mirror of
https://github.com/elastic/elasticsearch.git
synced 2025-04-25 07:37:19 -04:00
165e063b50
When running ML, sometimes it is best to automatically adjust the memory allotted for machine learning based on the nodesize and how much space is given to the JVM This commit adds a new static setting xpack.ml.use_auto_machine_memory_percent for allowing this dynamic calculation. The old setting remains as a backup just in case the limit cannot be automatically determined due to lack of information. Closes #63795
192 lines
9.2 KiB
Text
192 lines
9.2 KiB
Text
|
|
[role="xpack"]
|
|
[[ml-settings]]
|
|
=== Machine learning settings in Elasticsearch
|
|
++++
|
|
<titleabbrev>Machine learning settings</titleabbrev>
|
|
++++
|
|
|
|
[[ml-settings-description]]
|
|
// tag::ml-settings-description-tag[]
|
|
You do not need to configure any settings to use {ml}. It is enabled by default.
|
|
|
|
IMPORTANT: {ml-cap} uses SSE4.2 instructions, so it works only on machines whose
|
|
CPUs {wikipedia}/SSE4#Supporting_CPUs[support] SSE4.2. If you run {es} on older
|
|
hardware, you must disable {ml} (by setting `xpack.ml.enabled` to `false`).
|
|
|
|
// end::ml-settings-description-tag[]
|
|
|
|
[discrete]
|
|
[[general-ml-settings]]
|
|
==== General machine learning settings
|
|
|
|
`node.roles: [ ml ]`::
|
|
(<<static-cluster-setting,Static>>) Set `node.roles` to contain `ml` to identify
|
|
the node as a _{ml} node_ that is capable of running jobs. Every node is a {ml}
|
|
node by default.
|
|
+
|
|
If you use the `node.roles` setting, then all required roles must be explicitly
|
|
set. Consult <<modules-node>> to learn more.
|
|
+
|
|
IMPORTANT: On dedicated coordinating nodes or dedicated master nodes, do not set
|
|
the `ml` role.
|
|
|
|
|
|
`xpack.ml.enabled`::
|
|
(<<static-cluster-setting,Static>>) Set to `true` (default) to enable {ml} APIs
|
|
on the node.
|
|
+
|
|
If set to `false`, the {ml} APIs are disabled on the node. Therefore the node
|
|
cannot open jobs, start {dfeeds}, or receive transport (internal) communication
|
|
requests related to {ml} APIs. If the node is a coordinating node, {ml} requests
|
|
from clients (including {kib}) also fail. For more information about disabling
|
|
{ml} in specific {kib} instances, see
|
|
{kibana-ref}/ml-settings-kb.html[{kib} {ml} settings].
|
|
+
|
|
IMPORTANT: If you want to use {ml-features} in your cluster, it is recommended
|
|
that you set `xpack.ml.enabled` to `true` on all nodes. This is the default
|
|
behavior. At a minimum, it must be enabled on all master-eligible nodes. If you
|
|
want to use {ml-features} in clients or {kib}, it must also be enabled on all
|
|
coordinating nodes.
|
|
|
|
`xpack.ml.inference_model.cache_size`::
|
|
(<<static-cluster-setting,Static>>) The maximum inference cache size allowed.
|
|
The inference cache exists in the JVM heap on each ingest node. The cache
|
|
affords faster processing times for the `inference` processor. The value can be
|
|
a static byte sized value (i.e. "2gb") or a percentage of total allocated heap.
|
|
The default is "40%". See also <<model-inference-circuit-breaker>>.
|
|
|
|
[[xpack-interference-model-ttl]]
|
|
// tag::interference-model-ttl-tag[]
|
|
`xpack.ml.inference_model.time_to_live` {ess-icon}::
|
|
(<<static-cluster-setting,Static>>) The time to live (TTL) for models in the
|
|
inference model cache. The TTL is calculated from last access. The `inference`
|
|
processor attempts to load the model from cache. If the `inference` processor
|
|
does not receive any documents for the duration of the TTL, the referenced model
|
|
is flagged for eviction from the cache. If a document is processed later, the
|
|
model is again loaded into the cache. Defaults to `5m`.
