The benefits of automation are well understood: more agile service provisioning, faster time to insight when there are issues, and a reduction in human error as manual interaction is reduced. Much of the premise behind long-term SDN architectural advantages is steeped in the hope that SDN will help enable and ultimately promote automation. But while centralizing control has significant operational advantages, by itself, it doesn’t actually address the most important requirement for automation.
If automation is going to be more than just reducing keystrokes, there will have to be a rise of workflow state.
Successfully managing a network is an exercise in constant iteration through network state. Whenever something needs to be done, the architect or operator examines her current frame of reference to figure out the starting point. That frame of reference usually starts with some implicit understanding of how the network is designed. From there, she takes some action. Maybe she pings an endpoint, checks the state of a BGP neighbor, or examines some interface statistics. Whatever the first step, the point is that she knows when she starts that there is work after the first step.
The information gleaned from the first step yields additional understanding. Her frame of reference changes as she now knows more than before. With her new position in referential space, she takes the next step. And the next, and the next after that. Each step yields a different piece of information, and the process of iterating through a constantly changing referential space ultimately yields some outcome or resolution.
Byproducts of iterative workflows
There are two major byproducts of this iterative approach to workflow. The first is that the starting point is rarely based on an absolute understanding of fact. Rather it is an interpretation that the individual operator or architect creates based on a number of somewhat soft conditions – knowledge, experience, intuition, whatever. This means that for each task, the workflow is somewhat unique, depending on the operator and the environment.
The impact here is important. If workflows are unique based on the operator and the conditions (i.e., the referential space or frame of reference), then the outcomes driven by those workflows are difficult to repeat. Part of why networking is so hard is that so much of it borders on arcane dark art. Science demands repeatability, but the very nature of workflow management in networking makes that challenging.
The second byproduct of networking’s iterative nature is that workflows frequently depend on a set of chained tasks, each of which has a dependency on the preceding task. To make things worse, that dependency is actually rarely known at the start of a a workflow. It’s not that tasks cannot be predictably chained – first, you look at the physical layer, and then you move up stack perhaps. But each subsequent task is executed based on not just the previous task but also the output of the previous task. This creates a complex set of if/then statements in most workflows.
Part of the challenge in automation is providing the logic to navigate the conditional nature of networking workflows.
“Network engineers need to think like programmers”
With the rise of movements like DevOps, “network engineers thinking like programmers” has become a popular phrase. This is a very important change in how we handle network architecture and operations. But there are subtleties here that get lost in the cliche.
First, when people toss the phrase around, they often mean that network engineers need to pick up a scripting language (Python, Ruby, even Perl). Thinking like a software developer has very little to do with programming languages. Languages are a way of expressing intent, but it’s entirely possible to know Python and think nothing like a developer.
Second, when people refer to programming in the context of DevOps, they generally mean that network operators need to think about configuration less as a collection of commands and more like code. Once you make that shift, then you can think about things like source code management, automated testing, and rapid deployment.
But networking needs to do more than just treat configuration as code. DevOps has more to do with deploying and validating changes. It doesn’t fundamentally change how workflows are executed, and it barely touches more operational tasks like troubleshooting network conditions.
Before anyone picks a religious battle over DevOps here, my point is not that DevOps is bad. It’s just that DevOps by itself is not sufficient. And there are things that ought to be done that are separate from DevOps.
Tiny feedback loops
So if thinking like a programmer isn’t about learning a programming language and it’s more than treating configuration as code, what is it?
Software development is really about creating something out of lots of tiny feedback loops. When you write functions, you don’t just execute some task. You generally execute that task and then return a value. The value provides some immediate feedback about the outcome. In some cases, the function returns the value of a computation; in other cases, it simply returns an indication that the function succeeded or failed.
These values are obviously then used by other functions, which allows us to string together small building blocks into complex chains. The important part? These chains can then be repeatably executed in a deterministic way.
Networking workflows shouldn’t be that different. Each individual activity yields some value (sometimes a specific value as when looking at some counter, other times a success or failure as with a ping). The problem is that while networking commands frequently return information, it is up to the operator themselves to parse this information, analyze what it means, and then take the next action.
What we need if we really want to make automation happen in ways that extend beyond just scripting keystrokes is a means of creating deterministic networking workflows. For this to happen, we need people who construct workflows to think more like developers. Each activity within a workflow needs to be a tiny feedback loop with explicit workflow state that is programmatically passed between workflow elements.
We actually instinctively do this at times. XML, NETCONF, and the like have been used to encapsulate networking inputs and outputs for awhile with the intent of making things parseable and thus more automatable.
But we stopped short. We made the outputs more automation-friendly without ever really creating workflows. So while we can programmatically act on values, it only works if someone has automated a particular workflow. As an industry, we haven’t gotten to actually addressing the workflow problem.
Maybe it’s the highly conditional nature of networking combined with the uniqueness of individual networks. Or maybe it’s that outside of a few automation savants, our industry doesn’t generally think about workflows the way a software developer would.
The bottom line
Networking workflows rely way too heavily on an iterative pass through referential space. The reason change is so scary and troubleshooting so hard is that very little in networking is actually deterministic. But if we really want to improve the overall user experience en route to making workflows both repeatable and reliable, we do need to start thinking a bit more like developers. It all starts with a more explicit understanding of the workflows we rely on, and the expression of feedback via some form of workflow state.
And for everyone betting on abstractions, just know that abstracting a poorly-defined workflow results in an equally poor abstraction. We need to be starting elsewhere.
[Today’s fun fact: Only male fireflies can fly. Take that, females!]