Workflow evaluation graphs do not take into account workflow inputs

[peterhuene]

I found this example from the WDL 1.0 spec:

version 1.1

task my_task {
    input {
        Int i
    }

    command <<<>>>

    output {
        Int out = i
    }
}

workflow foo {
  input {
    Int x = 10
    Int y = t1.out
  }

  call my_task as t1 { input: i = x }
  call my_task as t2 { input: i = y }
}

If you were to call foo without a value for y, t2 would depend on t1 (via y) and would have to be executed in series.

If you were to call foo with a value for y, t1 and t2 could be executed in parallel as there is no dependency edge between them.

The issue in wdl-engine is that we currently build the evaluation graph based solely on the workflow statement and always introduce dependency edges from any of the input default expressions.

Instead, this process should take into account any inputs provided to the workflow and skip adding dependency edges from the default expressions if a value was provided.

[peterhuene]

The tricky part will be testing this with an assertion that the tasks ran in parallel, as workflow evaluation is a bit of black box in that regard.

I think perhaps a unit test on WorkflowEvaluationGraph where we check that an edge exists in the graph when not given the input vs. the edge does not exist when given the input will suffice.

[dead8309]

Hi @peterhuene can you please assign me this issue, as per our conversation on slack. I've made the necessary changes. I'm currently writing unit test for this.

Sending a patch in a few minutes. Thanks!