{ "@context": { "@language": "en", "@vocab": "https://schema.org/", "citeAs": "cr:citeAs", "column": "cr:column", "conformsTo": "dct:conformsTo", "cr": "http://mlcommons.org/croissant/", "rai": "http://mlcommons.org/croissant/RAI/", "data": { "@id": "cr:data", "@type": "@json" }, "dataType": { "@id": "cr:dataType", "@type": "@vocab" }, "dct": "http://purl.org/dc/terms/", "examples": { "@id": "cr:examples", "@type": "@json" }, "extract": "cr:extract", "field": "cr:field", "fileProperty": "cr:fileProperty", "fileObject": "cr:fileObject", "fileSet": "cr:fileSet", "format": "cr:format", "includes": "cr:includes", "isLiveDataset": "cr:isLiveDataset", "jsonPath": "cr:jsonPath", "key": "cr:key", "md5": "cr:md5", "parentField": "cr:parentField", "path": "cr:path", "recordSet": "cr:recordSet", "references": "cr:references", "regex": "cr:regex", "repeated": "cr:repeated", "replace": "cr:replace", "sc": "https://schema.org/", "separator": "cr:separator", "source": "cr:source", "subField": "cr:subField", "transform": "cr:transform", "samplingRate": "cr:samplingRate", "equivalentProperty": "cr:equivalentProperty" }, "@type": "sc:Dataset", "conformsTo": "http://mlcommons.org/croissant/1.0", "name": "ops-lite", "description": "A curated 500-case root-cause-analysis (RCA) evaluation set for microservice systems. Each case bundles a chaos-injection ground truth, a manifest-derived causal service graph, runtime environment snapshots, and 12 parquet metric tables (abnormal + normal windows). The corpus spans three open-source microservice testbeds: Train-Ticket (ts), Hotel Reservation from DeathStarBench (hs), and the OpenTelemetry Demo application (otel-demo). Faults are injected and observed by AegisLab; ground-truth causal graphs are produced by a manifest-driven layer-by-layer fault-propagation reasoner (rcabench-platform v3 internal/reasoning).", "url": "https://huggingface.co/datasets/anon-ops/ops-lite", "version": "1.0.0", "datePublished": "2026-05-04", "license": "https://www.apache.org/licenses/LICENSE-2.0", "citeAs": "@misc{opslite2026, title={ops-lite: A Curated RCA Evaluation Set for Microservice Systems}, year={2026}, howpublished={\\url{https://huggingface.co/datasets/anon-ops/ops-lite}}}", "creator": { "@type": "Organization", "name": "anon-ops" }, "keywords": [ "root-cause-analysis", "microservices", "observability", "AIOps", "chaos-engineering", "causal-graph", "benchmark" ], "isLiveDataset": false, "distribution": [ { "@type": "cr:FileObject", "@id": "repo", "name": "repo", "description": "The Hugging Face git repository.", "contentUrl": "https://huggingface.co/datasets/anon-ops/ops-lite/tree/refs%2Fconvert%2Fparquet", "encodingFormat": "git+https", "sha256": "https://github.com/mlcommons/croissant/issues/80" }, { "@type": "cr:FileSet", "@id": "parquet-files-for-config-default", "containedIn": { "@id": "repo" }, "encodingFormat": "application/x-parquet", "includes": "default/*/*.parquet" }, { "@type": "cr:FileObject", "@id": "manifest.jsonl", "name": "manifest.jsonl", "description": "Case-level manifest for the published dataset release.", "contentUrl": "https://huggingface.co/datasets/anon-ops/ops-lite/resolve/main/manifest.jsonl", "encodingFormat": "application/jsonlines", "sha256": "5d4d3960446408a6e43ea87c51a255aa7a39d43a1db6aeb731cf20b18f4fb7cd" } ], "recordSet": [ { "@type": "cr:RecordSet", "@id": "cases_index", "name": "cases_index", "description": "One record per curated RCA case.", "field": [ { "@type": "cr:Field", "@id": "cases_index/name", "name": "name", "description": "Stable case identifier; also the directory name under cases/.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.name" } } }, { "@type": "cr:Field", "@id": "cases_index/system", "name": "system", "description": "Microservice testbed: ts (Train-Ticket), hs (Hotel Reservation / DeathStarBench), or otel-demo (OpenTelemetry Demo).", "dataType": "sc:Text", "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.system" } } }, { "@type": "cr:Field", "@id": "cases_index/longest_path", "name": "longest_path", "description": "Longest fault-propagation path length in the ground-truth causal graph.