Learning from failure is extremely useful, especially in a world driven by demands for high reliability. And now, in the face of COVID-19, as organizations adjust to a rapidly changing environment and, where they're able, a newly all-remote workforce, addressing failure takes on a new set of challenges. How we choose to approach learning, and which failures we want to focus on, may not be clear. Two common avenues to learning are incident analysis and chaos engineering.
Recently, I was an editor on the O’Reilly book on Chaos Engineering. Around the same time, in informal conversations, I had heard folks discuss Incident Analysis and Chaos Engineering in opposition to each other; that is, believing one to be superior to the other in elucidating the important details that organizations can apply to build more robust solutions or improve their incident response. The reality, however, is that they are complementary. In general, Incident Analysis is broader in scope, and can be useful in directing and designing chaos experiments. This post will discuss Incident Analysis and Chaos Engineering and attempt to demonstrate how they can be considered in tandem to help your organization learn and grow.
Having worked across a diverse range of organizations (healthcare, e-commerce, fintech) and participated in several major incidents, I came to recognize that well-done incident analysis takes time. How much time it requires is always a debate, but suffice it to say that time-boxing ourselves to a few hours or constraining investigations to filling out a template does not result in high-quality analyses. To identify markers of quality in our analyses, it helps to define why we perform investigations in the first place. At a high level, I’ve observed three outcomes associated with incident analyses that govern their perceived value: pushing paper, technical teaching, and surfacing surprises.
Maybe it’s company policy to write an after-action review after an incident has occurred. Or, perhaps your organization must produce reports as part of regulatory requirements or fiduciary responsibility. Given this framing, incident analyses will be written to be filed, not read. At this level, organizations may leave a lot of useful information un-gathered, favoring compliance over critical inspection of events that are uniquely situated to teach them .
In these cases, the value of analysis is largely superficial: it’s part of a process to generate a paper trail. It’s a bureaucratic defensive procedure.
Imagine you operate an e-commerce site whose API fell over as requests stacked up during a database transaction ID wraparound event. Your analysis of an incident like this is an opportunity to provide useful details that update relevant documentation and procedures that can address technical gaps in people’s understanding of the architecture that powers your product. This is a good thing and it’s how we build our systems to be robust in the face of past, known conditions.
At this level of Incident Analysis, there is an acknowledgement that we can’t all possibly know everything. The value of analyses seems tangible and immediate by seeing failure as the chance to fill in some blanks in our processes, re-train ourselves, and possibly refine our practices. Still, there is even more to glean from our incidents.
At the “Technical Teaching” level of analysis, the notion of a “system” tends to be conceived as some set of well-known, discrete components that humans operate from a distance, wholly separate from it. In truth, our understanding of our systems is incomplete. There are conditions we cannot account for, nor predict, that will push our systems past their boundaries, surprising us.
The technical-teaching analysis reveals that there are gaps in even the most experienced engineer’s mental model about how the system functions. In fact, a common theme across incidents is that something, somewhere, was surprising. Surprise is borne from the events of an incident contrasting with our mental models of how the system operates. Surfacing surprise as a focus of analysis makes explicit the difference in mental models across a team of responders. If we delve into those surprises, we may be able to identify where those mental models and reality diverge. From Gary Klein's book, "Sources of Power":
“The surprise event challenges the model and triggers learning and model revision…”
There are powerful lessons to be learned here. While individual mental models may be partial and incomplete (Woods, 2017), collectively across the team they overlap and this provides more depth to understanding the system. Analysis that focuses on surprises helps identify the degrees by which those models overlap or are incomplete within teams and across an organization.
Beyond the specific details of an incident, there are likely other interesting surprises including tension between business goals and procedure (especially safety procedures) and differences between work-as-imagined and work-as-done. This tension influences performance. Here, organizations stand to gain from deep insights about the nature of their work that they can use to make positive improvements, such as implementing new response tactics or enhancing procedures to take advantage of efficiencies developed as workarounds to under-specified tasks.
