构建、启动和扩展 ChatGPT 图像

发布于 8 月 4 日

ChatGPT's Growth: It's the fastest-growing app, reaching 100M weekly active users in 12 months. New figures show accelerating growth, with 100 million new users signing up in the first week of March.
Launch Details: Higher-than-expected load from the start, going viral in India with 1 million new users per hour signing up. Initially planned to release to paying subscribers first, then to free users, but the latter was postponed due to high demand. Later, a gradual rollout to free users started.
Image Generation Work:
- How It Works: Converts description into image tokens, decoder renders tokens into an image through multiple passes, ensures integrity and safety, and finishes rendering. Users can iterate on generated images with new prompts.
- Technology Choices: Python for most code, FastAPI for building APIs, C for optimized parts, and Temporal for asynchronous workflows.
Handling Load: Rewrote image generation engine from synchronous to asynchronous to handle peak load. Traded off latency for availability. Despite unexpected high traffic, avoided hard outages and maintained availability by applying rate limits and increasing compute allocations.
Impact on Other Systems: File systems, databases, and authentication systems got overloaded. The team sped up isolation work and wrapped up non-ChatGPT endpoint isolation. Also sought quick performance wins by optimizing existing code.
Key to Success: Prioritizing reliability to ensure successful launch with higher-than-expected load.

阅读 15