I'll take a look, but one high-level thought: we should exclude operations from this that might have side effects. Specifically, remote function calls could do all sorts of things (send email, create files in GCS, etc.) without the return value being used.

I'll take a look, but one high-level thought: we should exclude operations from this that might have side effects. Specifically, remote function calls could do all sorts of things (send email, create files in GCS, etc.) without the return value being used.

There is a lot that needs to be done to properly model intended side-effects. Lazy evaluation can result in side effects being repeated, while filter pushdown (client side or bigquery side) can remove calls. I think side-effects should probably be explicitly indicated through a special api (maybe with no-output consumer udfs?), and we can then model as its own node type.