|
|
// end::interference-model-ttl-tag[]
|
|
|
|
`xpack.ml.max_inference_processors`::
|
|
(<<cluster-update-settings,Dynamic>>) The total number of `inference` type
|
|
processors allowed across all ingest pipelines. Once the limit is reached,
|
|
adding an `inference` processor to a pipeline is disallowed. Defaults to `50`.
|
|
|
|
`xpack.ml.max_machine_memory_percent`::
|
|
(<<cluster-update-settings,Dynamic>>) The maximum percentage of the machine's
|
|
memory that {ml} may use for running analytics processes. (These processes are
|
|
separate to the {es} JVM.) Defaults to `30` percent. The limit is based on the
|
|
total memory of the machine, not current free memory. Jobs are not allocated to
|
|
a node if doing so would cause the estimated memory use of {ml} jobs to exceed
|
|
the limit.
|
|
|
|
`xpack.ml.max_model_memory_limit`::
|
|
(<<cluster-update-settings,Dynamic>>) The maximum `model_memory_limit` property
|
|
value that can be set for any job on this node. If you try to create a job with
|
|
a `model_memory_limit` property value that is greater than this setting value,
|
|
an error occurs. Existing jobs are not affected when you update this setting.
|
|
For more information about the `model_memory_limit` property, see
|
|
<<put-analysislimits>>.
|
|
|
|
[[xpack.ml.max_open_jobs]]
|
|
`xpack.ml.max_open_jobs`::
|
|
(<<cluster-update-settings,Dynamic>>) The maximum number of jobs that can run
|
|
simultaneously on a node. Defaults to `20`. In this context, jobs include both
|
|
{anomaly-jobs} and {dfanalytics-jobs}. The maximum number of jobs is also
|
|
constrained by memory usage. Thus if the estimated memory usage of the jobs
|
|
would be higher than allowed, fewer jobs will run on a node. Prior to version
|
|
7.1, this setting was a per-node non-dynamic setting. It became a cluster-wide
|
|
dynamic setting in version 7.1. As a result, changes to its value after node
|
|
startup are used only after every node in the cluster is running version 7.1 or
|
|
higher. The maximum permitted value is `512`.
|
|
|
|
`xpack.ml.nightly_maintenance_requests_per_second`::
|
|
(<<cluster-update-settings,Dynamic>>) The rate at which the nightly maintenance task
|
|
deletes expired model snapshots and results. The setting is a proxy to the
|
|
[requests_per_second](https://www.elastic.co/guide/en/elasticsearch/reference/current/docs-delete-by-query.html#_throttling_delete_requests)
|
|
parameter used in the Delete by query requests and controls throttling.
|
|
Valid values must be greater than `0.0` or equal to `-1.0` where `-1.0` means a default value
|
|
is used. Defaults to `-1.0`
|
|
|
|
`xpack.ml.node_concurrent_job_allocations`::
|
|
(<<cluster-update-settings,Dynamic>>) The maximum number of jobs that can
|
|
concurrently be in the `opening` state on each node. Typically, jobs spend a
|
|
small amount of time in this state before they move to `open` state. Jobs that
|
|
must restore large models when they are opening spend more time in the `opening`
|
|
state. Defaults to `2`.
|
|
|
|
[discrete]
|
|
[[advanced-ml-settings]]
|
|
==== Advanced machine learning settings
|
|
|
|
These settings are for advanced use cases; the default values are generally
|
|
sufficient:
|
|
|
|
`xpack.ml.enable_config_migration`::
|
|
(<<cluster-update-settings,Dynamic>>) Reserved.
