ความทนทานของการเชื่อมต่อฝั่งไคลเอ็นต์

You configured replication, Sentinel, and Cluster on the server. But high availability only works if the client reacts correctly to failovers, moved slots, and dropped connections. This lesson covers client-side resilience.

Opening a TCP connection per command is slow. A connection pool reuses a set of connections across requests, bounded by a maximum size to protect the server.

Without timeouts, a stalled node can hang your whole app. Set both a connect timeout and a socket/command timeout so failed nodes fail fast.

Transient errors (a brief failover) should be retried, ideally with exponential backoff to avoid hammering a recovering node.

With Sentinel, the client asks Sentinel for the current master address rather than hardcoding it. After a failover, the client re-queries and reconnects to the new master.

For read-heavy workloads, route reads to replicas to offload the master. Be aware replicas may be slightly behind (eventual consistency).

In a cluster, a key may live on a different node. The server replies MOVED with the correct node. A cluster-aware client follows the redirect and updates its slot map.

During slot migration the server may reply ASK, a one-time redirect. The client should send ASKING then the command to the target node, without permanently updating its slot map.

Cluster-aware clients periodically refresh their slot-to-node map and on receiving redirects, so they keep routing to the right node as the cluster reshards.

Retrying writes is risky if the first attempt actually succeeded. Prefer idempotent operations (SET, INCR with a dedup key) so a retry cannot double-apply an effect.

Resilient clients combine pools, timeouts, backoff retries, Sentinel/cluster awareness, replica reads where safe, and idempotent writes. Together they turn server-side HA into end-to-end availability.

You learned client-side resilience: connection pools, connect and command timeouts, exponential backoff retries, Sentinel-aware master discovery, replica reads, and handling MOVED/ASK redirections in a cluster. Combine these with idempotent writes for true end-to-end availability.