Azure Cosmos DB RU cost optimization: stop over-provisioning throughput

Quick answer

Cosmos DB bills mostly by provisioned throughput — Request Units per second at ~$0.008 per 100 RU/s-hour, with a 400 RU/s floor (~$23/mo) — plus ~$0.25/GB storage, multiplied by region count. The overspend is almost always over-provisioned RU/s or unnecessary multi-region writes. Right-size RU/s, use autoscale for variable load and serverless for dev, limit regions, and reserve committed throughput (up to ~65% off).

Cosmos DB is fast and globally distributed, and it'll bill you for both whether you need them or not. The cost model is unusual — you don't pay per query or per CPU, you pay for provisioned throughput capacity measured in Request Units. Get the RU/s sizing and the region count right and Cosmos is reasonable; get them wrong and it's one of the priciest lines on an Azure bill.

The RU model

A Request Unit normalizes the cost of an operation: a 1 KB point read is 1 RU, a write is ~5 RU, and a complex query can be hundreds. You provision a number of RU/s; exceed it and requests get throttled with 429 errors. The pricing:

100 RU/s        ≈ $0.008/hour
400 RU/s (min)  ≈ $23/month
10,000 RU/s     ≈ $584/month
Storage         ≈ $0.25/GB-month
× number of regions (and × writes if multi-region write)

Pick the right throughput model

Provisioned (manual): cheapest for steady, predictable load. You set a fixed RU/s and pay for it continuously.
Autoscale: bills for the peak RU/s used within a 10x band (e.g. 1,000-10,000). Avoids throttling on variable load without provisioning for peak 24/7. Costs a bit more per RU than manual.
Serverless: bills per request (~$0.25 per million RU). Cheapest for spiky, low-volume, and dev workloads; not for sustained high throughput.

The two big overspends

Over-provisioned RU/s. Provisioning 10,000 RU/s when you use 1,000 wastes ~$500/month. Measure actual RU consumption and size to it with headroom; use autoscale if you can't predict the peak.
Unneeded multi-region writes. Each additional region multiplies throughput cost, and enabling multi-region writes multiplies it again. Multi-region is right for genuine global low-latency writes — but many deployments enable it by default and pay 2-3x for capacity they don't need.

Other levers

Reserved capacity: committing to throughput for 1-3 years is up to ~65% off — the biggest discount available for steady workloads.
TTL on containers: expire old documents automatically to keep storage (and per-region storage cost) down.
Partition key design: good partitioning means queries touch fewer partitions and cost fewer RUs — directly lowering the RU/s you need to provision.

FAQ

How is Azure Cosmos DB priced?

Primarily by provisioned throughput in Request Units per second (RU/s) at about $0.008 per 100 RU/s per hour, plus storage at ~$0.25/GB-month. The minimum provisioned throughput is 400 RU/s (~$23/month). Multi-region and multi-write multiply the throughput cost per region.

What is an RU in Cosmos DB?

A Request Unit is Cosmos DB's normalized cost of a database operation. A simple 1 KB point read is 1 RU; queries, writes, and larger items cost more. You provision RU/s capacity (or use serverless/autoscale), and throttling (429 errors) happens when demand exceeds provisioned RU/s.

Should I use provisioned, autoscale, or serverless throughput?

Provisioned (manual) is cheapest for steady, predictable load. Autoscale bills for the peak RU/s used within a 10x range and suits variable load without throttling. Serverless bills per request (~$0.25 per million RU) and is cheapest for spiky, low-volume, or dev workloads. Match the model to your traffic shape.

Why is my Cosmos DB bill so high?

Usually over-provisioned RU/s, multiple regions with multi-region writes, or autoscale set to a high maximum. Provisioning 10,000 RU/s 'to be safe' when you use 1,000 wastes ~$500/month. Multi-region write doubles or triples throughput cost. Right-sizing RU/s and limiting regions are the biggest levers.

How do I reduce Cosmos DB costs?

Right-size provisioned RU/s to real usage with headroom, use autoscale for variable load and serverless for dev, limit to the regions you actually need, set a sensible TTL to keep storage down, and design partition keys well so queries cost fewer RUs. Reserved capacity gives up to ~65% off committed throughput.

How does C3X estimate Cosmos DB cost?

C3X prices an azurerm_cosmosdb_account from its provisioned throughput and region configuration, so over-provisioned RU/s and multi-region multipliers show up as a concrete monthly number before deployment.

What to do next

The RU/s number and region count are set in your Terraform, which means you can catch over-provisioning before it ships. C3X prices an azurerm_cosmosdb_account from its provisioned throughput and regions, so a 10,000 RU/s default or an extra write region shows up as real dollars in review. The quickstart runs it on your stack in minutes.