elasticsearch-anomaly-detection

Elasticsearch Anomaly Detection

Create, open, and start ML anomaly detection jobs on time-series data. Choose the right count-family detector direction, configure bucket span and time field, wire the datafeed to the correct index, and confirm running state from stats — not from assumptions.

Environment Configuration

This skill executes Elasticsearch operations through the elastic CLI. If the elastic CLI is not installed, tell the user what it is needed for. Do not guess credentials, call the HTTP API directly, or attempt other workarounds.

This skill references operations in HTTP-shorthand form (e.g., GET /, GET /_cat/indices, GET /{index}/_mapping, GET /{index}/_settings/index.mode, POST /_query). The Operations table at the end of this document maps each shorthand to the equivalent elastic CLI command — always use the CLI rather than calling the HTTP API directly.

Prerequisite: ML anomaly detection requires a Platinum-equivalent license on self-managed clusters. Serverless projects include ML. The caller needs manage_ml to create and manage jobs.

Related skill: For interpreting anomaly scores, influencers, and model behavior after a job is running, use elasticsearch-anomaly-detection-explainer — not this skill.

Process

1. Discover the target index and time field. List candidate indices with GET /_cat/indices (pass a pattern when the user names one). Fetch field types for the chosen index with GET /{index}/_mapping. The decision: confirm the index exists, identify the time field (often @timestamp), and verify document volume is sufficient for baseline learning. Never guess index or field names — they vary across deployments.

2. Choose detector function and direction. Match the user's intent to a count-family detector in analysis_config.detectors:

   • Spike, surge, unusual increase in event volume → high_count (or count, which flags both directions but is acceptable when the user cares about spikes). Do not use low_count — it will miss spikes.
   • Drop, outage, absence of events, traffic stops → low_count. Do not use high_count — it will miss drops and silence.
   • Metric deviation (CPU, latency, a numeric field) → mean-family functions (mean, high_mean, low_mean) with field_name set — only when the user asks about a numeric metric, not raw event volume.