OpenAI Outage Hits 75 Percent Failure Rate Over Postgres Maintenance

A routine database update broke ChatGPT and the OpenAI API for 3.5 hours. Your app went down because their read replicas failed.

Tools ยท Source: OpenAI Status

What happened

On November 21, ChatGPT and the OpenAI API suffered a massive partial outage. Between 1:48 and 5:21 PT, request failures peaked at 75 percent. This massive drop hit both free and paid ChatGPT tiers. The outage lasted roughly three and a half hours. It swept across almost every single API endpoint. Chat Completions, Fine-tuning, Embeddings, Images, Batch, Audio, Moderations, Realtime, and Files all went down. On the consumer side, Conversations, Login, Search, File uploads, and Voice mode stopped working.

The root cause was not a complex AI failure. It was routine scheduled maintenance on a Postgres database. This specific database powers the Primary API service for OpenAI. That service is critical because it handles user authentication and service-to-service authentication across the entire OpenAI stack. During the maintenance window, both read replicas for this database failed at the exact same time. These replicas were heavily load-bearing.

When the replicas went offline, they fell out of sync. Replica lag increased without bound. Every single request path depending on that database failed immediately. The login and inter-service handshakes underneath the rest of the product simply vanished. OpenAI engineers made multiple unsuccessful attempts to revive the failed replicas. They also tried to bring up replacement replicas in other regions, but those attempts failed too. Ultimately, they mitigated the disaster by working directly with their unnamed service provider. Together, they brought two brand new replicas into service in the exact same region.

Key facts

Why it matters

This incident exposes a massive single point of failure for anyone building products on top of OpenAI. A single Postgres maintenance window took their cross-stack authentication offline for three and a half hours. If your AI product depends on the Primary API, your uptime is entirely at the mercy of their database architecture. You cannot engineer around their login portal going down. When their read replicas fail, your customers get error messages. You pay the price in lost trust and revenue while waiting for another company to fix their servers.

The second-order effect is a forced shift in how founders view reliance on a single model provider. OpenAI is now scrambling to fix their processes. They placed a moratorium on similar maintenance activities. They adjusted configuration settings to reduce connection churn. They also shipped a feature flag to shift traffic away from replicas on a percentage basis. Longer term, they promise to establish pre-maintenance review processes and ensure adequate buffer capacity as hot standbys. But the damage is already documented. Builders now know that even the most heavily funded AI company in the world can be taken down by basic database replica lag.

For builders

Your uptime depends on their database architecture

The Primary API service handles all authentication across the OpenAI stack. When their load-bearing Postgres replicas fail, your API calls drop immediately. You lose users and money when their infrastructure breaks.

Check the status page before debugging code

Do not waste time checking your own code or wifi connection when requests fail. Go straight to status.openai.com to see the real issue. The problem is likely a backend failure on their end.

Plan for multi-region failovers to fail

OpenAI tried to bring up replacement replicas in other regions during the outage. Those attempts did not work. You must assume that their regional failover mechanisms might not save your application during a hard crash.

My take

OpenAI builds the most advanced artificial intelligence in the world, but a basic Postgres maintenance task took them offline. It is a harsh reminder that traditional infrastructure matters just as much as model weights. I always tell founders to build fallbacks into their products. Even the biggest tech giants trip over database replicas, and you do not want your startup to die because someone else botched a routine update.

Original reporting: OpenAI Status. This is my rewrite and opinion.

More AI news for builders