Devil's Advocate in Operations

Around the world, Israeli intelligence agencies have a reputation for punching well above their weight. For a country of less than 10 million people, they compete with nations orders of magnitude larger and have routinely demonstrated an ability to achieve far more than their larger cousins.

There are many factors which play a role in this, but today I'd like to focus on a small keynote in that story which has, on numerous occasions, been the primary factor in advancing the reliability of the services I support. That factor is the role of the "Devil's Advocate".

History of the Devil's Advocate Unit

Following the Yom Kippur War and the failings which led to Israel being unprepared for the simultaneous attack by Egypt and Syria, an investigative commissionopen in new window undertook the task of identifying improvements which would allow the Israeli Defence Forces to be better prepared for a future conflict.

Their recommendations included the creation of a special unitopen in new window, tied to the Israeli Military Intelligence directorate, but not reporting to them. This unit's job is to (in a professional and critical fashion) examine unlikely scenarios and question the assumptions made by Israeli Military Intelligence and, where appropriate, present counter-proposals.

Important Takeaways

There are a few key points I want to call out here before we continue, because they distinguish this formal role from the often seen "Devil's advocate to disrupt progress" persona which we often run into.

  1. This is a formal role assigned to a group or individual.
  2. There is the expectation of professional, critical, evaluation of the counter proposals generated.
  3. The focus is on examining unlikely scenarios and the assumptions made by the broader group.
  4. The goal is to deliver the same end result as the broader group by assisting them in seeing things they may have missed.

With that in mind, let's talk operations...

Investigating a Problem

When you're working in any sufficiently complex environment, you are going to run into problems for which there is no clear answer and no clear path one can follow to derive the answer. Ideally, these situations happen rarely, but they do happen. If you're lucky, you're dealing with a communication issue and looping more people in will highlight the missing piece of information which unlocks the puzzle, if you're unlucky this will simply leave you with a larger group of people who are confused about how to proceed.

There is a critical point in this process where the group of investigators grows large enough that the available bandwidth outstrips the amount of information which can be analyzed under the current investigative directions and it is this point which I am most interested in today.

Coincidentally, it is this point that I found myself in 2 weeks ago, 2 weeks into an extremely complex investigation. With ~10 people actively engaged in trying to narrow down the cause of the particular issue we faced and several strong candidates with supporting statistical data, we still weren't making forward progress. As a group, we had plans for what questions we needed to answer next, what information we needed to gather to answer them and what we thought we could test to refine our theories, but we fundamentally lacked a coherent story about what was going wrong. Attempts to fabricate one led to the familiar Step 1?... Step 2??... Something we don't understand... Step 5! pattern.

Breaking the Cycle

In these situations, one of the most powerful roles you can introduce into the group is a person tasked with playing Devil's advocate. This person's job is not to dispute the information provided by other members of the group, but rather to examine the assumptions made, look for alternative theories which explain the data and pursue lines of inquiry which diverge from the established narrative.

The person tasked with this role should ideally be someone with solid systems thinking experience and the penchant to be creative. They should be able to rapidly identify supporting and disputing evidence and use this to quickly filter theories which are impossible while highlighting those which may be unlikely, but which have supporting data.

3 weeks into our investigation and it was this approach which delivered a coherent story, something we could start to reason about. We didn't know what Step 1 and Step 2 were, but we could now see a clear progression of Step 3... Step 4... Step 5! and the solution space had been compressed greatly for the areas we were confident about.

Reproducing the Success

Fundamentally, success in this domain is a product of two key realizations:

  1. You have multiple people with tons of experience looking at the problem and attempting to optimize a solution.
  2. Local-minima will naturally draw the search party unless we intentionally cast our gaze further afield.

With that in mind, the role of Devil's Advocate is to act as a scouting party for other nearby search opportunities and to rapidly determine whether directing the main search party's attention to these opportunities is a good or bad investment. If they determine that it is a good investment, they need to be able to efficiently and accurately provide context about the new search domain.

Identifying Search Domains

Identifying areas to search can be done in two ways: randomly casting out in a direction, or more procedurally evaluating useful avenues based on the findings of the main search party. Random can be useful, but it is very domain specific and I don't think I can add much value here. Procedural on the other hand, is much more structured and can enable engineers to more effectively build off one-another's progress.

I usually start off by working with the broader group to understand what their observations are and their current theories. You cannot possibly play Devil's advocate without understanding the mainstream position, so you need to get good at listening and asking questions.

Explain how $x starts to fail?

A good first question is to ask the group to explain not just what they believe the failure is, but what sequence of events leads to that failure. Have them walk you through the full dependency graph and highlight any supporting (or refuting) information they have which justifies their theory.

In my experience, the places with the most value are those where the search party cannot justify a portion of this dependency graph. That may be the result of them fixating on symptoms, mistaking a symptom for a cause or simply because they lack the information or experience to pursue the necessary avenue.

Take notes of all of the areas where you feel the justifications are vague, lack supporting data or are non-existent.

If it wasn't $x, what would it be instead?

When we're dealing with complex systems, there are often numerous possible dependencies which may impact a given node in our graph. In situations where you feel like a portion of the graph is "cloudy", start asking the engineers you're working with questions about what else could be contributing.

Perhaps it isn't dependency $x which is failing, but rather $a. Asking them to actively assume that a given dependency is in fact healthy (in situations where you do not have adequate confidence in either direction) can yield very interesting results. If possible, try removing this dependency intentionally and observing the system's behaviour, or looking to see whether $x and $a are both contributing to the failure.

You'll want to take notes of these options as well, because they're the things the search team is likely not directly looking at and are potentially a treasure trove of useful information. In my experience, most teams tend to land close to the core issue on their first guess and it is subtle shifts in the landscape caused by one or two assumptions which makes the difference between identifying the cause of a failure and fixing something related, but sub-critical.

If it was $x, why would we see $y?

With an idea of your dependency graph and its potential variables, your next step should be to start testing hypotheses. Look for metrics and logs which support, or refute, each assumption. Does the team believe that $x is failing? If so, what would confidently confirm that/confidently refute that? Look for this information and use to to aggressively prune the tree of theories which cannot be substantiated.

Sharing Context

It is crucial to remember that although you are playing Devil's advocate, you are still a fundamental part of the investigative team and you should remain in constant communication with them throughout this process. That means keeping abreast of new developments on their side, as well as sharing any information you find. Any success is going to be the result of a team effort and while playing the role of "scout" might help you deliver a key which unlocks further progress, attempting to develop the full potential of that key on your own is likely to delay the team's success.

Of course, sharing too much information and failing to curate stuff can also distract other engineers from the critical investigative work they are doing, so finding a good balance is the aim of the game. Where that balance lies will depend a lot on your team and the situation you find yourself in, but as a general rule, I will share information under these conditions:

  1. This is a new piece of information which nobody has seen before and it clearly relates to the issue at hand.
  2. I can now clearly tie this piece of information to a candidate narrative about the failure and it forms a crucial piece of that narrative.
A picture of Benjamin Pannell

Benjamin Pannell

Site Reliability Engineer, Microsoft

Dublin, Ireland