PD Microservices

Starting from v8.0.0, PD supports the microservice mode, which splits the timestamp allocation and cluster scheduling functions of PD into the following two independently deployed microservices. In this way, these two functions are decoupled from the routing function of PD, which allows PD to focus on the routing service for metadata.

tso microservice: provides monotonically increasing timestamp allocation for the entire cluster.
scheduling microservice: provides scheduling functions for the entire cluster, including but not limited to load balancing, hot spot handling, replica repair, and replica placement.

Each microservice is deployed as an independent process. If you configure more than one replica for a microservice, the microservice automatically implements a primary-secondary fault-tolerant mode to ensure high availability and reliability of the service.

Warning

Currently, the PD microservices feature is experimental. It is not recommended that you use it in production environments. This feature might be changed or removed without prior notice. If you find a bug, you can report an issue on GitHub.

Usage scenarios

PD microservices are typically used to address performance bottlenecks in PD and improve PD service quality. With this feature, you can avoid the following issues:

Long-tail latency or jitter in TSO allocations due to excessive pressure in PD clusters
Service unavailability of the entire cluster due to failures in the scheduling module
Bottleneck issues solely caused by PD

In addition, when the scheduling module is changed, you can update the scheduling microservice independently without restarting PD, thus avoiding any impact on the overall service of the cluster.

Note

If the performance bottleneck of a cluster is not caused by PD, there is no need to enable microservices, because using microservices increases the number of components and raises operational costs.

Restrictions

Currently, the tso microservice does not support dynamic start and stop. After enabling or disabling the tso microservice, you need to restart the PD cluster for the changes to take effect.
Only the TiDB component supports a direct connection to the tso microservice through service discovery, while other components need to forward requests to the tso microservice through PD to obtain timestamps.
Microservices are not compatible with the Data Replication Auto Synchronous (DR Auto-Sync) feature.
Microservices are not compatible with the TiDB system variable tidb_enable_tso_follower_proxy.
Because hibernate Regions might exist in a cluster, during a primary and secondary switchover of the scheduling microservice, the scheduling function of the cluster might be unavailable for a certain period (up to peer-stale-state-check-interval, which is five minutes by default) to avoid redundant scheduling.

Usage

Currently, PD microservices can be deployed using TiDB Operator.

For detailed information on using TiDB Operator, see the following documents:

When deploying and using PD microservices, pay attention to the following:

After you enable microservices and restart PD for a cluster, PD stops allocating TSO for the cluster. Therefore, you need to deploy the tso microservice in the cluster when you enable microservices.
If the scheduling microservice is deployed in a cluster, the scheduling function of the cluster is provided by the scheduling microservice. If the scheduling microservice is not deployed, the scheduling function of the cluster is still provided by PD.
The scheduling microservice supports dynamic switching, which is enabled by default (enable-scheduling-fallback defaults to true). If the process of the scheduling microservice is terminated, PD continues to provide scheduling services for the cluster by default.
If the binary versions of the scheduling microservice and PD are different, to prevent changes in the scheduling logic, you can disable the dynamic switching function of the scheduling microservice by executing pd-ctl config set enable-scheduling-fallback false. After this function is disabled, PD will not take over the scheduling service when the process of the scheduling microservice is terminated. This means that the scheduling service of the cluster will be unavailable until the scheduling microservice is restarted.

Tool compatibility

Microservices do not affect the normal use of data import, export, and other replication tools.

FAQs

How can I determine if PD becomes a performance bottleneck?
When your cluster is in a normal state, you can check monitoring metrics in the Grafana PD panel. If the TiDB - PD server TSO handle time metric shows a notable increase in latency or the Heartbeat - TiKV side heartbeat statistics metric shows a significant number of pending items, it indicates that PD becomes a performance bottleneck.