Addiction Graph as a Filter to Chaos Engineering (Category SRE)

Read the five-page work of Anatoly KrasnovskiyModel Discovery and Graph Simulation: A Lightweight Gateway to Chaos Engineering" The main idea is simple: before expensive experiments with failures, you can automatically collect dependency graphs from distributed tracks, add the number of replicas and cheaply run Monte Carlo along it.

The practical problem with such models is a manual description of architecture that quickly becomes obsolete. Here, the graph is extracted from Jaeger; in the future, the sources may be service mesh telemetry, Kubernetes/Terraform manifestos, API contracts, and SLO-as-Code. It turns out not another diagram, but an executable accessibility model that can be updated in CI/CD.

Checked the approach DeathStarBench Social Network: two deployment modes, five failure rates, simulation comparison with actual failure implementation (fault injection). The overall correlation between the model and the live experiment was approximately 0,992. Replication and failure rate 0,3 The assessments were almost identical: 0,3054 against 0,3054. But without replicas, the deviation was systematic: -24,4% upon 0,1 before +20,7% upon 0,5.

Separately, it should be noted that the model sees only mandatory synchronous calls and independent fail-stop failures. It does not take into account partial and gray refusals. (gray failures), correlated failures, queues, repeated attempts, load reset and asynchronous flows. And experienced architects know that the real radius of damage (blast radius) It is often determined not by the number of services or clusters, but by common correlated layers—the OS image, CNI, delivery system, or supply chain.

So for me, this is not a replacement for the chaos of engineering, but a nice and relatively cheap stage before it. Graph quickly shows suspicious chains, single points of failure (SPOF) Live experiments remain for places where the topology is incomplete or the behavior of the system is more difficult than the binary “work / fall”.

Work entered ICSE-NIER 2026 Received the Distinguished Paper Award (anyway code published). But for now, this is a test on one example. (proof by instance) and one benchmark. The practical first step for the platform team is to link tracks, dependency graphs, SLOs and the number of replicas into a verifiable artifact – and from it to form a short list of chaos scenarios with the highest risk.

P.S. Perhaps, we will write down with the author of the article its analysis and discussion of the foreign product, which the author did on the basis of this idea. If you like this idea, put

#Software #Engineering #Architecture #DevOps #PlatformEngineering #RnD