Introduces the PD scheduling component in a TiKV cluster.

The Placement Driver (PD) works as the manager in a TiKV cluster, and it also schedules Regions in the cluster. This article introduces the design and core concepts of the PD scheduling component.

Scheduling situations

In TiKV, data is organized as Regions, which are replicated on several stores. In all replicas of a Region, a leader is elected to responsible for the reading and writing traffics on the Region while followers are responsible for replicating Raft logs from the leader.

Now consider about the following situations:

  • To utilize storage space in a high-efficient way, multiple Replicas of the same Region need to be properly distributed on different nodes according to the Region size;
  • For multiple data center topologies, one data center failure only causes one replica to fail for all Regions;
  • When a new TiKV store is added, data can be rebalanced to it;
  • When a TiKV store fails, PD needs to consider:
    • Recovery time of the failed store.
      • If it’s short (for example, the service is restarted), whether scheduling is necessary or not.
      • If it’s long (for example, disk fault, data is lost, etc.), how to do scheduling.
    • Replicas of all Regions.
      • If the number of replicas is not enough for some Regions, PD needs to complete them.
      • If the number of replicas is more than expected (for example, the failed store re-joins into the cluster after recovery), PD needs to delete them.
  • Read/Write operations are performed on leaders, which can not be distributed only on a few individual stores;
  • Not all Regions are hot, so loads of all TiKV stores need to be balanced;
  • When Regions are in balancing, data transferring utilizes much network/disk traffic and CPU time, which can influence online services.

These situations can occur at the same time, which makes it harder to resolve. Also, the whole system is changing dynamically, so it is essential to collect all information about the cluster to adjust it appropriately by a scheduler. Thus, PD is introduced into the TiKV cluster.

Scheduling requirements

The above situations can be classified into two types:

  1. A distributed and highly available storage system must meet the following requirements:

    • The right number of replicas.
    • Replicas need to be distributed on different machines according to different topologies.
    • The cluster can perform automatic disaster recovery from TiKV peers’ failure.
  2. A good distributed system needs to have the following optimizations:

    • All Region leaders are distributed evenly on stores;
    • Storage capacity of all TiKV peers are balanced;
    • Hot spots are balanced;
    • Speed of load balancing for the Regions needs to be limited to ensure that online services are stable;
    • Maintainers are able to to take peers online/offline manually.

After the first type of requirements is satisfied, the system will be failure tolerable. After the second type of requirements is satisfied, resources will be utilized more efficiently and the system will have better scalability.

To achieve the goals, PD needs to collect information firstly, such as state of peers, information about Raft groups and the statistics of accessing the peers. Then we need to specify some strategies for PD, so that PD can make scheduling plans from these information and strategies. Finally, PD distributes some operators to TiKV peers to conduct scheduling plans.

Basic scheduling operators

All scheduling plans contain three basic operators:

  • Add a new replica
  • Remove a replica
  • Transfer a Region leader between replicas in a Raft group

They are implemented by the Raft commands AddReplica, RemoveReplica, and TransferLeader.

Information collection

Scheduling is based on information collection. In short, the PD scheduling component needs to know the states of all TiKV peers and all Regions. TiKV peers report the following information to PD:

  • State information reported by each TiKV peer:

    Each TiKV peer sends heartbeats to PD periodically. PD not only checks whether the store is alive, but also collects StoreState in the heartbeat message. StoreState includes:

    • Total disk space
    • Available disk space
    • The number of Regions
    • Data read/write speed
    • The number of snapshots that are sent/received (The data might be replicated between replicas through snapshots)
    • Whether the store is overloaded
    • Labels (See Perception of Topology)
  • Information reported by Region leaders:

    Each Region leader sends heartbeats to PD periodically to report RegionState, including:

    • Position of the leader itself
    • Positions of other replicas
    • The number of offline replicas
    • data read/write speed

PD collects cluster information by the two types of heartbeats and then makes decision based on it.

Besides, PD can get more information from an expanded interface to make a more precise decision. For example, if a store’s heartbeats are broken, PD can’t know whether the peer steps down temporarily or forever. It just waits for a while (by default 30min) and then treats the store as offline if there is still no heartbeat received. After that, PD balances all regions on the store to other stores.

But sometimes stores are manually set offline by a maintainer, so the maintainer can tell PD this by the PD control interface. Then PD can balance all regions immediately.

Scheduling strategies

After collecting the information, PD needs some strategies to make scheduling plans.

Strategy 1: The number of replicas of a Region needs to be correct

PD can know that the replica count of a Region is incorrect from the Region leader’s heartbeat. If it happens, PD can adjust the replica count by adding/removing replica(s). The reason for incorrect replica count can be:

  • Store failure, so some Region’s replica count is less than expected;
  • Store recovery after failure, so some Region’s replica count could be more than expected;
  • max-replicas is changed.

Strategy 2: Replicas of a Region need to be at different positions

Note that the “position” here is different from “machine”. In general, PD can only ensure that replicas of a Region are not at a same peer, thereby avoiding that the peer’s failure causes more than one replicas to become lost. However in production, you might have the following requirements:

  • Multiple TiKV peers are on one machine;
  • TiKV peers are on multiple racks, and the system is expected to be available even if a rack fails;
  • TiKV peers are in multiple data centers, and the system is expected to be available even if a data center fails;

The essence to these requirements is that peers can have the same “position”, which is the smallest unit for failure toleration. Replicas of a Region must not be in one unit. So, we can configure labels for the TiKV peers, and set location-labels on PD to specify which labels are used for marking positions.

Strategy 3: Replicas need to be balanced between stores

The size limit of a Region replica is fixed, so keeping the replicas balanced between stores is helpful for data size balance.

Strategy 4: Leaders need to be balanced between stores

Read and write operations are performed on leaders according to the Raft protocol, so that PD needs to distribute leaders into the whole cluster instead of several peers.

Strategy 5: Hot spots need to be balanced between stores

PD can detect hot spots from store heartbeats and Region heartbeats, so that PD can distribute hot spots.

Strategy 6: Storage size needs to be balanced between stores

When started up, a TiKV store reports capacity of storage, which indicates the store’s space limit. PD will consider this when scheduling.

Strategy 7: Adjust scheduling speed to stabilize online services

Scheduling utilizes CPU, memory, network and I/O traffic. Too much resource utilization will influence online services. Therefore, PD needs to limit the number of the concurrent scheduling tasks. By default this strategy is conservative, while it can be changed if quicker scheduling is required.

Scheduling implementation

PD collects cluster information from store heartbeats and Region heartbeats, and then makes scheduling plans from the information and strategies. Scheduling plans are a sequence of basic operators. Every time PD receives a Region heartbeat from a Region leader, it checks whether there is a pending operator on the Region or not. If PD needs to dispatch a new operator to a Region, it puts the operator into heartbeat responses, and monitors the operator by checking follow-up Region heartbeats.

Note that the “operators” here are only suggestions to the Region leader, which can be skipped by Regions. In other words, the Region Leader can decide whether to skip a scheduling operator or not based on its current status.