
What happened
Between 1:48 and 5:21 PT on November 21, requests to OpenAI's API and ChatGPT failed at rates peaking at 75%, after maintenance on a PostgresDB service caused both read replicas of the Primary API to fail.
Why it matters
The Primary API handles user authentication and service-to-service authentication across the entire OpenAI stack, so the replica failure took down login and inter-service handshakes underpinning the rest of the product, OpenAI's writeup says.
WHO IT HITSEnterprises and developers running products on OpenAI's API — especially those depending on Chat Completions, Embeddings, Realtime or Files endpoints — saw requests fail during the window, and teams relying on ChatGPT for login-gated workflows were affected too.
Summaries like this, every morning.
The failure centered on a component OpenAI describes as load-bearing: the two Postgres read replicas behind its Primary API service, which handles user authentication and service-to-service authentication across the entire OpenAI stack. When both replicas failed at the same time during scheduled maintenance, login and inter-service handshakes underneath the rest of the product went with them — which is why the blast radius extended from Chat Completions and Embeddings to Conversations, Search and Voice mode in ChatGPT.
Reviving the existing replicas did not work, according to the writeup. The fix was bringing two new replicas in the same region into service in coordination with the provider, which OpenAI does not name. That detail points to how tightly recovery depended on the underlying database provider's involvement, though the writeup does not specify the provider.
OpenAI's stated fixes split into immediate and longer-term steps. It has placed a moratorium on similar maintenance, retuned its database to reduce connection churn, and shipped a feature flag that can shift traffic away from replicas by percentage. Longer term, it says it will add a pre-maintenance review process and hold adequate buffer capacity as hot standbys during maintenance windows. Whether those measures prevent a repeat will only show at the next maintenance window, which the writeup does not date.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
FlashLabs updated Hermes Agent, its AI agent inside the FlashRev B2B sales platform, adding multiple sessions…

Nvidia reported $115.8 billion in fiscal Q3 2027 revenue, up 103% year over year, and guided to $130 billion f…

Chip export rules are tightening, Nvidia's leadership hit a shock twist and Beijing is hinting at rare earth a…

Artificial Analysis's evaluation found OpenAI's GPT-6.2 Astra and Sol give shorter, more direct answers than p…

OrcaRouter started providing an API for Qwen4 Max, Alibaba's latest flagship model, so developers can call it…

Nvidia's leadership turmoil, customer investigations and Q3 FY2027 numbers hit at once, and the ripple now run…
