Research Insights Made Simple 25: We model reliability by dependency graph (Category SRE)

Is it possible to know where a distributed system will break before the expensive crash experiment? In this release tonight 17:00 Together with Anatoly Krasnovskii we will analyze his work.Model Discovery and Graph Simulation: A Lightweight Gateway to Chaos Engineering", marked Distinguished Paper Award on ICSE-NIER 2026 (By the way, I handler before).

Anatoly is an engineer who went into science to save IT from shamanism. Za. 10For years in development, he was tired of the fact that complex systems are built on intuition and blind copying of “best practices”. He is now writing his PhD in Mathematics at Innopolis to learn how to prove their resilience mathematically rather than relying on empirical evidence, and runs his own channel on how complex systems actually work: @mb3rlab

The idea of Anatoly’s approach from this article is simple: automatically extract the mandatory call graph from distributed routes, add the number of replicas and use Monte Carlo to assess the system’s availability in case of failures. It turns out not to replace chaos engineering, but a cheap filter in front of it: the model helps to find suspicious chains, single points of failure and scenarios that are worth checking on a living system first.

We'll discuss with the author:

Why topology and number of replicas may suffice for a first approximation; How to automatically detect the Jaeger model and keep it up to date with the system What does a high correlation with live fault injection mean and why a single benchmark does not yet prove the universality of the method? Where model capabilities end: gray failures, correlated failures, queues, retries and asynchronous flows; How to embed such analysis in CI/CD, link it to SLO and turn it into a prioritization of chaos experiments Where is the line between useful simplification and dangerous false certainty?

For me, the big question of the release is a practical one: can we turn observability from a way of investigating an incident that has already happened to an executable reliability model -- and use it to break the system less often, but more accurately?

#Software #Engineering #Architecture #DevOps #Reliability #Research