Activate operator service
An operator service is the pre-built multimodal AI inference service of AnalyticDB for MySQL. Core operators—such as object detection and face recognition—are already hosted by the platform, so you do not have to procure GPU cards, build a Ray cluster, or maintain an inference framework yourself. A single create action gives you a highly available, autoscaling inference endpoint.
Prerequisites
Make sure all three conditions are met before you create an operator service:
-
An AnalyticDB for MySQL Enterprise Edition cluster is purchased and in the Running state. See Create a cluster.
-
The operator service is activated. Billing is based on actual token consumption or audio-duration usage. If you skip this step, creating an operator service will fail. Go to Activate operator service.
-
Your account has administrator permission on the AnalyticDB for MySQL cluster, or has been granted the operator service permissions.
An operator service attaches to your VPC through an elastic network interface (ENI). ENI attachment is enabled by default and cannot be turned off. Make sure your VPC has enough available IP addresses.
Create an operator service
A running operator service is not editable. To change the service type, specifications, or acceleration settings, you must delete the service and create a new one. Complete the planning below before you create.
Configuration planning
Service type
Only Multimodal Service is supported at present. A single multimodal service can optionally enable object detection acceleration and face recognition acceleration.
Ray worker configuration
Worker nodes run the actual inference workload. Plan the following fields:
|
Field |
Planning guidance |
|
Worker spec |
Compute spec of a single worker node. Default: medium (2 ACU). Pick a larger spec for heavy or complex inference workloads. |
|
Worker disk size (GB) |
Default 70 GB, used for model caching and intermediate data. Increase for very large inference workloads. |
|
Minimum count |
Floor of the worker fleet. Determines cold-start latency and minimum cost. Set to 2 or more for high-availability production workloads. |
|
Maximum count |
Ceiling of the worker fleet. Determines peak throughput. The system autoscales by load; setting this equal to the minimum leaves no room to grow. |
Acceleration capabilities (enable as needed)
-
Object Detection Acceleration—enable it when your business calls object-recognition operators.
-
Face Recognition Acceleration—enable it when your business calls face-recognition operators.
The two accelerations can be enabled independently or together. For each enabled acceleration, choose either GPU or CPU acceleration:
|
Dimension |
GPU Acceleration |
CPU Acceleration |
|
Best for |
Online inference and real-time detection that demand low latency and high throughput. |
Offline batch processing and low-QPS validation. |
|
Key configuration |
GPU model (currently |
Number of resources. The resource spec is automatically aligned with the Ray worker spec and is not configured separately. |
|
Cost |
Billed per GPU card—higher. |
Billed per ACU—lower. |
Network configuration
The elastic network interface (ENI) is enabled by default and cannot be turned off. The service attaches to the VPC of your cluster. Confirm in advance that the VPC has enough available IP addresses.
Platform-managed fields (no input needed)
-
Ray Resource Group Name: auto-generated, in the format
resource{random}. -
Ray Head Configuration: fixed at medium (2 ACU) with a 70 GB disk.
Procedure
-
Log on to the AnalyticDB for MySQL console and click the target cluster ID in the cluster list.
-
In the left-side navigation pane of the cluster, expand AI Applications and click Operator Service.
-
In the upper-right corner, click Create Operator Service.
-
In the create panel, fill in each field according to your earlier planning:
-
Service Name: used to identify the service. Encode the business purpose, for example
mm-doorface-prod. -
Service Type: Multimodal Service.
-
Ray Worker configuration: set spec, disk size, minimum count, and maximum count.
-
Object Detection Acceleration and Face Recognition Acceleration: enable as needed and choose GPU Acceleration or CPU Acceleration.
-
Network Configuration: confirm the ENI switch is on (default).
-
-
Click OK. The service enters the Creating state, and typically becomes Running in 5 to 10 minutes.
-
After the service is Running, copy the VPC endpoint from the Service Endpoint column and call the service from your application. See Call the operator service for details.
Call the operator service
After the service enters the Running state, get one of the two endpoints from the Service Endpoint column or the detail panel:
|
Endpoint type |
Use when |
|
VPC endpoint |
Your application is in the same VPC. Preferred—lower latency and traffic stays off the public Internet. |
|
Public endpoint |
Your application is on the public Internet, or you need a quick smoke test. Apply your own access controls and source-IP restrictions. |
For protocol details, parameters, and sample code, choose the use-case guide that matches your scenario:
-
Voiceprint retrieval: voiceprint library management, 1:1 voiceprint matching, and 1:N voiceprint search for speaker identification, voiceprint authentication, and service quality inspection.