Chaos Engineering has a lot of attention in software at the moment and rightly so. As software engineering matures, and code is run in more environments, we must find improved ways of building confidence in its operation. Our software is designed to express specific characteristics and behave in expected ways. It reflects our current ideas of the problems it is intended to solve. Because we live and work in dynamic environments, eventually things will change enough that our initial designs may no longer serve us. Well-designed Chaos Engineering experiments help us evaluate our systems under realistic conditions and loads to help identify where entropy is slipping in and where our initial expectations are holding up.
As far as learning from failure is concerned, Chaos Engineering experiments provide a similar value as do the sorts of incident analyses that provide technical teachings because the results are tangible and immediate. Chaos experiments are narrowly scoped and the deliverable is clear.
In a few years, I suspect that Chaos Engineering will become a regular practice in software engineering, much like writing unit tests. In other words, it will be seen less as a hot new idea and more of a normal part of rigorous engineering.
Incident Analysis or Chaos Engineering
In terms of deciding how to make better use of time and people as part of a learning-from-failure mindset, it would certainly seem like focusing on chaos engineering is more efficient than deeply analyzing incidents. And this is understandable; incidents are big and scary, with unknown contours. Chaos experiments include the comfort of constrained variables and the certainty of science!
Unfortunately, framing this conversation as an either/or proposition sets up a false dichotomy – one that attempts to establish greater value of one over the other. Even attempting to frame the conversation as which is more efficient in achieving learning from failure misses the mark. A more nuanced perspective is that incident analysis and chaos engineering are complementary and when paired, may produce powerful results.
Recouping our Investments
Experiments are not executed for their own sake. We run them because we are looking for tangible and immediate results. How do we decide which to run? Further, how do we prioritize one over another? From where do we get the motivation to create them? The answer, simply, is past incidents.
The folks at Adaptive Capacity Labs have this to say about incidents:
Incidents are unplanned investments, and … [y]our challenge is to maximize the ROI…
Incidents, therefore, inherently have value. Failures we have experienced direct our attention to areas that warrant further inspection including, possibly, via chaos experiments. A high-quality Incident Analysis will yield useful insights, including pointing to areas where an organization can focus to learn more.
If an incident is an investment, the direction and ideas for chaos experiments derived from its analysis are one way we can recoup our costs.
Imagine looking across a history of incidents – what patterns might we see that could further inform which experiments to run? The combined insights of multiple analyses, then, would make it possible for an organization to design more targeted exercises, thereby extracting greater value.
In other words, through incident analysis, we can develop well-targeted chaos experiments that help us maximize the returns on our incidents-as-investments. This is how we should assess incident analysis. It is never about a report or correctly identifying a failing component – it is always about surfacing useful context that directs our attention in meaningful ways.
A Vision for the Future
In 1995, Boeing set out to test the wing deflection of their 777 airplane. The test subjected the wings to stresses well beyond what it was expected to encounter in operation. The test was designed to find the wings’ literal breaking point. And they did, spectacularly, as seen in this video:
I have a vision for the future of software engineering:
Every organization will know the value of incident analysis and that knowledge will be expressed in day-to-day work like it was for Boeing during this test 25 years ago.
Certainly, Boeing did not lightly decide to build a multi-million dollar vehicle simply to break it. It took much more to motivate a company as massive as Boeing to expend the time, energy, people, and, notably, serious money to design and build the experiment, the plane, and the test rig required to see this through. They deliberately chose to build a plane and snap its wings. That choice was driven by what they and the rest of the aviation industry has been learning from analyzing incidents.
Our job now, as software engineers, managers, or investigators, is to develop the skills necessary to analyze our incidents and draw out the deep insights they expose. Sometimes those insights will lead to wing-snapping experiments; sometimes they’ll be less exciting. Either way, they will all be extremely valuable in pushing the state of the art forward.
 See the “Challenges to Making Progress” section in John Allspaw’s presentation People Are The Adaptable Element of Complex Systems From Lean Agile Scotland 2019.