Question

You are designing a Cosmos DB container for an IoT telemetry workload. Devices write readings continuously, and there are only 12 device types. You initially chose deviceType as the partition key, but during load testing you observe severe throttling (429 errors) concentrated on a few logical partitions, while total provisioned throughput is far from being consumed. The application always queries telemetry filtered by a specific deviceType over a time range. Which change to the partition key strategy best resolves the hot-partition problem while keeping queries efficient?

Accepted Answer

Create a synthetic partition key by concatenating deviceType with a bucket suffix (e.g., deviceType_0 through deviceType_9) to spread writes across more logical partitions.

Answer

Switch the partition key to a random GUID generated per reading so every document lands in a unique logical partition.

Answer

Keep deviceType as the partition key but increase provisioned throughput with autoscale to eliminate the 429 errors.

Answer

Change the partition key to the reading's timestamp so writes are always distributed by time.