norun9/re2ob-pyg

RE2-OB PyG

Preprocessed RE2-OB dataset as PyTorch Geometric Data objects, ready for use with ADA-TGN.

RE2-OB is a microservice fault injection dataset based on Online Boutique. Metrics are collected from Prometheus/cAdvisor/Istio at 1-second granularity and aggregated into 10-second snapshots.

Dataset Structure

Each experiment dict contains:

Key	Type	Description
experiment_id	str	{service}_{fault}, e.g. checkoutservice_cpu
rep	int	Replication index (1, 2, or 3)
root_cause	str	Fault-injected service (same as the service part of experiment_id)
normal_pyg	list[Data]	PyG Data objects before inject_time.txt (normal period)
anomaly_pyg	list[Data]	PyG Data objects from inject_time.txt onward (anomaly period)

* inject_time.txt is a file in each RE2-OB experiment directory containing a single Unix timestamp indicating when the fault was injected.

Each Data object:

Attribute	Shape	Description
x	[10, 12]	Node feature matrix
edge_index	[2, 14]	Inter-service call graph
y	[1]	Graph-level label (0 = normal, 1 = anomaly)
node_y	[10]	Node-level label (1 for root cause node only)

Nodes and Features

10 nodes — Online Boutique services (redis excluded: no Istio sidecar):

adservice, cartservice, checkoutservice, currencyservice, emailservice, frontendservice, paymentservice, productcatalogservice, recommendationservice, shippingservice

12 features per node (based on SRE's Four Golden Signals):

Group	Features
Latency (4)	istio_latency_{50,90,95,99}
Memory (3)	container_memory_{usage_bytes,rss,working_set_bytes}
CPU (2)	container_cpu_{user,system}_seconds_total
Network (3)	istio_request_total, container_network_{receive,transmit}_bytes_total

14 edges — directed inter-service call dependencies (checkoutservice→*, frontendservice→*, recommendationservice→productcatalogservice).

Experiments

60 experiments: 5 services × 4 fault types × 3 replications

• Services: checkoutservice, currencyservice, emailservice, productcatalogservice, recommendationservice
• Fault types: CPU stress, memory stress, network delay, network packet loss

Preprocessing

• Snapshot: 10-second aggregation windows
• Latency / Memory / Network: P99 scaling → asinh(x / 0.5) × 2.0 → clip(−10, +10)
• CPU: diff (rate) → Gaussian smoothing (σ = 2.0) → divide by 0.5

To reproduce or customise preprocessing, run notebooks/feature_extraction.ipynb.

Source Dataset

Raw RE2-OB dataset: Zenodo #14590730

License

Apache 2.0