Why can we not provide deps to graph assets?

[CompRhys]

I would like to make a graph asset depend on another asset in a non_argument_deps manner but there's not an obvious way to do this?

It seems like a strange omission from the API and I can't work out what would block it?

[OwenKephart]

Hi @CompRhys ! Good question. From a pure API perspective, it definitely would be a natural thing to have. The explanation for why it's non-trivial to add is essentially:

• GraphDefinitions are not functionally distinct from an "arbitrary bag of OpDefinitions" in the core orchestration layer
• This means that a GraphDefinition itself does not have its own independent inputs -- each individual op inside of the graph will execute as soon as its inputs are satisfied.
  • For example, if you had a graph that had one op with a mapped input and one op with no dependencies, if you executed a run that contained that graph, the op with no dependencies would execute immediately, regardless of the upstream dependencies of the other op / the graph as a whole.
• So, if you wanted to declare new (non-argument) inputs to the "graph as a whole", this would involve wiring up those graph inputs to all root ops within your graph (such that no ops would execute before all those inputs were satisfied), which would involve creating new copies of those op definitions that have additional inputs to map those new graph inputs into.

This sort of messiness is what's prevented us from adding support for that argument so far.

[CompRhys]

Thanks for the explanation. Essentially I would like to be able to show that the graph asset depends on an earlier partitioned asset being run in the global asset tree but the graph asset uses different partition logic and so pulls in the data using a resource rather than taking the outputs of the earlier partitioned asset via an an i/o manager. The first op within the graph asset takes in the input file and then creates a Dynamic out to apply different subroutines to process the data.

Perhaps the solution I want would be that I can put a deps on an op but then I don't think that would end up reflected in the asset view?

[OwenKephart]

Hi, just saw this reply -- you can use a Nothing-type input to define an input to an op that is not explicitly loaded using an io manager: https://docs.dagster.io/guides/build/ops/graphs#defining-nothing-dependencies

This might be what you want here. That Nothing-type input can then be mapped to the upstream asset key. Under the hood, this is what happens automatically when you set deps with an @asset decorator, it's just that we can't do this automatically in the same way for graphs.