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.longest_path" } } }, { "@type": "cr:Field", "@id": "cases_index/n_svc", "name": "n_svc", "description": "Number of services in the case's causal graph.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.n_svc" } } }, { "@type": "cr:Field", "@id": "cases_index/n_edge", "name": "n_edge", "description": "Number of edges in the case's causal graph.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.n_edge" } } }, { "@type": "cr:Field", "@id": "cases_index/n_alarm_svc", "name": "n_alarm_svc", "description": "Number of services that produced detector alarms during the abnormal window.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.n_alarm_svc" } } }, { "@type": "cr:Field", "@id": "cases_index/root_services", "name": "root_services", "description": "Ground-truth root-cause services (the chaos injection target(s)).", "dataType": "sc:Text", "repeated": true, "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.root_services[*]" } } }, { "@type": "cr:Field", "@id": "cases_index/chaos_family", "name": "chaos_family", "description": "Coarse chaos family: JVM*, HTTP*, Network*, Pod*, *Stress, DNS, Time, hybrid_clean, hybrid_kill.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.chaos_family" } } }, { "@type": "cr:Field", "@id": "cases_index/primary_kind", "name": "primary_kind", "description": "Specific chaos type for non-hybrid cases, otherwise 'hybrid'.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.primary_kind" } } }, { "@type": "cr:Field", "@id": "cases_index/subtypes", "name": "subtypes", "description": "Sorted unique chaos_types present in the case (one per leg for hybrid cases).", "dataType": "sc:Text", "repeated": true, "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.subtypes[*]" } } }, { "@type": "cr:Field", "@id": "cases_index/hybrid", "name": "hybrid", "description": "True if the case combines two or more chaos legs.", "dataType": "sc:Boolean", "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.hybrid" } } }, { "@type": "cr:Field", "@id": "cases_index/has_kill_leg", "name": "has_kill_leg", "description": "True if any leg of the injection is a pod-kill style fault.", "dataType": "sc:Boolean", "source": { "fileObject": { "@id": "manifest.jsonl" }, "extract": { "jsonPath": "$.has_kill_leg" } } } ] } ], "rai:dataCollection": "Each case originates from a controlled chaos-engineering experiment on one of three open-source microservice testbeds (Train-Ticket, Hotel Reservation from DeathStarBench, and the OpenTelemetry Demo application). Faults are injected via Chaos Mesh through AegisLab, an RCA benchmarking platform, while the system processes synthetic load. Service-level metrics (latency, error rate, resource), traces, and logs are recorded for both an 'abnormal' window (during fault) and a 'normal' baseline window. Only cases for which the AegisLab detector confirmed end-to-end observability of the injected fault enter the candidate pool.", "rai:dataCollectionType": [ "Synthetic / Simulation", "Automated machine instrumentation" ], "rai:dataCollectionTimeframe": { "startDate": "2026-04-20", "endDate": "2026-05-03" }, "rai:dataAnnotationProtocol": "Ground-truth causal graphs are produced automatically by a manifest-driven fault-propagation reasoner (rcabench-platform v3, internal/reasoning). For each case the reasoner reads the chaos injection's fault contract from a registered fault manifest and enumerates layer-by-layer the services that the contract says can be affected, terminating when no further nodes match the contract's entry signatures. The resulting service graph is what the fault contract itself prescribes as propagation - it is not a post-filtered exhaustive walk. Cases whose resulting graph is cyclic, has longest_path <= 1, or whose injection target is a frontend / load-generator service are dropped. The kept pool is reduced to 500 by a greedy selector with hard caps on system, chaos family, and root service.", "rai:dataAnnotationPlatform": "rcabench-platform v3 internal/reasoning module (manifest-driven layer expansion); curation script in the AegisLab repository.", "rai:dataAnnotationAnalysis": "Annotation is fully algorithmic and deterministic given the fault