lincyaw/openrca2-lite-v1

OpenRCA-2.0-Lite v1

A 635-case curated subset of OpenRCA-2.0 for root-cause analysis (RCA) on microservice systems. Stratified across (system × fault_type × root-cause service × propagation-path skeleton) with end-to-end causal-chain verification.

At a glance

• Cases: 635
• Systems: hotel-reserv, online-boutique, train-ticket, sock-shop, social-network
• Fault types: CPUStress, NetworkPartition, NetworkLatency, NetworkLoss, MemoryStress, IOStress, PodFailure, PodKill, ContainerKill, MysqlCorrupt, RedisCorrupt, …
• Per-case telemetry: OpenTelemetry-derived parquet for traces, logs, metrics (counter/sum/histogram), split into pre-injection (normal_*) and post-injection (abnormal_*) windows
• Per-case ground truth: injection.json (root-cause services + fault parameters), causal_graph.json (alarm propagation skeleton), result.json (verified causal chain), conclusion.parquet, env.json, label.txt

Layout

openrca2_lite_v1/
├── MANIFEST.json                                 # 635 case index
├── <system>__<fault_type>__<case_id>/
│   ├── injection.json                            # ground truth (RC service, fault params)
│   ├── causal_graph.json                         # alarm-node graph (used by RCA agents)
│   ├── result.json                               # verified causal chain
│   ├── env.json                                  # cluster + injection metadata
│   ├── label.txt                                 # human-readable label
│   ├── conclusion.parquet                        # SLO-violation summary
│   ├── normal_traces.parquet                     # pre-injection traces
│   ├── normal_logs.parquet                       # pre-injection logs
│   ├── normal_metrics.parquet                    # pre-injection gauge metrics
│   ├── normal_metrics_sum.parquet                # pre-injection sum metrics
│   ├── normal_metrics_histogram.parquet          # pre-injection histogram metrics
│   ├── abnormal_traces.parquet                   # post-injection traces
│   ├── abnormal_logs.parquet                     # post-injection logs
│   ├── abnormal_metrics.parquet                  # post-injection gauge metrics
│   ├── abnormal_metrics_sum.parquet              # post-injection sum metrics
│   └── abnormal_metrics_histogram.parquet        # post-injection histogram metrics
└── ...

MANIFEST.json schema:

{
  "n_cases": 635,
  "curation": { "strategy": "...", "cap_per_type": 30, ... },
  "cases": [
    {
      "name": "hotel-reserv__CPUStress__batch-01KQJ5SQVREZ80KHA4XAZHGPAD",
      "origin": "new",                 // "old" = carried from openrca2-lite, "new" = added in v1 curation
      "system": "hotel-reserv",
      "fault_type": "CPUStress",
      "rc_service": "hotel-reserv-search",
      "rc_services": ["hotel-reserv-search"],
      "alarm_services": ["frontend", "search"],
      "skeleton": ["hotel-reserv-search", "search", "frontend", "search"],
      "n_skeletons": 3
    },
    ...
  ]
}

Origin tags

• old (403 cases): carried over from the original openrca2-lite set
• new (232 cases): added in v1 curation, ground truth stored in AegisLab batch schema (list-of-dicts)

Download

# Hugging Face CLI
hf download lincyaw/openrca2-lite-v1 --repo-type dataset --local-dir ./openrca2_lite_v1

# Or via huggingface_hub (Python)
from huggingface_hub import snapshot_download
snapshot_download("lincyaw/openrca2-lite-v1", repo_type="dataset", local_dir="./openrca2_lite_v1")

The dataset is ~3 MB total.

Use with ThinkDepthAI (LangGraph RCA agent)

git clone https://github.com/<thinkdepthai-repo>
cd ThinkDepthAI
cp .env.example .env
# Edit .env: set UTU_LLM_API_KEY and LITE_V1_ROOT to the snapshot path above

uv sync
uv run python scripts/seed_lite_v1_db.py     # populate eval.db from MANIFEST
uv run rca llm-eval run config/eval/openrca2_lite.yaml -a thinkdepthai

Evaluation reports per-case correctness, causal-graph F1, and path reachability (judge-side).

Citation

@misc{openrca2lite_v1,
  title  = {OpenRCA-2.0-Lite v1: A Curated Microservice RCA Benchmark},
  author = {OperationsPAI},
  year   = {2026},
  url    = {https://huggingface.co/datasets/lincyaw/openrca2-lite-v1}
}

License

CC BY 4.0