tskisner
commented
8 years ago Looking at this now as part of the refactor. @sbailey, your text above mentions wanting to "gracefully recover" if some of the inputs to a pipeline step do not exist. This goes against our pipeline planning philosophy so far. We don't currently have the concept of "optional" dependencies. Either they are a dependency or not. If one frame fails sky subtraction, then it seems like that is an error that should be fixed before proceeding with later steps. Currently the pipeline runs each macroscopic "step" in sequence, so all sky files have completed or failed before the stdstar step.
The current
desispec.pipelinelogic incorrectly treatsdesi_fit_stdstars.pyas working on an individual frame, like all of the other calibration steps. However,desi_fit_stdstars.pysimultaneously fits all 3 cameras of a spectrograph so that it finds the best template for each standard star. It only needs to be run once per spectrograph, not once per frame. Inspecting test output from the current pipeline shows thatstdstars-[brz]{n}-{expid}.fitsare identical.The pipeline logic is being refactored, but this throws a new wrinkle in the dependency tree since
desi_fit_stdstars.pycurrently depends upon 10 input files over 3 frames:i.e.
desi_fit_stdstars.pyneeds to wait for all 3 different frames to finish with sky subtraction before it can proceed. I think the currently pipeline may accidentally enforce that by at least running all sky steps before moving on to the stdstar steps.desi_fit_stdstars.pyshould also gracefully fall back to using 2 channels if a 3rd channel is missing. At first I thought that that the pipeline should check those inputs and not bother running unless all 3 exist, but that creates a fragility if one CCD fails. It might be better for the pipeline not to enforce that, but just require 2/3 to exist and then proceed, lettingdesi_fit_stdstars.pyitself handle a missing channel.