Memcached

Brad Fitzpatrick wrote Memcached in 2003 to rescue LiveJournal from database overload, removing persistence and querying to create an in-memory key-value cache. That design means modern systems treat the daemon as a disposable buffer where cache misses are normal operational behavior. Written in C with an event-driven core, it serves high-throughput read paths with predictable, sub-millisecond response times.

Why Memcached instead of Memcached Cloud

Memcached Cloud runs the Memcached protocol on managed Redis Enterprise infrastructure, adding multi-zone replication, automated failover, and cloud consoles at a commercial markup. For engineering teams integrating caching into existing deployment pipelines, self-hosted Memcached avoids external vendor lock-in, account overhead, and cross-VPC networking latency. Running instances directly on Kubernetes worker nodes or internal virtual machines keeps cache lookups inside the local private network.

The trade-off is operational ownership versus infrastructure complexity. Memcached Cloud treats cache nodes like stateful databases that require active synchronization and failover management. In contrast, native Memcached relies on client-side consistent hashing: when a node restarts, clients route around it and repopulate keys from primary databases. Self-hosting requires provisioning instances and setting memory caps, but it avoids treating ephemeral cache data like an expensive distributed database.

What it does well

• Multithreaded execution: Distributes connections, command parsing, and lookups across worker threads using libevent, saturating multi-core CPUs from a single daemon.
• Slab memory allocation: Pre-allocates memory in fixed slab classes, preventing system memory fragmentation and allocator pauses during high write churn.
• Built-in programmable proxy: Routes requests, pools connections, and balances traffic across cache clusters using embedded Lua scripts without separate proxy services.
• Flash tiering with Extstore: Offloads large values to NVMe or SSD storage while retaining keys in RAM, reducing memory costs for large datasets.
• Shared-memory restarts: Keeps cached data intact in shared memory across daemon restarts, preventing database stampedes during routine updates.
• Extended meta protocol: Supports pipelining, return-value suppression, and optimistic locking via CAS tokens in single round-trips to conserve network bandwidth.