The API exposed by APM Server is a regular HTTP JSON API. To achieve high availability, you can place multiple instances of APM Server behind a regular HTTP load balancer, for example HAProxy or nginx.

If an instance of APM Server should fail, the load balancer will remove it from the load balancing rotation after a short period of time.

APM Server endpoint /healthcheck always returns a HTTP 200. You can configure your load balancer to send HTTP requests to this endpoint to determine if a given APM Server is up and running.

APM Server maintains a small buffer internally. Data coming from agents is kept here until it can be delivered to Elasticsearch. Under normal circumstance, data is forwarded to Elasticsearch immediately. In situations in which Elasticsearch is down briefly or in situations where there is a sudden influx of data and Elasticsearch cannot ingest it all at once, the buffer acts as a temporary storage until the data can be ingested by Elasticsearch.

This also means that if a given APM Server process fails, for example because the machine its running on experiences an issue, the data that has not yet been forwarded to Elasticsearch is lost.