LangGraph Streaming (Python)

Overview

An engineer ships stream_mode="values" to a token-level chat UI because it "seemed the most complete." Every single token causes the full graph state — message history, scratchpad, plan — to be re-sent and re-rendered. At ~60 tokens/sec the browser overdraws, the React reconciler can't keep up, the tab freezes, and users blame the model. The correct answer was stream_mode="messages", which emits an AIMessageChunk delta per token (typically 5-50 bytes) — one token's worth of DOM work. This is pain-catalog entry P19 and it is the #1 LangGraph integration mistake in the 1.0 generation.

Then the same UI ships to Cloud Run and hangs forever. No error. No logs. The server is emitting tokens; they just never reach the browser. Default proxy buffering (Nginx, Cloud Run's HTTP/1.1 path, Cloudflare Free) holds the last chunk waiting for more bytes. This is P46 — SSE streams from LangGraph drop the final end event over proxies that buffer — and the fix is three headers: X-Accel-Buffering: no, Cache-Control: no-cache, Connection: keep-alive.