Cloudflare extends Workers update grace period from 30s to 20 minutes for long-running AI inference
irvinebroque · x · 2026-10-13
A Cloudflare engineer detailed a change aimed at long-running agents and AI inference: the drain grace period for Workers runtime updates has been extended from 30 seconds to up to 20 minutes.
- Why: Inference requests can take minutes, and the old 30s drain window would cut off streaming responses and long requests during deploys.
- How it works: New HTTP requests go straight to the updated runtime, while previous-version processes keep handling existing work for up to 20 minutes. Streaming AI responses can finish on the old version, and existing WebSocket sessions can stay connected until the grace period ends.
- Caveats: The grace period applies per process; unfinished requests are still terminated when it ends. Developers should still design for WebSocket reconnection, safe retries, and state recovery, and Hibernation API connections may drop earlier.
A practical infra improvement for anyone building long-running agents or streaming AI products on Workers.
More from coding & agent
- OpenAI publishes full Annotations extensibility docs for WebMCP agents — OpenAIDevs · 2026-10-13
- OpenAI ships WebMCP annotations to give browser agents richer page context — OpenAIDevs · 2026-10-13
- Anthropic explains Claude Code pitfall: PreToolUse hooks returning "allow" silently bypass the safety classifier — lydiahallie · 2026-10-13
- Anthropic now lets subscription use inside third-party agents built on Claude Agent SDK — lucasmeijer · 2026-10-13
- Simon Willison on Python 3.15: lazy imports, comprehension unpacking, sampling profiler — charliermarsh · 2026-10-13
- Cursor Composer predictions hit ~60% acceptance rate, users back to tab-tab-tab flow — rudrank · 2026-10-13