# ops-lite Canonical source: Introducing source: [ops-lite](../../wiki/sources/ops-lite-2026.md) ## Dataset Type ops-lite is a curated 500-case root-cause-analysis evaluation set for microservice systems. Its distinguishing feature is manifest-driven causal-graph ground truth per case. ## System Structure The corpus spans three open-source testbeds: - Train-Ticket: 320 cases, 44 Java/Spring Cloud application services. - Hotel Reservation / DeathStarBench: 142 cases, 9 Go/gRPC application services. - OpenTelemetry Demo: 38 cases, 15 polyglot application services. ## Data Structure Each case includes: - `injection.json`: chaos-injection ground truth. - `causal_graph.json`: causal service graph derived from the fault contract. - `env.json`, `result.json`, and `label.txt`: runtime snapshot and labels. - 12 parquet metric tables split into abnormal and normal windows. The top-level `manifest.jsonl` has one JSON record per case and includes system ID, longest path, service/edge counts, root services, chaos family, primary kind, subtypes, and hybrid flags. ## Reported Scale The Hugging Face card reports 500 released cases selected from a 1464-case raw pool, mean longest path 3.18, mean edge count 4.06, mean service count 3.97, eight chaos families, and three systems. ## Inputs And Outputs Inputs are normal and abnormal metric windows, runtime environment snapshots, and service graph context. Outputs are root-cause services and causal graph/path predictions, which can be scored against `causal_graph.json` and manifest labels. ## Actions Or Interventions Chaos injections are controlled benchmark events. They are useful as exogenous fault conditions, but the released evaluation set does not provide operator remediation decisions. ## Access And License Notes The Hugging Face dataset card lists Apache-2.0. The card also notes that Train-Ticket, Hotel Reservation / DeathStarBench, and OpenTelemetry Demo retain their own upstream licenses. This knowledge base records metadata only and does not mirror case payloads.