Redis vs Memcached: differences and which to choose

"Put Redis in front of it" has become a reflex, and interviewers sometimes ask why not Memcached. The honest answer is that for a plain look-aside cache of strings, both work, and the choice comes down to what else you need.

Memcached stores opaque values by key, across many threads, and nothing more. Redis stores typed data structures, can persist and replicate them, and runs scripts against them. That extra power is why Redis gets used for rate limiters, queues and leaderboards, not just caching.

Data model. Memcached maps a key to a blob of bytes; your application serialises and deserialises. Redis understands types, so it can increment a field inside a hash, add to a sorted set and fetch the top ten, or pop from a list, all atomically on the server without reading the value back.

Threading. Memcached is multithreaded, so a single large node scales with its cores. Redis executes commands one at a time on a main thread, which is what makes each command atomic and makes Lua scripts and MULTI/EXEC transactions simple to reason about. Since Redis 6, optional I/O threads take over socket reads and writes. To use many cores for commands, you run more Redis processes, usually as a cluster.

Persistence and replication. Memcached has neither built in; a restarted node is empty, and high availability means the client tolerates misses. Redis offers RDB snapshots, an append-only file (AOF) with a configurable fsync policy, asynchronous replicas, Sentinel for automatic failover and Redis Cluster for sharding across 16,384 hash slots. Asynchronous replication means a failover can lose the last writes acknowledged by the old primary, which matters if Redis holds anything you cannot rebuild.

Eviction. Memcached evicts least-recently-used items when memory is full, by default (the -M flag makes it return errors instead); evicting is its job. Redis lets you choose: allkeys-lru, allkeys-lfu, the volatile-* policies that only evict keys with a TTL, or noeviction, the default, which returns errors on writes when memory is full. For a cache, set an eviction policy explicitly; for anything else, monitor memory.

Licensing and Valkey. Memcached is BSD-licensed. Redis was BSD-licensed until March 2024, when Redis Ltd. moved new versions to a source-available dual licence (RSALv2 and SSPLv1). The Linux Foundation then started Valkey, a fork of Redis 7.2.4 under the BSD licence, which several cloud providers now offer. Redis 8 added the AGPLv3 as a third licence option. For most teams using a managed service this changes little, but it matters if you embed or redistribute it, so check the licence of the exact version you run.

When Memcached still wins. A very large, simple cache of rendered pages or serialised objects, where multithreading lets a few big nodes replace many Redis processes, and where losing a node only costs misses. Some of the largest published caching deployments, such as the one described in Facebook's memcache paper, are built this way.