Model Discovery and Graph Simulation: A Lightweight Gateway to Chaos Engineering

TL;DR

This paper introduces a lightweight, automated approach to model discovery from existing observability data, enabling fast and low-risk resilience estimation and chaos engineering for complex systems.

Contribution

It presents a novel automated model discovery method that synthesizes service dependency graphs from existing artifacts, facilitating practical resilience testing without handcrafted models.

Findings

Discovered models closely match live fault-injection results.

Median error near zero at mid-range failure rates with replication.

No-replication models show biases, indicating missing mechanisms.

Abstract

Chaos engineering reveals resilience risks but is expensive and operationally risky to run broadly and often. Model-based analyses can estimate dependability, yet in practice they are tricky to build and keep current because models are typically handcrafted. We claim that a simple connectivity-only topological model - just the service-dependency graph plus replica counts - can provide fast, low-risk availability estimates under fail-stop faults. To make this claim practical without hand-built models, we introduce model discovery: an automated step that can run in CI/CD or as an observability-platform capability, synthesizing an explicit, analyzable model from artifacts teams already have (e.g., distributed traces, service-mesh telemetry, configs/manifests) - providing an accessible gateway for teams to begin resilience testing. As a proof by instance on the DeathStarBench Social Network, we extract the dependency graph from Jaeger and estimate availability across two deployment modes and five failure rates. The discovered model closely tracks live fault-injection results; with replication, median error at mid-range failure rates is near zero, while no-replication shows signed biases consistent with excluded mechanisms. These results create two opportunities: first, to triage and reduce the scope of expensive chaos experiments in advance, and second, to generate real-time signals on the system's resilience posture as its topology evolves, preserving live validation for the most critical or ambiguous scenarios.