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.

   The decision: pick one primary detector whose direction matches the anomaly type. For volume spike/drop questions on document counts, stay in the count family — mean detectors are unsuited to "how many events" questions.

3. Set immutable job shape before creation. These fields cannot change after PUT /_ml/anomaly_detectors/{job_id}:

   • analysis_config.bucket_span — use the interval the user specifies (e.g. 15m for 15-minute buckets). Match the granularity of anomalies they care about; too short is noisy, too long is slow to detect.
   • data_description.time_field — the time field from the mapping (commonly @timestamp).
   • analysis_config.detectors — the function and direction from step 2.

   Example job body for a volume-spike detector:

   {
     "analysis_config": {
       "bucket_span": "15m",
       "detectors": [{ "function": "high_count" }]
     },
     "data_description": { "time_field": "@timestamp" }
   }
   

   Example for an outage / drop detector:

   {
     "analysis_config": {
       "bucket_span": "15m",
       "detectors": [{ "function": "low_count" }]
     },
     "data_description": { "time_field": "@timestamp" }
   }
   

4. Create the job. Call PUT /_ml/anomaly_detectors/{job_id} with the job id the user requested (or a descriptive id you propose). The job starts in closed state — creating it does not start analysis.

5. Create the datafeed. Call PUT /_ml/datafeeds/datafeed-{job_id} immediately after job creation. Set job_id to the same id, indices to the target index (exact name or pattern from step 1), and a query that selects the relevant documents (typically match_all). The datafeed id convention is datafeed-{job_id}.

   {
     "job_id": "{job_id}",
     "indices": ["{index}"],
     "query": { "match_all": {} }
   }
   

6. Open the job, then start the datafeed — in that order. This sequence is mandatory; do not skip or reorder:

   1. POST /_ml/anomaly_detectors/{job_id}/_open — transitions the job to opened.
   2. POST /_ml/datafeeds/datafeed-{job_id}/_start — transitions the datafeed to started.

   Opening before the datafeed exists fails. Starting the datafeed before opening the job fails. Do not report success after only creating resources — the job is not running until both are active.

7. Confirm running state from stats. Verify the outcome with:

   • GET /_ml/anomaly_detectors/{job_id}/_stats — expect state: "opened".
   • GET /_ml/datafeeds/datafeed-{job_id}/_stats — expect state: "started".

   Optionally call GET /_ml/anomaly_detectors/{job_id} to confirm configuration (detectors, bucket_span, time_field, datafeed indices). Report both stats states explicitly — "created" is not the same as "opened" and "started".

Teardown

When stopping or deleting a job, reverse the startup order:

1. POST /_ml/datafeeds/datafeed-{job_id}/_stop — stop the datafeed first.
2. POST /_ml/anomaly_detectors/{job_id}/_close — then close the job.

Stop the datafeed before closing the job. Close the job before resetting or deleting it.

Guidelines

• Required lifecycle order (create): job → datafeed → open job → start datafeed. Every new job follows this sequence.
• Detector direction is the highest-impact decision for volume anomalies. Re-read the user's wording: "spike", "surge", and "unusual increase" → high direction; "drop", "outage", "stops", "absence" → low direction.
• Immutable fields (bucket_span, detectors, time_field) require delete-and-recreate if wrong — validate mapping and intent before the first PUT.
• Datafeed index must match the user's target. Point indices at the exact index or pattern they named — not a nearby guess.
• Entity-level analysis (by_field_name, over_field_name, partition_field_name) and advanced tuning live in references/anomaly-detection-reference.md.

Full Reference

For API paths, request/response fields, score semantics, and field interactions, read references/anomaly-detection-reference.md.

Operations