To enhance and customize your OpenSearch cluster for machine learning (ML), you can add and modify several configuration settings for the ML Commons plugin in your 'opensearch.yml' file.
To learn more about static and dynamic settings, see [Configuring OpenSearch]({{site.url}}{{site.baseurl}}/install-and-configure/configuring-opensearch/index/).
By default, ML tasks and models only run on ML nodes. When configured without the `data` node role, ML nodes do not store any shards and instead calculate resource requirements at runtime. To use an ML node, create a node in your `opensearch.yml` file. Give your node a custom name and define the node role as `ml`:
To set up a cluster with a dedicated ML node, see the sample [Docker compose file](https://github.com/opensearch-project/ml-commons/blob/main/docs/docker/docker-compose.yml).
If `true`, ML Commons tasks and models run ML tasks on ML nodes only. If `false`, tasks and models run on ML nodes first. If no ML nodes exist, tasks and models run on data nodes.
We suggest running ML workloads on a dedicated ML node rather than on data nodes. Starting with OpenSearch 2.5, ML tasks run on ML nodes only by default. To test models on a data node, set `plugins.ml_commons.only_run_on_ml_node` to `false`.
`round_robin` dispatches ML tasks to ML nodes using round robin routing. `least_load` gathers runtime information from all ML nodes, like JVM heap memory usage and running tasks, and then dispatches the tasks to the ML node with the lowest load.
When returning runtime information with the [Profile API]({{site.url}}{{site.baseurl}}/ml-commons-plugin/api/profile/), ML Commons will run a regular job to sync newly deployed or undeployed models on each node. When set to `0`, ML Commons immediately stops sync-up jobs.
Controls how many predict requests are monitored on one node. If set to `0`, OpenSearch clears all monitoring predict requests in cache and does not monitor for new predict requests.
This setting gives you the ability to register models using a URL. By default, ML Commons only allows registration of [pretrained]({{site.url}}{{site.baseurl}}//ml-commons-plugin/pretrained-models/) models from the OpenSearch model repository.
This setting gives you the ability to register a model using a local file. By default, ML Commons only allows registration of [pretrained]({{site.url}}{{site.baseurl}}//ml-commons-plugin/pretrained-models/) models from the OpenSearch model repository.
The default URL value for this trusted URL setting is not secure. For security, use you own regex string to the trusted repository that contains your models, for example `https://github.com/opensearch-project/ml-commons/blob/2.x/ml-algorithms/src/test/resources/org/opensearch/ml/engine/algorithms/text_embedding/*`.
Sets a circuit breaker that checks all system memory usage before running an ML task. If the native memory exceeds the threshold, OpenSearch throws an exception and stops running any ML task.
Values are based on the percentage of memory available. When set to `0`, no ML tasks will run. When set to `100`, the circuit breaker closes and no threshold exists.
Starting with OpenSearch 2.5, ML Commons runs a native memory circuit breaker to avoid an out-of-memory error when loading too many models. By default, the native memory threshold is 90%. If memory usage exceeds the threshold, ML Commons returns an error. For testing purposes, you can disable the circuit breaker by setting `plugins.ml_commons.native_memory_threshold` to 100.
This setting automatically redeploys deployed or partially deployed models upon cluster failure. If all ML nodes inside a cluster crash, the model switches to the `DEPLOYED_FAILED` state, and the model must be deployed manually.
This setting sets the limit for the number of times a deployed or partially deployed model will try and redeploy when ML nodes in a cluster fail or new ML nodes join the cluster.
This setting sets the ratio of success for the auto-redeployment of a model based on the available ML nodes in a cluster. For example, if ML nodes crash inside a cluster, the auto redeploy protocol adds another node or retires a crashed node. If the ratio is `0.7` and 70% of all ML nodes successfully redeploy the model on auto-redeploy activation, the redeployment is a success. If the model redeploys on fewer than 70% of available ML nodes, the auto-redeploy retries until the redeployment succeeds or OpenSearch reaches [the maximum number of retries](#set-retires-for-auto-redeploy).
When set to `true`, this setting enables the ability to run Python-based models supported by OpenSearch, such as [Metrics correlation]({{site.url}}{{site.baseurl}}/ml-commons-plugin/algorithms/#metrics-correlation).
