databricksmodelservingprovisionedthroughput Resource

June 12, 2026 ยท View on GitHub

API Documentation

This resource allows you to manage Foundation Model provisioned throughput endpoints in Databricks.

~> This resource is currently in private preview, and only available for enrolled customers.

-> This resource can only be used with a workspace-level provider!

Example Usage

Creating a Foundation Model provisioned throughput endpoint

resource "databricks_model_serving_provisioned_throughput" "llama" {
  ai_gateway {
    usage_tracking_config {
      enabled = true
    }
  }
  config {
    served_entities {
      entity_name             = "system.ai.llama-4-maverick"
      entity_version          = "1"
      provisioned_model_units = 100
    }
  }
}

Argument Reference

The following arguments are supported:

  • name - (Required) The name of the model serving endpoint. This field is required and must be unique across a workspace. An endpoint name can consist of alphanumeric characters, dashes, and underscores. NOTE: Changing this name will delete the existing endpoint and create a new endpoint with the updated name.
  • config - The model serving endpoint configuration.
    • served_entities - A list of served entities for the endpoint to serve.
    • traffic_config - A single block represents the traffic split configuration amongst the served models.
  • tags - Tags to be attached to the serving endpoint and automatically propagated to billing logs.
  • ai_gateway - (Optional) A block with AI Gateway configuration for the serving endpoint. Note: only external model endpoints are supported as of now.
  • budget_policy_id - (Optiona) The Budget Policy ID set for this serving endpoint.
  • email_notifications - (Optional) A block with Email notification setting.
  • provider_config - (Optional) Configure the provider for management through account provider. This block consists of the following fields:
    • workspace_id - (Required) Workspace ID which the resource belongs to. This workspace must be part of the account which the provider is configured with.

served_entities Configuration Block

  • name - The name of a served entity. It must be unique across an endpoint. A served entity name can consist of alphanumeric characters, dashes, and underscores. If not specified for an external model, this field will be created from the entity_name and entity_version
  • entity_name - The full path of the UC model to be served, given in the form of catalog_name.schema_name.model_name.
  • entity_version - The version of the model in UC to be served.
  • provisioned_model_units - The number of model units to be provisioned.

traffic_config Configuration Block

  • routes - (Required) Each block represents a route that defines traffic to each served entity. Each served_entity block needs to have a corresponding routes block.
    • served_entity_name - (Required) The name of the served entity this route configures traffic for. This needs to match the name of a served_entity block.
    • traffic_percentage - (Required) The percentage of endpoint traffic to send to this route. It must be an integer between 0 and 100 inclusive.

tags Configuration Block

  • key - The key field for a tag.
  • value - The value field for a tag.

ai_gateway Configuration Block

  • guardrails - (Optional) Block with configuration for AI Guardrails to prevent unwanted data and unsafe data in requests and responses. Consists of the following attributes:
    • input - A block with configuration for input guardrail filters:
      • invalid_keywords - List of invalid keywords. AI guardrail uses keyword or string matching to decide if the keyword exists in the request or response content.
      • valid_topics - The list of allowed topics. Given a chat request, this guardrail flags the request if its topic is not in the allowed topics.
      • safety - the boolean flag that indicates whether the safety filter is enabled.
      • pii - Block with configuration for guardrail PII filter:
        • behavior - a string that describes the behavior for PII filter. Currently only BLOCK value is supported.
    • output - A block with configuration for output guardrail filters. Has the same structure as input block.
  • rate_limits - (Optional) Block describing rate limits for AI gateway. For details see the description of rate_limits block above.
  • usage_tracking_config - (Optional) Block with configuration for payload logging using inference tables. For details see the description of auto_capture_config block above.
  • inference_table_config - (Optional) Block describing the configuration of usage tracking. Consists of the following attributes:
    • enabled - boolean flag specifying if usage tracking is enabled.

email_notifications Block

  • on_update_failure - (Optional) a list of email addresses to be notified when an endpoint fails to update its configuration or state.
  • on_update_success - (Optional) a list of email addresses to be notified when an endpoint successfully updates its configuration or state.

Attribute Reference

In addition to all the arguments above, the following attributes are exported:

  • id - Equal to the name argument and used to identify the serving endpoint.
  • serving_endpoint_id - Unique identifier of the serving endpoint primarily used to set permissions and refer to this instance for other operations.

Access Control

  • databricks_permissions can control which groups or individual users can Manage, Query or View individual serving endpoints.

Timeouts

The timeouts block allows you to specify create and update timeouts. The default right now is 10 minutes for both operations.

timeouts {
  create = "30m"
}

Import

The model serving provisioned throughput resource can be imported using the name of the endpoint:

import {
  to = databricks_model_serving_provisioned_throughput.this
  id = "<model-serving-endpoint-name>"
}

Alternatively, when using Terraform version 1.4 or earlier, import using the terraform import command:

terraform import databricks_model_serving_provisioned_throughput.this <model-serving-endpoint-name>

The following resources are often used in the same context: