OpenSearch/docs/reference/monitoring/collectors.asciidoc

[role="xpack"]
[testenv="basic"]
[[es-monitoring-collectors]]
== Collectors

[IMPORTANT]
=========================
{metricbeat} is the recommended method for collecting and shipping monitoring
data to a monitoring cluster.

If you have previously configured legacy collection methods, you should migrate
to using {metricbeat} collection methods. Use either {metricbeat} collection or
legacy collection methods; do not use both.

Learn more about <<configuring-metricbeat>>.
=========================

Collectors, as their name implies, collect things. Each collector runs once for
each collection interval to obtain data from the public APIs in {es} and {xpack}
that it chooses to monitor. When the data collection is finished, the data is
handed in bulk to the <<es-monitoring-exporters,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 {es} {monitor-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.

[options="header"]
|=======================
| Collector       | Data Types | Description
| Cluster Stats   | `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  | `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.
| Shards          | `shards`
| 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.
| Jobs            | `job_stats`
| 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
`xpack.ml.enabled` set to true (default) and a license level that supports {ml}.
| Node Stats      | `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
{monitor-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 {es} {monitor-features} use a single threaded scheduler to run the
collection of {es} 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 {kib}, 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.

NOTE: 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>>.

[discrete]
[[es-monitoring-stack]]
==== Collecting data from across the Elastic Stack

{es} {monitor-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. {es} monitoring data is not
collected and all monitoring data from other sources such as {kib}, 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.

WARNING: Because this stack-level "collector" lives outside of the collection
interval of {es} {monitor-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 {kib},
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 {kib} 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-monitoring-works>>.