BOOM: Benchmark of Observability Metrics
Source
- Dataset metadata snapshot: source.md
- Metadata JSON: metadata.json
- Official Hugging Face: https://huggingface.co/datasets/Datadog/BOOM
- Official leaderboard: https://huggingface.co/spaces/Datadog/BOOM
- Official code: https://github.com/DataDog/toto/tree/main/boom
- Introducing paper: Toto
Core Claim
BOOM is Datadog’s observability forecasting benchmark for evaluating models on high-cardinality operational metrics. It is the main dataset reason that Toto belongs in high-dimensional forecasting discussions even though its dimensionality regime is smaller than Time-HD.
Dataset Notes
- BOOM contains about 350 million time-series points across 2,807 metric queries.
- The Hugging Face card reports 32,887 variates, with each dataset entry containing one metric query and up to 100 variates.
- Metric-query groups become related variates in one multivariate time series.
- Domain labels include application usage, infrastructure, database, networking, and security.
- Metric types include gauge, rate, distribution, and count.
- The Toto paper also defines BOOMlet as a smaller representative subset with 32 metric queries, 1,627 variates, and about 23 million observation points.
Why It Matters
BOOM is the strongest current dataset anchor for observability-style high-dimensional forecasting in this repository. It captures high cardinality, nonstationarity, missing intervals, sparse spikes, heavy tails, and scale changes in grouped metric series.
Limitations
- BOOM is passive forecasting data. It does not include deployments, rollbacks, autoscaling changes, traffic-control commands, remediations, or other operator actions as forecast-conditioning channels.
- It comes from Datadog internal pre-production monitoring, so transfer to other observability stacks should be checked empirically.