What are the benefits of graph compilation in LangGraph?

[onyourmark]

I'm trying to understand LangGraph's architecture and noticed that graphs must be compiled before execution:
pythongraph = StateGraph(AgentState)

... add nodes and edges ...

app = graph.compile() # Required compilation step
app.invoke(data)
This compilation requirement creates certain constraints, like needing to declare all conditional paths upfront:
python# Must map all possible paths at compile time
graph.add_conditional_edges("node", router, {
"path_a": "node_a",
"path_b": "node_b"
})
I'm curious about the design decision to require compilation:

What specific benefits does compilation provide in LangGraph's use case?

Is it primarily for performance optimization?
For validation before execution?
For graph visualization?

Were interpreted/runtime evaluation approaches considered?

What tradeoffs led to choosing compilation?

Given that LLM calls typically take 100s of milliseconds, does routing optimization from compilation provide measurable benefits?

I'm asking because I'm trying to understand graph orchestration patterns better, and LangGraph's compilation approach is different from some other workflow systems I've seen. Would love to understand the reasoning behind this architectural choice.
Thanks for creating this framework!