A Question about graph composition

[Meguminnnnnnnn]

There is a graph：

In the graph, nodeC depends on the output of nodeA and nodeB, nodeB depends on the output of nodeA. I need to add a 'Passthrough' node to pass nodeA's output to nodeC.

Why don't LangGraph execute a node when all its preceding nodes are ready， like tensorflow, so that we can compose graph like:

I‘m concerned about why LangGraph decide to use the mode that execute the next node of each current node but not the mode like tensorflow?

[hwchase17]

not sure i understand, how are you defining your state/graph? depending on how you are defining your state you may just be able to to do a -> b -> c

[Meguminnnnnnnn]

Yes, I understand your thought. This question is just a leading one. It made me think, why does LangGraph operate according to the principle of "executing the node following each one", rather than "executing it when all the upstream nodes of a node are completed", similar to TensorFlow.
I think that the latter, compared to the former, does not have obvious weaknesses and can better handle concurrent situations and data transmission issues as in the above example.
What are the reasons for your choosing the former?

[nfcampos]

Hi, what you're looking for can be achieved by adding an edge with two sources, like this

builder.add_edge(['A', 'B'], 'C')

This edge will wait for both sources to trigger the destination C.

Note that if you declare separate edges, eg.

builder.add_edge('A', 'C')
builder.add_edge('B', 'C')

then we interpret that as meaning you want to go to C twice, once after A finishes, another time after B finishes. If they happen to finish in the same step we'd dedupe that and visit C only once.