-
Image retrieval: image search, object detection, and face matching.
-
Video key-event extraction and retrieval: extract and search key events inside videos.
-
PDF intelligent parsing: extract structured data such as text and tables from PDF documents.
A single operator service endpoint can serve several of the scenarios above. You do not need a separate operator service per scenario.
Manage the service lifecycle
States
An operator service moves through four observable states:
|
State |
Meaning |
What to do |
|
Creating |
The underlying AI resource group is being brought up. Usually takes 5 to 10 minutes. |
Wait. The endpoint cannot accept calls yet. |
|
Running |
The service is healthy and serving inference traffic. |
Copy the endpoint and start calling it. |
|
Creation Failed |
Resource provisioning failed. The detail panel does not expose the failure reason today. |
Delete and recreate the service. If it keeps failing, see the FAQ. |
|
Deleting |
The platform is releasing the underlying AI resource group. |
Wait until the service disappears from the list. The service name is reserved until then. |
Find an operator service
Combine the following three filters to narrow the list:
|
Filter |
How it works |
|
Service Name |
Fuzzy match on the service name. |
|
Service Type |
For example, Multimodal Service. |
|
Service Status |
Running, Creating, and so on. |
View service details
In the Actions column, click Details. A side panel opens with:
-
Basic Information: service name, type, status, creation time, last-updated time.
-
Service endpoint: public and VPC endpoints (populated only after the service is Running).
-
Resource Configuration: Ray resource group name, Ray head config, Ray worker config (spec, disk, min and max count), and acceleration config (shown only when enabled).
-
Network Configuration: ENI attachment status.
About the AI resource group
Each operator service automatically provisions a dedicated AI resource group (type AI RayCluster) that hosts the Ray head and workers.
-
You can see this resource group on the Cluster Management > Resource Management > Resource Groups tab. Its name matches the Ray Resource Group Name in the service detail panel.
-
The group is tightly bound to one operator service. Do not modify or delete it manually—doing so breaks the service.
-
When you delete the operator service, its resource group is released automatically. No manual cleanup is needed.
Delete an operator service
Deletion is irreversible. The service stops immediately and callers lose access to its endpoints. Make sure that dependent applications have been cut over or shut down first.
-
In the Actions column, click Delete.
-
In the Delete operator service service-name? confirmation dialog, double-check the name and click OK.
-
The service enters the Deleting state and disappears from the list after the resource group is released (typically 1 to 2 minutes).
FAQ
The service stays in Creating too long, or moves to Creation Failed
The three most common causes:
-
The operator service is not activated. Go to Activate operator service and complete the order (billed by actual usage afterwards). Recreate the operator service after activation.
-
VPC ENI quota or available IPs are exhausted. Check the VPC CIDR for free IPs, and switch zones or expand the VPC if needed.
-
The cluster is not in the Running state. In the cluster list, make sure the cluster shows Running (not Creating Network, Upgrading, and so on) before you try again.
After fixing the root cause, delete the failed service and recreate it. Creation Failed does not recover on its own.
The endpoint is unreachable
Walk through these checks in order:
-
Confirm that the service status is Running (Creating cannot serve traffic).
-
Verify that the caller and endpoint type match: a VPC endpoint accepts traffic only from inside the same VPC. Use the public endpoint for callers on the public Internet.
-
For a public-endpoint call, confirm that the caller host's outbound security group permits the target port.
-
Re-read the use-case guide (Image retrieval or Video key-event retrieval) and confirm that you are passing the right path and parameters.
Workers do not scale out as expected
-
Scale-out triggers depend on current load and worker resource utilization. Low-QPS workloads may stay at the minimum count.
-
Verify that the maximum count is greater than the minimum count—equal values leave no room to grow.
-
To change the minimum or maximum, delete the service and create a new one.
The resource group is still visible after deletion
Release takes 1 to 2 minutes. If it is still there after 5 minutes, refresh Cluster Management > Resource Management > Resource Groups. If the group does not go away, submit a ticket.