cugraph Python tests should download test data only once

[rlratzel]

cugraph Python tests that use the dataset package may end up downloading the same test data multiple times when the download=True arg is set, as seen below:

E = dataset.get_edgelist(download=True)

This could be needed as a workaround for supporting pytest-xdist where different parallel runs are deleting downloaded test data as part of a cleanup step (just a guess). Ideally tests can read tests data from a previous download step, or even better, if a CI run has access to the source repo which includes test data, no download step would be needed at all.

[nv-rliu]

To my knowledge, each CI job pulls a fresh cugraph repo , which means it will have all the datasets currently present in the repo without explicit downloads. Also during the recent effort to support xdist, the tests were vetted to ensure that only those testing the download functionality should be relying on download=True and the other tests can read from what's already there. Also, certain tests may use datasets not present in the repo for various reasons (not tracked yet, size, etc), which will then also be downloaded at runtime.

In summary, the download step is there and turned on when we test for that functionality. Otherwise, it is left off.

[rlratzel]

To my knowledge, each CI job pulls a fresh cugraph repo , which means it will have all the datasets currently present in the repo without explicit downloads. Also during the recent effort to support xdist, the tests were vetted to ensure that only those testing the download functionality should be relying on download=True and the other tests can read from what's already there. Also, certain tests may use datasets not present in the repo for various reasons (not tracked yet, size, etc), which will then also be downloaded at runtime.

In summary, the download step is there and turned on when we test for that functionality. Otherwise, it is left off.

I was initially thinking these tests used only datasets in the repo and we shouldn't need download=True, but since that's not true, that option makes sense. We should open a separate issue if we want those datasets not in the repo to be downloaded using the download script CI calls instead.