Redis

Atomicity Traps

• GET then SET is not atomic—another client can modify between; use INCR, SETNX, or Lua
• SETNX for locks: SET lock:resource {token} NX EX 30—NX = only if not exists
• WATCH/MULTI/EXEC for optimistic locking—transaction aborts if watched key changed
• Lua scripts are atomic—use for complex operations: EVAL "script" keys args

Pub/Sub Limitations

• Messages not persisted—subscribers miss messages sent while disconnected
• At-most-once delivery—no acknowledgment, no retry
• Use Streams for reliable messaging—XREAD BLOCK + XACK pattern
• Pub/Sub across cluster: message goes to all nodes—works but adds overhead

Persistence Configuration

• RDB (snapshots): fast recovery, but data loss between snapshots—default every 5min
• AOF (append log): less data loss, slower recovery—appendfsync everysec is good balance
• Both off = pure cache—acceptable if data can be regenerated
• BGSAVE for manual snapshot—doesn't block but forks process, needs memory headroom

Memory Management (Critical)

• maxmemory must be set—without it, Redis uses all RAM, then swap = disaster
• Eviction policies: allkeys-lru for cache, volatile-lru for mixed, noeviction for persistent data
• INFO memory shows usage—monitor used_memory vs maxmemory
• Large keys hurt eviction—one 1GB key evicts poorly; prefer many small keys

Clustering

• Hash slots: keys distributed by hash—same slot required for multi-key operations
• Hash tags: {user:1}:profile and {user:1}:sessions go to same slot—use for related keys
• No cross-slot MGET/MSET—error unless all keys in same slot
• MOVED redirect: client must follow—use cluster-aware client library

Common Patterns

• Cache-aside: check Redis, miss → fetch DB → write Redis—standard caching
• Write-through: write DB + Redis together—keeps cache fresh
• Rate limiter: INCR requests:{ip}:{minute} with EXPIRE—simple fixed window
• Distributed lock: SET ... NX EX + unique token—verify token on release

Connection Management

• Connection pooling: reuse connections—creating is expensive
• Pipeline commands: send batch without waiting—reduces round trips
• QUIT on shutdown—graceful disconnect
• Sentinel or Cluster for HA—single Redis is SPOF

Common Mistakes

• No TTL on cache keys—memory grows until OOM
• Using as primary database without persistence—data loss on restart
• Blocking operations in single-threaded Redis—KEYS * blocks everything; use SCAN
• Storing large blobs—Redis is RAM; 100MB values are expensive
• Ignoring maxmemory—production Redis without limit will crash host