SYMPTOM
Cosmos DB 429: request rate too large, RU/s exhausted

CAUSE
The container's provisioned RU/s is below what the workload asks for. 429 with a RetryAfterMs header is throttling, not an outage; the fix is query efficiency or more throughput, not retries alone.

CONFIRM IT
CONFIRM: portal metrics show normalized RU consumption at 100% on the throttled partition; SDK response headers carry the request charge and retry-after. Check whether one hot partition key drives it (uneven partition distribution).

FIX
FIX: short term: honor retry-after with backoff (the SDK does this by default - do not disable it). Real fix: add a partition-key filter to cross-partition queries, or raise RU/s (`az cosmosdb sql container throughput update`), or enable autoscale. For a hot partition, the partition key design is wrong and needs a migration.

VERIFY
VERIFY: normalized RU consumption drops below 70% at peak and 429s stop in metrics.