elasticsearch-exporter

Elasticsearch doesn't speak Prometheus. It exposes its internals over its own REST API — GET /_nodes/stats, /_cluster/health, /_cat/indices — as JSON, in Elastic's own shape. Prometheus can't scrape that.

The elasticsearch-exporter is a small sidecar-or-Deployment-shaped translator that sits between the two:

• it polls those ES REST endpoints on an interval,
• flattens the JSON into Prometheus text-format metrics,
• and serves them on /metrics (conventionally :9114) for Prometheus to scrape.

The canonical implementation is prometheus-community/elasticsearch_exporter (formerly justwatchcom/elasticsearch_exporter), packaged as the prometheus-elasticsearch-exporter Helm chart. Elastic also ships a first-party alternative path — Metricbeat's elasticsearch module, or the newer Elastic Agent integration — but those ship into Elasticsearch/Kibana's own monitoring cluster, not into Prometheus, so they don't help a Grafana-alerting-on-Prometheus setup.

The metrics it produces are for example:

• metric
• what it gives you 

• search_jvm_memory_used_bytes{area="heap"}
• actual JVM heap in use — the thing that actually predicts an ES OOM
• elasticsearch_jvm_memory_max_bytes{area="heap"}
• configured heap ceiling (-Xmx), so you can take a real ratio
• elasticsearch_jvm_gc_collection_seconds_*
• GC pressure; sustained old-gen GC is the pre-OOM tell
• elasticsearch_breakers_tripped
• circuit breakers firing — ES rejecting work to avoid OOM
• elasticsearch_cluster_health_status
• green/yellow/red, unassigned shards                                 │

The exporter is a separate deployable — ECK does not install it.