Follower Read

When a read hotspot appears in a Region, the Region leader can become a read bottleneck for the entire system. In this situation, enabling the Follower Read feature can significantly reduce the load of the leader, and improve the throughput of the whole system by balancing the load among multiple followers. This document introduces the use and implementation mechanism of Follower Read.

Overview

The Follower Read feature refers to using any follower replica of a Region to serve a read request under the premise of strongly consistent reads. This feature improves the throughput of the TiDB cluster and reduces the load of the leader. It contains a series of load balancing mechanisms that offload TiKV read loads from the leader replica to the follower replica in a Region. TiKV's Follower Read implementation provides users with strongly consistent reads.

Usage

To enable TiDB's Follower Read feature, modify the value of the tidb_replica_read variable as follows:

set [session | global] tidb_replica_read = '<target value>';

Scope: SESSION | GLOBAL

Default: leader

This variable is used to set the expected data read mode.

  • When the value of tidb_replica_read is set to leader or an empty string, TiDB maintains its default behavior and sends all read operations to the leader replica to perform.

  • When the value of tidb_replica_read is set to follower, TiDB selects a follower replica of the Region to perform all read operations.

  • When the value of tidb_replica_read is set to leader-and-follower, TiDB can select any replicas to perform read operations. In this mode, read requests are load balanced between the leader and follower.

  • When the value of tidb_replica_read is set to prefer-leader, TiDB prefers to select the leader replica to perform read operations. If the leader replica is obviously slow in processing read operations (such as caused by disk or network performance jitter), TiDB will select other available follower replicas to perform read operations.

  • When the value of tidb_replica_read is set to closest-replicas, TiDB prefers to select a replica in the same availability zone to perform read operations, which can be a leader or a follower. If there is no replica in the same availability zone, TiDB reads from the leader replica.

  • When the value of tidb_replica_read is set to closest-adaptive:

    • If the estimated result of a read request is greater than or equal to the value of tidb_adaptive_closest_read_threshold, TiDB prefers to select a replica in the same availability zone for read operations. To avoid unbalanced distribution of read traffic across availability zones, TiDB dynamically detects the distribution of availability zones for all online TiDB and TiKV nodes. In each availability zone, the number of TiDB nodes whose closest-adaptive configuration takes effect is limited, which is always the same as the number of TiDB nodes in the availability zone with the fewest TiDB nodes, and the other TiDB nodes automatically read from the leader replica. For example, if TiDB nodes are distributed across 3 availability zones (A, B, and C), where A and B each contains 3 TiDB nodes and C contains only 2 TiDB nodes, the number of TiDB nodes whose closest-adaptive configuration takes effect in each availability zone is 2, and the other TiDB node in each of the A and B availability zones automatically selects the leader replica for read operations.
    • If the estimated result of a read request is less than the value of tidb_adaptive_closest_read_threshold, TiDB can only select the leader replica for read operations.
  • When the value of tidb_replica_read is set to learner, TiDB reads data from the learner replica. If there is no learner replica in the Region, TiDB returns an error.

Implementation mechanism

Before the Follower Read feature was introduced, TiDB applied the strong leader principle and submitted all read and write requests to the leader node of a Region to handle. Although TiKV can distribute Regions evenly on multiple physical nodes, for each Region, only the leader can provide external services. The other followers can do nothing to handle read requests but receive the data replicated from the leader at all times and prepare for voting to elect a leader in case of a failover.

To allow data reading in the follower node without violating linearizability or affecting Snapshot Isolation in TiDB, the follower node needs to use ReadIndex of the Raft protocol to ensure that the read request can read the latest data that has been committed on the leader. At the TiDB level, the Follower Read feature simply needs to send the read request of a Region to a follower replica based on the load balancing policy.

Strongly consistent reads

When the follower node processes a read request, it first uses ReadIndex of the Raft protocol to interact with the leader of the Region, to obtain the latest commit index of the current Raft group. After the latest commit index of the leader is applied locally to the follower, the processing of a read request starts.

Follower replica selection strategy

Because the Follower Read feature does not affect TiDB's Snapshot Isolation transaction isolation level, TiDB adopts the round-robin strategy to select the follower replica. Currently, for the coprocessor requests, the granularity of the Follower Read load balancing policy is at the connection level. For a TiDB client connected to a specific Region, the selected follower is fixed, and is switched only when it fails or the scheduling policy is adjusted.

However, for the non-coprocessor requests, such as a point query, the granularity of the Follower Read load balancing policy is at the transaction level. For a TiDB transaction on a specific Region, the selected follower is fixed, and is switched only when it fails or the scheduling policy is adjusted. If a transaction contains both point queries and coprocessor requests, the two types of requests are scheduled for reading separately according to the preceding scheduling policy. In this case, even if a coprocessor request and a point query are for the same Region, TiDB processes them as independent events.