RCABench OpenRCA2-Lite Dataset

A curated subset of 500 cases from the RCABench dataset, tagged as openrca2-lite.

Description

This dataset contains root cause analysis (RCA) benchmark cases for cloud-native microservice systems. Each case represents an incident scenario with telemetry data (logs, metrics, traces) and a ground-truth root cause.

Data Fields

Field	Type	Description
id	int	Unique case identifier
source	string	Datapack name (e.g., ts0-mysql-corrupt-kwx8n5)
question	string	Incident description with SLO-violating API endpoints
answer	string	Ground-truth root cause services (comma-separated)
tags	list[string]	Dataset tags (openrca2, openrca2-lite, fse, fse26)
datapack_name	string	Original datapack identifier
ground_truth	list[string]	Root cause services as a list
fault_type	string	Type of injected fault (e.g., NetworkCorrupt)
fault_category	string	Fault category (e.g., NetworkChaos)
difficulty_spl	int	Shortest path length from root cause to alert
difficulty_n_svc	int	Number of services in the system
difficulty_n_edge	int	Number of edges in the service graph
api_reports	list[string]	Full API endpoint SLO violation reports

Telemetry Data

Each case has a corresponding directory under data/<source>/ containing converted telemetry files:

• normal_logs.parquet / abnormal_logs.parquet - Log data
• normal_metrics.parquet / abnormal_metrics.parquet - Metric data
• normal_metrics_histogram.parquet / abnormal_metrics_histogram.parquet - Histogram metrics
• normal_metrics_sum.parquet / abnormal_metrics_sum.parquet - Sum metrics
• normal_traces.parquet / abnormal_traces.parquet - Distributed trace data
• env.json - Environment configuration
• injection.json - Fault injection metadata
• causal_graph.json - Ground-truth causal graph
• conclusion.parquet - Conclusion data

Usage

from datasets import load_dataset
import json

dataset = load_dataset("lincyaw/rca", split="train")

for case in dataset:
    print(f"Case: {case['source']}")
    print(f"Question: {case['question']}")
    print(f"Answer: {case['answer']}")
    print(f"Fault Type: {case['fault_type']}")
    
    # Access telemetry data
    import polars as pl
    traces = pl.read_parquet(f"data/{case['source']}/abnormal_traces.parquet")

Statistics

• Total cases: 500
• Fault categories: NetworkChaos, ResourceStress, Exception, ContainerChaos, etc.
• Systems: TrainTicket microservice architecture
• Telemetry data size: ~4.5 GB (converted parquet files)

License

Same as RCABench original dataset.

Citation

@inproceedings{rcabench2025,
  title={RCABench: A Benchmark for Root Cause Analysis in Cloud-Native Systems},
  year={2025}
}