2.3. Azure Managed Redis in AI Solutions

💡 First Principle: The cheapest, fastest query is the one you never execute. Redis exists to intercept repeated work — repeated reads, repeated similarity searches, even repeated model calls — and answer them from memory in under a millisecond. In an AI system where a single model call costs seconds and cents, interception compounds into dramatic savings.

Why care? Because AI workloads are pathologically repetitive: the same FAQ questions, the same session lookups, the same embeddings recomputed. Azure Managed Redis — the current managed offering built on Redis Enterprise — is where that repetition goes to die. The exam tests two roles: classic cache (2.3.1) and vector-capable semantic store (2.3.2).

The mental model: Redis is your system's short-term memory. Fast, small, and forgetful by design — entries expire, and under memory pressure the eviction policy discards the least valuable. Long-term memory is Cosmos DB or PostgreSQL; Redis holds what's hot right now.

⚠️ Common Misconception: "Redis stores data, so it can be the system of record." Data in Redis can vanish legitimately — TTL expiry, eviction under pressure. Everything in Redis must be rebuildable from a durable store. Scenarios proposing Redis as the sole store of conversations or embeddings are describing data loss with extra steps.