Request Units and Query Cost Optimization

What are Request Units?

Amara’s cost problem

📡 Amara at SensorFlow provisioned 100,000 RU/s for her sensor readings container. After a month, she notices some queries cost 500+ RU each — 500× more than a point read. With 500M daily events, inefficient queries are burning through her budget.

Factors that affect RU cost

Factor	Impact	Example
Document size	Larger documents cost more to read/write	1KB = ~1 RU read; 10KB = ~10 RU read
Property count	More indexed properties = higher write cost	5 indexed props vs 20 indexed props
Consistency level	Strong/Bounded = 2× read cost	Session read = 1 RU; Strong read = 2 RU
Query complexity	Cross-partition, no index, aggregations cost more	Point read = 1 RU; cross-partition scan = 500+ RU
Index utilisation	Queries using indexes cost far less	Indexed filter = 3 RU; full scan = 300 RU
Response size	More results = more RU	10 items = 5 RU; 1000 items = 50 RU

Reading RU cost from response headers

Every Cosmos DB response includes the RU charge:

// Point read — check RU cost
ItemResponse<SensorReading> response = await container.ReadItemAsync<SensorReading>(
    id: "reading-123",
    partitionKey: new PartitionKey("device-A")
);
double ruCost = response.RequestCharge;
Console.WriteLine($"Point read cost: {ruCost} RU");  // ~1 RU for 1KB

// Query — check RU cost
QueryDefinition query = new QueryDefinition(
    "SELECT * FROM c WHERE c.deviceId = @deviceId AND c.temperature > @threshold")
    .WithParameter("@deviceId", "device-A")
    .WithParameter("@threshold", 80);

FeedIterator<SensorReading> iterator = container.GetItemQueryIterator<SensorReading>(query);

double totalRU = 0;
while (iterator.HasMoreResults)
{
    FeedResponse<SensorReading> page = await iterator.ReadNextAsync();
    totalRU += page.RequestCharge;
    Console.WriteLine($"Page RU: {page.RequestCharge}, Items: {page.Count}");
}
Console.WriteLine($"Total query cost: {totalRU} RU");

💡 Exam tip: RequestCharge is per page

For queries that return multiple pages, RequestCharge is reported per page, not for the entire query. You must sum the charges across all pages to get the total cost. The exam tests this — a query returning 5 pages of results has a total cost that’s the sum of all 5 page charges.

Five optimisation techniques

1. Use point reads when possible

// Point read: ~1 RU for 1KB (cheapest possible operation)
var response = await container.ReadItemAsync<SensorReading>(
    id: "reading-123",
    partitionKey: new PartitionKey("device-A")
);

// vs. Query for the same document: ~3 RU
// SELECT * FROM c WHERE c.id = 'reading-123' AND c.deviceId = 'device-A'

Rule: If you know the id and partition key, always use a point read instead of a query.

2. Scope queries to a single partition

-- Single-partition query (cheap): targets one partition
SELECT * FROM c WHERE c.deviceId = 'device-A' AND c.temperature > 80

-- Cross-partition query (expensive): fans out to ALL partitions
SELECT * FROM c WHERE c.temperature > 80

3. Project only needed fields

-- Returns entire document (all 20+ fields)
SELECT * FROM c WHERE c.deviceId = 'device-A'

-- Returns only needed fields (smaller response = fewer RU)
SELECT c.deviceId, c.temperature, c.timestamp FROM c WHERE c.deviceId = 'device-A'

4. Optimise indexing policy

Ensure your query’s WHERE and ORDER BY fields are indexed (covered in Module 18).

5. Use pagination with MaxItemCount

QueryRequestOptions options = new QueryRequestOptions
{
    MaxItemCount = 100  // limit items per page to control memory and RU per page
};

The capacity calculator

Microsoft provides an online Cosmos DB Capacity Calculator to estimate RU/s needs:

Input	What to provide
Document size	Average size in KB
Reads/sec	Point reads + queries per second
Writes/sec	Creates + updates per second
Regions	Number of read/write regions
Indexing	Default or custom policy

💡 Exam tip: 429 response handling

When you exceed provisioned RU/s, Cosmos DB returns a 429 TooManyRequests response with a Retry-After header (in milliseconds). The official SDKs have built-in retry logic that automatically waits and retries.

Exam questions may ask: “What should you do when getting 429 errors?” — the answer depends on context:

• Transient: Let the SDK retry (built-in)
• Persistent: Increase provisioned RU/s or enable autoscale
• Query-specific: Optimise the query to reduce per-request RU cost

🎬 Video walkthrough

🎬 Video coming soon

Request Units and Query Cost — DP-420 Module 19

Request Units and Query Cost — DP-420 Module 19

~16 min

Flashcards

Question

How much does a point read of a 1KB document cost?

Click or press Enter to reveal answer

Answer

~1 RU. A point read (using id + partition key) is the cheapest possible operation. It's the baseline — 1 RU for 1KB. Larger documents cost proportionally more.

Click to flip back

💡 Show hint

Question

How do you get the RU cost of a Cosmos DB operation?

Click or press Enter to reveal answer

Answer

From the response's RequestCharge property (response.RequestCharge in C#) or the x-ms-request-charge HTTP header. For multi-page queries, sum the RequestCharge across all pages.

Click to flip back

💡 Show hint

Question

Why do Strong and Bounded Staleness consistency reads cost 2× RU?

Click or press Enter to reveal answer

Answer

They require quorum confirmation — the read must be verified against multiple replicas before returning. Session, Consistent Prefix, and Eventual reads go to a single replica, costing 1× RU.

Click to flip back

💡 Show hint

Question

What is the cheapest way to read a single document if you know its id and partition key?

Click or press Enter to reveal answer

Answer

A point read (ReadItemAsync). It costs ~1 RU for 1KB — cheaper than any SQL query for the same document. Always prefer point reads over queries when you have both the id and partition key.

Click to flip back

💡 Show hint

Knowledge Check

Knowledge Check

Amara notices a query costing 500 RU per execution. The query is: SELECT * FROM c WHERE c.temperature > 80. What's the most likely cause of the high cost?

AThe temperature field is not indexedBThe query is cross-partition — it scans all partitions because there's no partition key filterCThe container is using serverless modeDThe query returns too few results

Check Answer

Knowledge Check

A developer reads a document using: SELECT * FROM c WHERE c.id = 'reading-123' AND c.deviceId = 'device-A'. The query costs 3.5 RU. How can they reduce the cost?

AUse a point read (ReadItemAsync) instead of a query — it costs ~1 RUBAdd a composite index on id and deviceIdCChange the consistency level to Strong for faster readsDIncrease the container's provisioned RU/s

Check Answer

Knowledge Check

Amara's query returns results across 5 pages. Each page reports RequestCharge of 20 RU. What is the total RU cost?

A20 RU — the charge is for the entire queryB100 RU — sum of all page chargesC4 RU — 20 divided by 5 pagesDCannot determine without knowing document sizes

Check Answer

---

Next up: Integrated Cache and Dedicated Gateway — how to add a server-side cache layer that reduces RU consumption for read-heavy workloads.