Elastic Agent and Metricbeat are the recommended methods for collecting and shipping monitoring data to a monitoring cluster.

If you have previously configured legacy collection methods, you should migrate to using Elastic Agent or Metricbeat collection. Do not use legacy collection alongside other collection methods.

Collectors, as their name implies, collect things. Each collector runs once for each collection interval to obtain data from the public APIs in Elasticsearch and X-Pack that it chooses to monitor. When the data collection is finished, the data is handed in bulk to the exporters to be sent to the monitoring clusters. Regardless of the number of exporters, each collector only runs once per collection interval.

There is only one collector per data type gathered. In other words, for any monitoring document that is created, it comes from a single collector rather than being merged from multiple collectors. The Elasticsearch monitoring features currently have a few collectors because the goal is to minimize overlap between them for optimal performance.

Each collector can create zero or more monitoring documents. For example, the index_stats collector collects all index statistics at the same time to avoid many unnecessary calls.

Collector Data Types Description

Cluster Stats


Gathers details about the cluster state, including parts of the actual cluster state (for example GET /_cluster/state) and statistics about it (for example, GET /_cluster/stats). This produces a single document type. In versions prior to X-Pack 5.5, this was actually three separate collectors that resulted in three separate types: cluster_stats, cluster_state, and cluster_info. In 5.5 and later, all three are combined into cluster_stats. This only runs on the elected master node and the data collected (cluster_stats) largely controls the UI. When this data is not present, it indicates either a misconfiguration on the elected master node, timeouts related to the collection of the data, or issues with storing the data. Only a single document is produced per collection.

Index Stats

indices_stats, index_stats

Gathers details about the indices in the cluster, both in summary and individually. This creates many documents that represent parts of the index statistics output (for example, GET /_stats). This information only needs to be collected once, so it is collected on the elected master node. The most common failure for this collector relates to an extreme number of indices — and therefore time to gather them — resulting in timeouts. One summary indices_stats document is produced per collection and one index_stats document is produced per index, per collection.

Index Recovery


Gathers details about index recovery in the cluster. Index recovery represents the assignment of shards at the cluster level. If an index is not recovered, it is not usable. This also corresponds to shard restoration via snapshots. This information only needs to be collected once, so it is collected on the elected master node. The most common failure for this collector relates to an extreme number of shards — and therefore time to gather them — resulting in timeouts. This creates a single document that contains all recoveries by default, which can be quite large, but it gives the most accurate picture of recovery in the production cluster.



Gathers details about all allocated shards for all indices, particularly including what node the shard is allocated to. This information only needs to be collected once, so it is collected on the elected master node. The collector uses the local cluster state to get the routing table without any network timeout issues unlike most other collectors. Each shard is represented by a separate monitoring document.



Gathers details about all machine learning job statistics (for example, GET /_ml/anomaly_detectors/_stats). This information only needs to be collected once, so it is collected on the elected master node. However, for the master node to be able to perform the collection, the master node must have set to true (default) and a license level that supports machine learning.

Node Stats


Gathers details about the running node, such as memory utilization and CPU usage (for example, GET /_nodes/_local/stats). This runs on every node with monitoring features enabled. One common failure results in the timeout of the node stats request due to too many segment files. As a result, the collector spends too much time waiting for the file system stats to be calculated until it finally times out. A single node_stats document is created per collection. This is collected per node to help to discover issues with nodes communicating with each other, but not with the monitoring cluster (for example, intermittent network issues or memory pressure).

The Elasticsearch monitoring features use a single threaded scheduler to run the collection of Elasticsearch monitoring data by all of the appropriate collectors on each node. This scheduler is managed locally by each node and its interval is controlled by specifying the xpack.monitoring.collection.interval, which defaults to 10 seconds (10s), at either the node or cluster level.

Fundamentally, each collector works on the same principle. Per collection interval, each collector is checked to see whether it should run and then the appropriate collectors run. The failure of an individual collector does not impact any other collector.

Once collection has completed, all of the monitoring data is passed to the exporters to route the monitoring data to the monitoring clusters.

If gaps exist in the monitoring charts in Kibana, it is typically because either a collector failed or the monitoring cluster did not receive the data (for example, it was being restarted). In the event that a collector fails, a logged error should exist on the node that attempted to perform the collection.

Collection is currently done serially, rather than in parallel, to avoid extra overhead on the elected master node. The downside to this approach is that collectors might observe a different version of the cluster state within the same collection period. In practice, this does not make a significant difference and running the collectors in parallel would not prevent such a possibility.

For more information about the configuration options for the collectors, see Monitoring collection settings.

Collecting data from across the Elastic Stackedit

Elasticsearch monitoring features also receive monitoring data from other parts of the Elastic Stack. In this way, it serves as an unscheduled monitoring data collector for the stack.

By default, data collection is disabled. Elasticsearch monitoring data is not collected and all monitoring data from other sources such as Kibana, Beats, and Logstash is ignored. You must set xpack.monitoring.collection.enabled to true to enable the collection of monitoring data. See Monitoring settings.

Once data is received, it is forwarded to the exporters to be routed to the monitoring cluster like all monitoring data.

Because this stack-level "collector" lives outside of the collection interval of Elasticsearch monitoring features, it is not impacted by the xpack.monitoring.collection.interval setting. Therefore, data is passed to the exporters whenever it is received. This behavior can result in indices for Kibana, Logstash, or Beats being created somewhat unexpectedly.

While the monitoring data is collected and processed, some production cluster metadata is added to incoming documents. This metadata enables Kibana to link the monitoring data to the appropriate cluster. If this linkage is unimportant to the infrastructure that you’re monitoring, it might be simpler to configure Logstash and Beats to report monitoring data directly to the monitoring cluster. This scenario also prevents the production cluster from adding extra overhead related to monitoring data, which can be very useful when there are a large number of Logstash nodes or Beats.

For more information about typical monitoring architectures, see How it works.