Redis Read/Write Separation Architecture: Star vs. Chain Replication and Its Advantages

Redis traditionally serves both reads and writes from the master node, which limits scalability because all traffic passes through a single master. To address read‑heavy scenarios, Alibaba Cloud offers a read/write separation service that adds read‑only replica nodes while keeping full compatibility with existing Redis cluster specifications.

The architecture introduces a new read‑only replica role. The redis‑proxy routes read requests to replicas and write requests to the master, while the HA component monitors node health and performs failover or replica reconstruction as needed.

Two replication topologies are described:

Star replication : each replica synchronizes directly with the master, resulting in low replication latency but increasing CPU and network load on the master as the number of replicas grows, which limits overall cluster performance.

Chain replication : replicas are arranged in a chain, so the master only syncs with the first replica; this allows virtually unlimited numbers of replicas and near‑linear performance scaling, though latency grows toward the chain’s end and a failure can affect downstream nodes.

Redis read/write separation offers several advantages:

Transparent compatibility : the service uses the same redis‑proxy as standard clusters, allowing seamless migration between regular and read/write‑separated deployments.

High availability : HA continuously monitors all DB nodes, automatically promoting a new master on failure and rebuilding failed replicas, while redis‑proxy adjusts replica weights to minimize impact.

High performance : for read‑dominant workloads, adding read‑only replicas scales read capacity without burdening the master, delivering up to 600,000 QPS and 192 MB/s per shard while preserving full command compatibility.

The article also notes that read/write separation introduces eventual consistency, as replicas may lag behind the master, and that Alibaba Cloud mitigates full‑sync occurrences with an optimized binlog replication mechanism.