TimeoutError: signal timed out when using the map-reduce paradigm via the Send API even for low parallelization (k=20)

[wallacelim]

When trying to implement the map-reduce pattern (with close reference to this guide), I found that increasing the parallelization factor (i.e., number of concurrent Send calls) to ~20 will seemingly always throw a TimeoutError: signal timed out.

To reproduce, simply run the example code in the guide linked above, except with the following line changed:

subjects_prompt = """Generate a comma separated list of **20** examples related to: {topic}."""

Things I've tried that made no difference

• Running on LangGraph cloud vs locally (via LangGraph Studio)
• Switching between Python 3.11 and 3.12
• Refactoring the parallel nodes as subgraphs instead of plain functions

In my mind, 20 is a surprisingly low ceiling (was hoping for at least ~1000), and so I wanted to know if this is expected? Even better, I'd love to understand why.

Thanks in advance!