Scale Replica
Zilliz Cloud supports cluster-level replication. Each replica is an exact copy of the resources and data in a cluster. Using replicas can increase query throughput and availability.
For users experiencing QPS bottlenecks, adding replicas can distribute the query workload, enhancing overall query throughput. To proactively optimize performance, you can monitor Query CU Computation on the metrics page to determine when replica scaling is needed.
Note that adding replicas will not increase the cluster capacity because the capacity is only determined by the number of query CUs of each cluster. If you want to increase the cluster capacity, please refer to Scale Cluster.
This guide outlines the procedures of configuring replicas for a cluster in Zilliz Cloud.
Limits
You can configure replicas for an existing Dedicated cluster as long as the following conditions are met:
-
The cluster has 8 query CUs or more
-
The product of the cluster query CU count x replica count should not exceed 256.
Updating the replica configurations may lead to slight service jitter. Please exercise caution.
Manual scaling
You can adjust the number of replicas for an existing Dedicated cluster either manually on the console or programmatically.
The following demo shows how to configure replicas on the Zilliz Cloud web console.
When clicking Save in the Scale Cluster Replicas dialog box, you will be prompted to check the resource quota for your project. If the resources are sufficient, the dialog box will disappear after the check is complete, otherwise, you can
Click Go To Project Resource Settings to edit resource settings for the project, or
Click Back to Last Step to change your cluster settings.
During the process, some additional resources will be required for rolling; these resources will be released after use.
You can also use the RESTful API to manually adjust the number of replicas in a cluster. For details, refer to Modify Cluster Replica.
export TOKEN="YOUR_API_KEY"
export CLUSTER_ID="inxx-xxxxxxxxxxxxxxx"
curl --request POST \
--url "${BASE_URL}/v2/clusters/${CLUSTER_ID}/modify" \
--header "Authorization: Bearer ${TOKEN}" \
--header "Accept: application/json" \
--header "Content-Type: application/json" \
-d '{
"replica": 2
}'
Scheduled scaling
You can configure scaling of replicas based on a predefined time schedule via the Zilliz Cloud web console or RESTful API.
The interval between schedules should be greater than 30 minutes.
For details about how to use the advanced mode to write cron expressions, see Cron Expression.
The following demo shows how to enable replica auto-scaling.
You can also use the RESTful API to configure replica scheduled scaling. For details, refer to Modify Cluster.
export TOKEN="YOUR_API_KEY"
export CLUSTER_ID="inxx-xxxxxxxxxxxxxxx"
curl --request POST \
--url "${BASE_URL}/v2/clusters/${CLUSTER_ID}/modify" \
--header "Authorization: Bearer ${TOKEN}" \
--header "Accept: application/json" \
--header "Content-Type: application/json" \
-d '{
"autoscaling": {
"replica": {
"schedules": [
{
"cron": "10 0 0 0 0 ?",
"target": 2
}
]
}
}
}'
Dynamic scaling
Zilliz Cloud supports dynamic scaling of replicas to help you maintain performance while eliminating manual intervention. When enabled, the system automatically adjusts the replica count based on the real-time CU computation metric, ensuring your workload is served efficiently without service disruption.
When setting up dynamic scaling, you can configure the following bounds:
-
Minimum Replica: Defaults to the current count.
-
Maximum Replica: Defaults to 1× the current CU size. The maximum replica cannot exceed 10. Contact support if you need this limit increased.
Selecting a maximum replica below the current value triggers an immediate scale-in.
Selecting a minimum replica above the current value triggers an immediate scale-out.
Trigger conditions
-
Scale Out: Triggered when CU computation exceeds 60% for 2 minutes.
-
Scale In: Triggered when CU computation stays below 40% for 10 minutes.
Scaling size calculation
The following formula explains how Zilliz Cloud calculates the target replica count for a dynamic scaling event. The dynamic scaling formula aims to maintain your CU computation at a target value of 50%.
Target Replica Count = Current Replica Count × (Current Metric Value / Target Metric Value)
Variable Name | Description |
|---|---|
Target Replica Count | The new replica count that the system aims to scale to. |
Current Replica Count | The current replica count of the cluster. |
Current Metric Value | The current measured value of the CU computation metric. |
Target Metric Value | Expected CU compuation value after scaling, which is 50%. |
For example, if replica dynamic scaling is enabled and the following conditions are met:
-
Current Replica Count: 1
-
Cluster CU Computation: Above 60% for 10 minutes
A dynamic scaling event will be triggered. The target query CU number is calculated as:
1 × (60 / 50) = 1.2
This value is then rounded up to 2, resulting in a new replica count of 2.
Procedures
The following demo shows how to configure dynamic auto-scaling on the Zilliz Cloud web console.
In addition, you can configure dynamic scaling using RESTful API. For details, refer to Modify Cluster.
export TOKEN="YOUR_API_KEY"
export CLUSTER_ID="inxx-xxxxxxxxxxxxxxx"
curl --request POST \
--url "${BASE_URL}/v2/clusters/${CLUSTER_ID}/modify" \
--header "Authorization: Bearer ${TOKEN}" \
--header "Accept: application/json" \
--header "Content-Type: application/json" \
-d '{
"autoscaling": {
"replica": {
"min": 1,
"max": 2
}
}
}'
View scaling progress
Once a manual scaling request is sent or a scheduled or dynamic scaling event is triggered, a job record will be generated. You can check the progress on the Jobs page.
When a scaling job is in progress, you cluster status will change to "Modifying". Once the scaling job is successful, the cluster status will change to "Running".