The replicas were the load-bearing part
The Primary API service is not one backend among many. In OpenAI's own description, it provides user authentication and service-to-service authentication across the entire OpenAI stack.
That is why a database maintenance job became a partial outage across the company. The two read replicas behind that service were, in the write-up's word, load-bearing: when both failed together, the request paths that depended on them failed with them, and replica lag increased, in OpenAI's phrase, unboundedly, after the replicas went offline and out of sync.
The recovery sequence is the part worth reading closely. Engineers tried to revive either of the failed replicas and could not. They tried to bring up replacements in other regions and could not. What worked was bringing two new replicas into service in the same region, working with the service provider. The write-up gives the window as 1:48 to 5:21 PT, without an a.m. or p.m. marker; the 75% figure is a peak failure rate on paid and free ChatGPT, not a rate sustained across the window.
What OpenAI says it has changed
The write-up separates what is done from what is planned, and the desk will keep them separate. Already implemented: a moratorium on similar maintenance activities until further notice; adjusted configuration settings to reduce connection churn; a feature flag that lets OpenAI shift traffic away from replicas on a percentage rollout; and a test directing 100% of queries to the primary instance, which OpenAI says verified its hypothesis about why the primary failed.
Still to come: a pre-maintenance review process, so that any maintenance activity has a rapid-recovery plan before it runs; adequate buffer capacity kept as hot standbys during maintenance; buffer capacity in the regions where critical services are hosted; a short-term strategy and automation for recovering from replica lag, with longer-term architectural changes behind it; and a stack-wide review of redundancy and fail-safe mechanisms.
Read the second list as the gap list. The hot standbys during maintenance, the recovery plan reviewed in advance, and the replica-lag automation are all items OpenAI says it will implement after an incident in which maintenance on the authentication path took down both of its replicas at once.
The practical read
Anyone building on a single AI provider already knows the model can go down. This incident is about a duller dependency. Applications that route their own users' sign-in, billing checks or agent permissions through a provider's platform inherit that provider's maintenance calendar as well as its uptime. For about three and a half hours on Saturday, November 21, the failing component at OpenAI was not a model. It was the authentication path in front of everything.
Sources
- [1] OpenAI, post-incident write-up, "Elevated Errors on API and ChatGPT," status.openai.com, incident of November 21, 2026Read source
- [2] AI Weekly alert, "OpenAI: Postgres replica failure peaked API, ChatGPT errors at 75%", November 22, 2026 (secondary)Read source