Alex Open Research Wiki

GAIA / MicroSS

2 min read

  • Normal
  • Not read

GAIA / MicroSS

Source

Core Claim

GAIA is an AIOps dataset collection for anomaly detection, log analysis, fault localization, and related tasks. MicroSS is the graph/system-relevant subset: a business-simulation microservice system with metrics, traces, business logs, and anomaly-injection records.

Dataset Notes

  • MicroSS contains more than 6500 metrics, more than 7000000 log items, and detailed trace data collected over an initial two-week window.
  • The trace schema includes service names, host IPs, trace IDs, span IDs, parent IDs, URLs, status codes, and messages.
  • Companion Data contains 406 anomaly-detection and metric-prediction series, including 279 labeled series, plus about 218736 log records.
  • MicroSS is system-level, but the public layout is not a single graph object. Service call structure is reconstructed from traces.

Why It Matters

GAIA/MicroSS is useful when a benchmark needs AIOps-style metrics, logs, traces, business logs, and anomaly-injection records from a whole microservice scenario. It is less graph-native than ChronoGraph but richer in trace/log context.

Gotchas

  • Companion Data should not be treated as a whole-service-graph dataset; it is closer to single-series KPI and log-task data.
  • License metadata is inconsistent: the repository LICENSE file is GPL-2.0, while the README license section says Apache 2.0.
  • Anomaly injections are exogenous benchmark events, not logged operator remediations.

Foundation TSFM Relevance

Agenda slotVerdictEvidenceMissing pieces
Benchmarks: what level of modeling is tested?partially closesMicroSS combines metrics, traces, business logs, system logs, and anomaly-injection records from a microservice scenario, making it a small observability benchmark for the digital-world robot north star.It is not packaged as a clean graph object and does not include remediation action histories.
Context interfaceadjacentTrace records include service names, host IPs, trace IDs, spans, parent IDs, URLs, status codes, and messages.Context must be reconstructed from raw traces/logs; no canonical channel metadata schema is provided.
Control and counterfactualswarningFault/anomaly injections are logged benchmark conditions.Injections are exogenous events, not operator actions or controllable interventions for policy learning.

Graph View

Backlinks