Overview
OceanBase AI Database adopts a storage-compute separation architecture. It provides data durability through high-availability LogService, ensuring zero data loss (RPO = 0), and guarantees data high availability through multi-replica compute nodes. This topic describes the disaster recovery deployment solutions supported by the storage-compute separation architecture. You can flexibly choose a solution based on your IDC conditions and performance/availability requirements.
Deployment solutions
You can adjust the deployment mode of a cluster by flexibly adjusting the IDC distribution across zones and the replica distribution across zones, thereby achieving different disaster recovery levels.
The following table describes the disaster recovery levels supported by OceanBase AI Database.
Deployment mode |
Number of replicas |
RTO |
RPO |
Disaster recovery scenarios |
Disaster recovery capability |
|---|---|---|---|---|---|
| Single IDC | Single replica on compute node Three LogService replicas |
minutes | 0 | For scenarios where multi-IDC deployment is not feasible. Within the same IDC, deploy replicas of the same data set on servers with equivalent disaster recovery capabilities, such as those in the same rack or powered by the same power supply. |
|
| Single IDC | Two replicas on compute nodes Three LogService replicas |
Within 8 seconds | 0 | For scenarios where multi-IDC deployment is not feasible. Within the same IDC, deploy replicas of the same data set on servers with equivalent disaster recovery capabilities, such as those in the same rack or powered by the same power supply. |
|
| Three IDCs in the same city | Single replica on compute node Three LogService replicas |
minutes | 0 | This solution is designed for scenarios where a city has three IDCs. Replicas are deployed across different IDCs, with each IDC serving as a zone. The network latency between IDCs typically ranges from 0.5 to 2 ms. |
|
| Three IDCs in the same city | Two replicas on compute nodes Three LogService replicas |
Within 8 seconds | 0 | This solution is designed for scenarios where a city has three IDCs. Replicas are deployed across different IDCs, with each IDC serving as a zone. The network latency between IDCs typically ranges from 0.5 to 2 ms. |
|
All deployment solutions rely on the multi-replica high-availability LogService that is based on the Paxos consensus protocol. The working mechanism is as follows:
Write path: The leader of a compute node directly writes logs to multiple LogService replicas through Paxos for persistence.
Read path: A follower compute node reads logs from the nearest LogService replica and locally replays hot data.
Fault tolerance: When a minority of replicas are unavailable (a three-replica LogService cluster allows one replica to be unavailable), the read and write operations of the compute node are not affected, and the compute node continues to provide services normally.
Single-IDC deployment
Single-replica compute node
Deployment topology: Single compute node replica + Three LogService replicas, all deployed in the same IDC.
Disaster recovery mechanism: When a compute node becomes unavailable, a new compute node restores data and metadata from object storage and restores incremental data from LogService to restore full service capability.
Metric |
Value |
|---|---|
| RPO | 0 (No data loss) |
| RTO | minutes |
| Disaster recovery granularity | Machine level |
Application scenarios: Business scenarios where the recovery time requirement is relatively relaxed, and simplicity of deployment, resource conservation, and better performance are prioritized.
Two-replica compute node
Deployment topology: Two compute node (OBServer) replicas + Three LogService replicas, all deployed in the same IDC.
Disaster recovery mechanism: When the server hosting the Leader fails, the system automatically performs a disaster recovery switchover, promoting a Follower to Leader to continue providing services without manual intervention.
Metric |
Value |
|---|---|
| RPO | 0 (No data loss) |
| RTO | < 8s |
| Disaster recovery granularity | Machine level |
Application scenarios: Business scenarios with high recovery time requirements, where failover needs to be completed within seconds, and better performance is prioritized.
Three-IDC deployment within the same city
Single-replica compute node
Deployment topology: Single compute node replica + Three LogService replicas, with the three replicas of the LogService deployed across IDCs.
Disaster recovery mechanism: When any IDC becomes unavailable, LogService still has a majority of replicas available, ensuring log durability and integrity. The compute node can restore full service capability from object storage and LogService.
Metric |
Value |
|---|---|
| RPO | 0 (No data loss) |
| RTO | minutes |
| Disaster recovery granularity | Data center level |
Application scenarios: Suitable for scenarios with multiple IDCs that require protection against IDC-level failures but have relatively loose recovery time requirements.
Two-replica compute node
Deployment topology: Two compute node replicas are deployed across IDCs, and three LogService replicas are deployed across IDCs.
Disaster recovery mechanism: When any IDC becomes unavailable, LogService still has a majority of replicas available. If the IDC where the compute node Leader resides becomes unavailable, the system automatically performs a disaster recovery switchover, promoting a replica from another IDC to become the Leader and continue providing service.
Metric |
Value |
|---|---|
| RPO | 0 (No data loss) |
| RTO | < 8s |
| Disaster recovery granularity | Data center level |
Application scenarios: Recommended as the primary solution for core production environments with extremely high availability and recovery time requirements.
Considerations
- Compute nodes and LogService currently only support deployment within the same city.
- Compute nodes support single- and two-replica deployments; LogService requires a three-replica deployment with at least three nodes.
- Compute nodes and LogService nodes can be deployed on the same server or on different servers independently.