2.1. Azure Cosmos DB for NoSQL

💡 First Principle: Cosmos DB doesn't sell you servers — it sells you operations, priced in Request Units. Every read, write, and query has an RU price determined by how much work it forces the engine to do. Once you internalize "RU = the currency of work," every optimization topic in this section (indexing policy, consistency level, point reads, partition scoping) becomes the same question: how do I buy the same outcome for fewer RUs?

Why care? Because RU waste is real money at scale, and Microsoft knows it — which is why nearly every Cosmos DB exam question is secretly an RU-economics question. A chatbot storing millions of conversation turns can differ in cost by 10x based purely on indexing policy and query patterns.

The mental model: treat RUs like a foreign currency with posted exchange rates. A 1 KB point read costs ~1 RU — the reference exchange rate. A write costs ~5x more (it must update indexes). A cross-partition query costs a multiple of a scoped one (it visits every partition). A strong-consistency read costs double (it must confirm with a quorum). You never ask "is this operation possible?" — you ask "what does it cost, and is there a cheaper way to buy the same result?"

⚠️ Common Misconception: "RU consumption depends only on item size." Size is just one input. Indexing (write RUs grow with every indexed path), consistency level (strong and bounded staleness reads cost roughly double), and query shape (cross-partition fan-out) all multiply the bill. The exam loves scenarios where the fix is an indexing policy or consistency change, not a smaller document.