• Store : each tikv server maps to a store
  • Pending: learner/follower log lags behind leader by a large margin

  • PD use standard revision + lease for leader failover

  • Scheduler: generate the operator, which encapsulates the steps related to rebalancing for a single region,i.e., one region has mutliple steps
    • balance-leader-scheduler
    • balance-region-scheduler
    • hot-region-scheduler

type Scheduler interface {
	GetName() string
	GetResourceKind() ResourceKind
	Schedule(cluster *clusterInfo) Operator
}

After operator is generated, it will be put in a queue and execute sequentially (possibly in parallel), by sending the operator step to the region leader as a response to the heartbeat

  • It will wait for timeout with ack

type Operator interface {
	GetRegionID() uint64
	GetResourceKind() ResourceKind
	Do(region *regionInfo) (*pdpb.RegionHeartbeatResponse, bool)
}