Analysis
OpenAI's second Habitat engineering report examines the Azure Cosmos DB layer behind its application storage platform. The authors describe how shared physical partitions allowed heavy workloads to disrupt unrelated features, including core access flows.
Moving workloads into separate containers reduced the reach of incidents, but introduced capacity-planning overhead and launch-time scaling problems. The report describes a subsequent return to multi-tenancy by default, with data movement capability.
Our reading: isolation and efficiency require active management as a service grows. A database choice alone cannot supply that operating model. This is an engineering account of one large deployment; its architecture and reported scale should not be treated as a prescription or a performance promise for smaller systems.
Source: Rapidly scaling online storage to serve over 1 billion ChatGPT users (Part II) · 2026-10-23
Why it matters
Assess noisy neighbors, launch capacity and recovery behavior alongside ordinary request latency.