Network design discussions often involve anecdotal evidence, and the arguments for preferring something follow up with “We should do X because at Y place, we did this.”. This is alright in itself as we want to bring the experience to avoid repeating past mistakes in the future. Still, more often than not, it feels like we have memorized the answers, and without reading and understanding the question properly, we want to write down the answer. Ideally, we want to understand the problem and solution space and put that into the current context, discussing various tradeoffs and picking the best solution in the given context. Our best solution for the same problem may change as the context changes. Also, this problem is everywhere. For example: Take a look at this twitter thread

Maybe one way to approach on how to think is to adopt stochastic thinking and add qualifications while making a case if we don’t have all the facts. The best engineers I have seen do apply similar thought processes. As world-class poker player Annie Duke points out in Thinking in Bets, even if you start at 90%, your ego will have a much easier time with a reversal than if you have committed to absolute, eternal certainty.

Now we aren’t going to fix human behavior anytime soon. However, purely from an analytical perspective, for us to be able to quantify things, many times, we need to rely on modeling and simulation. But often, those frameworks don’t exist. One such area was quantifying routing protocol implementation behavior for a given topology until Sibyl came out to help us. At a high level, It proposes various metrics to measure routing protocol implementation, allows you to spin up the topology of a given size with a given protocol implementation, and then measures those metrics for a given failure.

A Markov chain or Markov process is a stochastic model describing a sequence of possible events in which the probability of each event depends only on the state attained in the previous event. It is named after the Russian mathematician Andrey Markov.

In case of Linear Models, we assume a linear relationship between the mean of the response variable and a set of explanatory variables with inference assuming that response variable has a Normal conditional distribution with constant variance. The Generalized Linear Model permits the distribution for the Response Variable other than the normal and permits modeling of non-linear functions of the mean. Linear models are special case of GLM.

Introduction