# ChronoGraph Canonical source: Official code and data: Introducing source: [ChronoGraph](../../wiki/sources/chronograph-2025.md) ## Dataset Type ChronoGraph is a graph-structured multivariate time-series dataset for production microservice forecasting and anomaly detection. Nodes are services, edges are service-to-service dependencies, node features are service health metrics, and edge features are interaction metrics. ## Temporal And Graph Structure The paper reports six months of production telemetry collected at a 30-minute interval. After alignment and filtering, each of 708 services has 8005 time steps and five service-level metrics. The official README reports a directed temporal graph with 708 nodes, 1529 edges, five temporal node features, eight temporal edge features, and node-level anomaly or disruption labels. ## Data Structure - `edges.csv` stores the static directed topology with `source` and `target` service identifiers. - `node_features.json` stores nested service-level metric trajectories with `values` and `steps`. - `edge_features_part{i}.json` stores nested edge-level trajectories keyed by `source_id->target_id`, again with `values` and `steps`. - Incident labels come from expert-labeled disruption windows. ## Inputs And Outputs Inputs are past node metric trajectories, past edge metric trajectories, and the service dependency graph. Forecasting outputs are future service metric values. Anomaly-detection outputs are disruption scores or incident-window detections. ## Reported Baselines The paper evaluates Prophet, Chronos-Bolt Base, TabPFN-TS, Autoencoder, Isolation Forest, One-Class SVM, and an ensemble combining Prophet, Isolation Forest, and Autoencoder. The reported baseline suite is mostly topology-agnostic, so ChronoGraph is better understood as a benchmark that motivates graph-aware modeling than as a solved graph-model leaderboard. ## Actions Or Interventions ChronoGraph has no logged operator-action channel. Incident windows are labels or exogenous disruption events, not controllable interventions. It is useful for passive graph/time-series representation learning and forecasting, but it is not an action-conditioned operations world-model dataset by itself. ## Access And License Notes The official GitHub repository is public and contains an Apache-2.0 LICENSE file. This knowledge base records metadata only and does not mirror the dataset payload.