lincyaw/openrca2-v1-500

OpenRCA2 v1 (500 cases, 2026-05-02 snapshot)

Curated 500-case RCA evaluation set built 2026-05-02 from a unified pool of older FSE/openrca2 train-ticket cases plus the most recent aegisctl detector_success runs across train-ticket, hotel-reservation, and otel-demo. Each case carries the full telemetry parquets, the injection metadata, and a regenerated service-level causal_graph.json.

Layout

<case_name>/
├── injection.json            # ground-truth fault list (engine_config[*].app/chaos_type/…)
├── causal_graph.json         # service-level GT graph (regenerated 2026-05-02)
├── env.json                  # case environment / namespace metadata
├── conclusion.parquet        # alarm conclusions used for selection
├── abnormal_metrics.parquet  # OTLP metrics during the fault window
├── abnormal_metrics_histogram.parquet
├── abnormal_metrics_sum.parquet
├── abnormal_traces.parquet
├── abnormal_logs.parquet
├── normal_metrics.parquet    # baseline (pre-fault) windows
├── normal_metrics_histogram.parquet
├── normal_metrics_sum.parquet
├── normal_traces.parquet
└── normal_logs.parquet

MANIFEST.json at the root lists all 500 cases with system / chaos_family / hybrid flag / service-graph stats.

Composition

	ts	hs	otel-demo	total
old (FSE/openrca2)	188	0	0	188
new (aegisctl detector_success)	126	132	54	312
total	314	132	54	500

Service-level longest-path distribution (DAG depth + 1): mean 3.46, p25=2, p50=3, p75=4, max=9.

Two injection.json formats

• New (aegisctl detector_success): engine_config is a list of dicts with app, chaos_type, target_service, direction, class, method. Use this directly.
• Old (FSE/openrca2): engine_config is an opaque JSON-encoded string and fault_type is numeric. The authoritative ground truth is injection.json::ground_truth. Loaders should fall back to that field when engine_config is not a list-of-dicts.

The rcabench-platform[llm-eval] v2 evaluator (rcabench_platform.v3.sdk.evaluation.v2.extract_gt_faults) handles both formats transparently.

Picking pipeline (reproducibility)

1. aegisctl inject list --state detector_success --all -o ndjson → 48-h gap query.
2. aegisctl inject download-batch → fill the gap.
3. Symlink-wrap source dirs into a canonical <pool>/<case>/converted/ layout.
4. cli/reason.py reason batch --max-hops 15 --force → regenerate causal_graph.json.
5. Score + filter (drop no_graph, cyclic, frontend_inj, empty_graph); 705 candidates survive.
6. Greedy pick under diversity caps (system, chaos_family, root_service, hybrid:leaf ratio) → 500.

The causal_graph.json shipped here is the regenerated one; the reasoning's result.json and label.txt are intentionally omitted from this snapshot.

Loading example

from huggingface_hub import snapshot_download

local = snapshot_download(
    repo_id="lincyaw/openrca2-v1-500",
    repo_type="dataset",
    local_dir="./data/openrca2_v1_500",
)

import json
manifest = json.load(open(f"{local}/MANIFEST.json"))
for case in manifest["cases"][:3]:
    case_dir = f"{local}/{case['name']}"
    injection = json.load(open(f"{case_dir}/injection.json"))
    # use rcabench_platform.v3.sdk.evaluation.v2.evaluate_v2(...) for scoring

Citation

If you use this dataset, please cite the upstream rcabench / OpenRCA work.