Network variation is hurting us

Network devices like switches, routers, firewalls and load-balancers ship with many powerful features. These features can be configured by each engineer to fit the unique needs of every network. This flexibility is extremely useful and, in many ways, it's what makes networking cool. But there comes a point at which this flexibility starts to backfire and become a source of pain for network engineers.

Variation creeps up on you.  It can start with harmless requests for some non-standard connectivity, but I've seen those requests grow to the point where servers were plugging straight into the network core routers.  In time, these one-off solutions start to accumulate and you can lose sight of what the network ‘should’ look like.  Every part of the network becomes its own special snowflake.

I’m not judging here. I've managed quite a few networks and all of them end up with high-degrees of variation and technical debt. In fact, it takes considerable effort to fight the storm of snowflakes. But if you want a stable and useful network you need to drive out variation. Of course you still need to meet the demands of the business, but only up to a point. If you're too flexible you will end up hurting your business by creating a brittle network which cannot handle changes.

Your network becomes easier and faster to deploy, monitor, map, audit, understand and fix if you limit your network to a subset of standard components. Of course there are great monitoring tools to help you manage messy networks, but you’ll get greater value from your tools when you point them towards a simple structured network.

What’s so bad about variety?

Before we can start simplifying our networks we have to see the value in driving out that variability. Here are some thoughts on how highly variable (or heterogeneous) networks can make our lives harder as network engineers:

  • Change control - Making safe network change is extremely difficult without standard topologies or configurations. Making a change safely requires a deep understanding of the current traffic flows - and this will take a lot of time. Documentation makes this easier, but a simple standardized topology is best. The most frustrating thing is that when you do eventually cause an outage, the lessons learned from your failed change cannot be applied to other dissimilar parts of your network.
  • Discovery time can be high. How do you learn the topology of your network in advance of problems occurring? A topology mapping tool can be really helpful to reduce the pain here, but most people have just an outdated visio diagram to rely on.
  • Operations can be a nightmare in snowflake networks.  Every problem will be a new one, but probably one that could have been avoided - it's likely that you'll go slowly mad. Often you'll start troubleshooting a problem and then realize, ‘oh yeah, I caused this outage with the shortcut I took last week. Oops’.  By the way, it’s a really good sign when you start to see the same problems repeatedly. Operations should be boring, It means you can re-orient your Ops time towards 80/20 analysis of issues, rather that spending your days firefighting.
  • Stagnation -  You won't be able to improve your network until you simplify and standardize your network. Runbooks are fantastic tools for your Ops and Deployment teams, but the runbook will be useless if the steps are different for every switch in your network. Think about documenting a simple task...if network Y do step1, except if feature Z enabled then do something else, except if it’s raining or if it's a leap year.  You get the message.
  • No-Automation - If your process it too complicated to capture in a runbook you shouldn't automate it. Simplify your network, then your process, then automate.



Network variation can be a real source of pain for us engineers. In this post we looked at the pain it causes and why we need to simplify and standardize our networks. In Part 2 we'll look at the root causes for these complicated, heterogenous networks and how we can begin tackling the problem.