Deploy the model in your cluster

After you import the model and vocabulary, you can use Kibana to view and manage their deployment across your cluster under Machine Learning > Model Management. Alternatively, you can use the start trained model deployment API.

You can deploy a model multiple times by assigning a unique deployment ID when starting the deployment. It enables you to have dedicated deployments for different purposes, such as search and ingest. By doing so, you ensure that the search speed remains unaffected by ingest workloads, and vice versa. Having separate deployments for search and ingest mitigates performance issues resulting from interactions between the two, which can be hard to diagnose.

It is recommended to fine-tune each deployment based on its specific purpose. To improve ingest performance, increase throughput by adding more allocations to the deployment. For improved search speed, increase the number of threads per allocation.

When you deploy the model, its allocations are distributed across available machine learning nodes. Model allocations are independent units of work for NLP tasks. To influence model performance, you can configure the number of allocations and the number of threads used by each allocation of your deployment. Alternatively, you can enable Adaptive allocations to automatically create and remove model allocations based on the current workload of the model (you still need to manually set the number of threads).

Throughput can be scaled by adding more allocations to the deployment; it increases the number of inference requests that can be performed in parallel. All allocations assigned to a node share the same copy of the model in memory. The model is loaded into memory in a native process that encapsulates libtorch, which is the underlying machine learning library of PyTorch. The number of allocations setting affects the amount of model allocations across all the machine learning nodes. Model allocations are distributed in such a way that the total number of used threads does not exceed the allocated processors of a node.

The threads per allocation setting affects the number of threads used by each model allocation during inference. Increasing the number of threads generally increases the speed of inference requests. The value of this setting must not exceed the number of available allocated processors per node.

You can view the allocation status in Kibana or by using the get trained model stats API. If you want to change the number of allocations, you can use the update trained model stats API after the allocation status is started. You can also enable Adaptive allocations to automatically create and remove model allocations based on the current workload of the model.

The numbers of threads and allocations you can set manually for a model remain constant even when not all the available resources are fully used or when the load on the model requires more resources. Instead of setting the number of allocations manually, you can enable adaptive allocations to set the number of allocations based on the load on the process. This can help you to manage performance and cost more easily. When adaptive allocations are enabled, the number of allocations of the model is set automatically based on the current load. When the load is high, a new model allocation is automatically created. When the load is low, a model allocation is automatically removed.

You can enable adaptive allocations by using:

If the new allocations fit on the current machine learning nodes, they are immediately started. If more resource capacity is needed for creating new model allocations, then your machine learning node will be scaled up if machine learning autoscaling is enabled to provide enough resources for the new allocation. The number of model allocations cannot be scaled down to less than 1. And they cannot be scaled up to more than 32 allocations, unless you explicitly set the maximum number of allocations to more. Adaptive allocations must be set up independently for each deployment and inference endpoint.

Each allocation of a model deployment has a dedicated queue to buffer inference requests. The size of this queue is determined by the queue_capacity parameter in the start trained model deployment API. When the queue reaches its maximum capacity, new requests are declined until some of the queued requests are processed, creating available capacity once again. When multiple ingest pipelines reference the same deployment, the queue can fill up, resulting in rejected requests. Consider using dedicated deployments to prevent this situation.

Inference requests originating from search, such as the text_expansion query, have a higher priority compared to non-search requests. The inference ingest processor generates normal priority requests. If both a search query and an ingest processor use the same deployment, the search requests with higher priority skip ahead in the queue for processing before the lower priority ingest requests. This prioritization accelerates search responses while potentially slowing down ingest where response time is less critical.