`loadListing` behavior in workflow steps

[LanderOtto]

I have encountered in cwltool a situation where loadListing behavior seems inconsistent when a Directory is passed from a Workflow input to a CommandLineTool step with different loadListing requirements.

In my scenario the workflow defines an input Directory with loadListing: deep_listing, and a step (CommandLineTool) takes that Directory as input but defines loadListing: no_listing.

The directory object inside the CommandLineTool still contains the listing, even though the step explicitly requests no_listing.
However, I expect the step input to respect its own loadListing: no_listing definition rather than inheriting the upstream state.

Is the current behavior of cwltool correct?

Here an example.
I expect out_none: true instead i got out_none: false

#!/usr/bin/env cwl-runner
class: Workflow
cwlVersion: v1.2
inputs:
  d2:
    type: Directory
    loadListing: deep_listing
outputs:
  out_none:
    type: boolean
    outputSource: step_none/out
steps:
  step_none:
    in:
      d: d2
    out: [out]
    run: 
      class: CommandLineTool
      cwlVersion: v1.2
      requirements:
        InlineJavascriptRequirement: {}
      inputs:
        d:
          type: Directory
          loadListing: no_listing
      outputs:
        out:
          type: boolean
          outputBinding:
            outputEval: $(inputs.d.listing === undefined)
      baseCommand: "true"

[LanderOtto]

@mr-c what do you think?

[mr-c]

I guess this happened because cwltool did indeed load the listing for the Directory at the workflow level, even though it wasn't used at the workflow or workflow step level. It then serialized that object when evaluating the outputEval as it as in the same memory process. I also tested this with toil-cwl-runner and a Slurm-managed cluster and the same error occurred; here toil must have pickled the directory listing and passed it to the tool.

Lets look at the CWL v1.2 standard; is retaining an existing directory listing against the normative text?

https://www.commonwl.org/v1.2/CommandLineTool.html#LoadListingEnum

Specify the desired behavior for loading the listing field of a Directory object for use by expressions.

Symbols

symbol	description
no_listing	Do not load the directory listing.
shallow_listing	Only load the top level listing, do not recurse into subdirectories.
deep_listing	Load the directory listing and recursively load all subdirectories as well.

(emphasis is mine)

So we are told that the desired behavior is not to load, but we aren't being required to trim the listing.

Now, I agree that this can produce inconsistent results and lead to variation between implementations, which is not great.

Now I ask you, should we change this in CWL v1.3 to require trimming the location field?