiences, before we can truly learn how a failure can occur within the system as we think we understand it. Instead of blaming practioners for errors in judgement or execution that doesn't follow the rules, the New View strives to learn why and how things happened the way they did.

The New View is very much systems-level thinking: it steps outside event details to understand the system's stocks, it's flows (input/output) and the interaction of its various components.

1. Transcending Paradigms

I can point to nothing technical, managerial, nor organizational here. Here Meadows espouses being open to learning new things, to change, to one's own personal growth as we try to better understand, build, and operate the systems around us. I can think of no better way to end than to quote Meadows directly:

It is to "get" at a gut level the paradigm that there are paradigms,
and to see that that itself is a paradigm, and to regard that whole
realization as devastatingly funny.

Final Thought

Toward the end of the book, Meadows began to touch on the fact that systems are not cold, hard constructs, especially because they do not operate independently of people. To that end, understanding them requires a degree of humility, a desire to learn, and a realization that it's never as simple as pushing buttons or pulling levers. There are many contexts we need to be comfortable with. In short, building, understanding, and changing systems is part science and part art.

[M]astery has less to do with pushing leverage points than it does
with strategically, profoundly, madly, letting go and dancing with
the system.

That is beautiful to me.

Footnotes

[1] Thanks to @mrtazz, @stack72, and @kAnderson610 for the recommendation. Return