manifest, so per-annotator agreement is not applicable. Quality is governed by the manifest itself: every chaos type used in this corpus has an entry signature and propagation rule registered in rcabench-platform's manifest, and the graph for any case can be regenerated from the injection.json and the manifest version. Aggregate graph-shape statistics for the released set are: mean longest_path 3.18, mean n_edge 4.06, mean n_svc 3.97.", "rai:annotationsPerItem": "1 (deterministic, manifest-derived).", "rai:annotatorDemographics": "Not applicable - annotations are produced by software, not humans.", "rai:machineAnnotationTools": [ "rcabench-platform v3 internal/reasoning (manifest-driven causal-graph reasoner)" ], "rai:personalSensitiveInformation": "None. The dataset contains only synthetic load and fault telemetry from open-source microservice testbeds running in isolated lab clusters. No real user data, PII, payment data, or production traffic is present. Service / endpoint names come from the public Train-Ticket, Hotel Reservation / DeathStarBench, and OpenTelemetry Demo projects.", "rai:dataBiases": "The corpus is biased toward fault types that the AegisLab detector can confirm end-to-end (predominantly latency / error-rate / resource-stress / pod-availability faults). Topology-only or silent-data-corruption faults are under-represented. System coverage is also uneven: ts (Train-Ticket) is over-represented (320 / 500) compared to hs (142) and otel-demo (38), reflecting both the richer fault surface of the Train-Ticket testbed and the manifest reasoner being stricter about JVM / *Stress entry signatures, which disproportionately drops otel-demo cases. Hybrid (multi-leg) cases account for 43 % of the corpus, which is higher than typical production incident distributions.", "rai:dataUseCases": "Recommended uses: (1) benchmarking microservice RCA algorithms - root-cause ranking, propagation-path inference, alarm clustering, anomaly detection on service metrics; (2) ablation studies on graph-shape difficulty (longest_path, n_edge); (3) supervision data for graph-structured RCA models. Not recommended: (1) training or evaluating production-incident triage on real customer traffic - the synthetic load and lab setting do not reflect production tail behavior; (2) general-purpose anomaly detection unrelated to microservice fault propagation; (3) any task requiring real user, employee, or business data.", "rai:dataLimitations": "Each case captures a single fault window in an isolated lab cluster, so the dataset cannot exercise (a) concurrent independent faults, (b) noisy-neighbour effects from co-tenant workloads, (c) long-running gradual degradations beyond the recording window, or (d) fault propagation through queues / async pipelines that the underlying testbeds do not model. Causal graphs are only as accurate as the manifest's propagation rules - faults whose real-world effect exceeds the manifest's declared contract will have ground-truth graphs that under-represent the actual blast radius. The 500-case sample is intentionally small for fast benchmark turnaround; statistical claims based on this corpus alone should be treated as evaluation evidence, not population estimates.", "rai:dataSocialImpact": "The dataset is intended to advance reproducible RCA research for microservice systems. Risks are low: there is no human or operational data. Possible indirect risk: an RCA algorithm tuned only on this benchmark may transfer poorly to production fault distributions, potentially giving operators false confidence. Users are encouraged to report any observed gaps between benchmark performance and production performance via the dataset's HF discussion page.", "rai:dataReleaseMaintenancePlan": "Versioned releases on Hugging Face under anon-ops/ops-lite. The pipeline that produced this release (chaos injection in AegisLab + manifest-driven reasoner in rcabench-platform v3) is open-source, so any reissue can be reproduced from the same fault manifest commit. We plan to refresh the corpus when the fault manifest gains new chaos families. No deprecation timeline is set; older versions remain accessible via repo history." }