|
|
|
|
`xpack.ml.max_anomaly_records`::
|
|
(<<cluster-update-settings,Dynamic>>) The maximum number of records that are
|
|
output per bucket. The default value is `500`.
|
|
|
|
`xpack.ml.max_lazy_ml_nodes`::
|
|
(<<cluster-update-settings,Dynamic>>) The number of lazily spun up {ml} nodes.
|
|
Useful in situations where {ml} nodes are not desired until the first {ml} job
|
|
opens. It defaults to `0` and has a maximum acceptable value of `3`. If the
|
|
current number of {ml} nodes is greater than or equal to this setting, it is
|
|
assumed that there are no more lazy nodes available as the desired number
|
|
of nodes have already been provisioned. If a job is opened and this setting has
|
|
a value greater than zero and there are no nodes that can accept the job, the
|
|
job stays in the `OPENING` state until a new {ml} node is added to the cluster
|
|
and the job is assigned to run on that node.
|
|
+
|
|
IMPORTANT: This setting assumes some external process is capable of adding {ml}
|
|
nodes to the cluster. This setting is only useful when used in conjunction with
|
|
such an external process.
|
|
|
|
`xpack.ml.process_connect_timeout`::
|
|
(<<cluster-update-settings,Dynamic>>) The connection timeout for {ml} processes
|
|
that run separately from the {es} JVM. Defaults to `10s`. Some {ml} processing
|
|
is done by processes that run separately to the {es} JVM. When such processes
|
|
are started they must connect to the {es} JVM. If such a process does not
|
|
connect within the time period specified by this setting then the process is
|
|
assumed to have failed. Defaults to `10s`. The minimum value for this setting is
|
|
`5s`.
|
|
|
|
xpack.ml.use_auto_machine_memory_percent::
|
|
(<<static-cluster-setting,Static>>) If this setting is `true`, the
|
|
`xpack.ml.max_machine_memory_percent` setting is ignored. Instead, the maximum
|
|
percentage of the machine's memory that can be used for running {ml} analytics
|
|
processes is calculated automatically and takes into account the total node size
|
|
and the size of the JVM on the node. The default value is `false`. If this
|
|
setting differs between nodes, the value on the current master node is heeded.
|
|
+
|
|
TIP: If you do not have dedicated {ml} nodes (that is to say, the node has
|
|
multiple roles), do not enable this setting. Its calculations assume that {ml}
|
|
analytics are the main purpose of the node.
|
|
+
|
|
IMPORTANT: The calculation assumes that dedicated {ml} nodes have at least
|
|
`256MB` memory reserved outside of the JVM. If you have tiny {ml}
|
|
nodes in your cluster, you shouldn't use this setting.
|
|
|
|
[discrete]
|
|
[[model-inference-circuit-breaker]]
|
|
==== {ml-cap} circuit breaker settings
|
|
|
|
`breaker.model_inference.limit`::
|
|
(<<cluster-update-settings,Dynamic>>) Limit for the model inference breaker,
|
|
which defaults to 50% of the JVM heap. If the parent circuit breaker is less
|
|
than 50% of the JVM heap, it is bound to that limit instead. See
|
|
<<circuit-breaker>>.
|
|
|
|
`breaker.model_inference.overhead`::
|
|
(<<cluster-update-settings,Dynamic>>) A constant that all accounting estimations
|
|
are multiplied by to determine a final estimation. Defaults to 1. See
|
|
<<circuit-breaker>>.
|
|
|
|
`breaker.model_inference.type`::
|
|
(<<static-cluster-setting,Static>>) The underlying type of the circuit breaker.
|
|
There are two valid options: `noop` and `memory`. `noop` means the circuit
|
|
breaker does nothing to prevent too much memory usage. `memory` means the
|
|
circuit breaker tracks the memory used by inference models and can potentially
|
|
break and prevent `OutOfMemory` errors. The default is `memory`.
|