Integration test

[tcompa]

As of our last meeting today:

Let's add an integration test which runs on a small dataset (I would aim at something like 2 wells of 2x2 sites, with just a couple of Z levels and a handful channels). The plan is to run yokogawa_to-zarr + illumination_correction + maximum_intensity_projection, and then to compare the three zarr files with reference ones.

Some more work is probably needed before adding this test becomes compelling. At least on these issues:

• Complete the first version of the illumination-correction task (with custom selection of correction images), Illumination correction task fractal-client#62.
• Sort out the channel assignment, OME-Zarr parsing has non-deterministic assignments of channels per well fractal-client#61.
• Reach some stability in the way a workflow is defined, related to the "pipeline file" which we should discuss in https://github.com/fractal-analytics-platform/mwe_fractal/issues/68.

[tcompa]

As of e079db2c0dd067b2f67253e1646cfe69d074c519, we have a simple test that runs on fake images (see tests/Example/run_fake_workflow.sh).

By now it just runs as an example. The plan is to transform it into an actual test (doing everything in a python script, rather than in a bash script), where we run a complex workflow (on a tiny fake dataset) and compare results at each step with some reference results.

[tcompa]

As of 40d00de04e6197f5426da841aedebf042c21050f we have a python script in the tests folder (test_workflow_on_fake_data.py), which runs a trivial workflow on fake data. At the moment it simply tests that the run gets through without errors/exceptions, but later we'll add an actual validation of output against some known values.

[tcompa]

From fractal-analytics-platform/fractal-client#69 (comment):

Very good question! We could save a small dataframe with actual measurements and use a np.assert_almost_equal check on them or something similar (e.g. https://pandas.pydata.org/docs/reference/api/pandas.testing.assert_frame_equal.html?)
For initial checks of the dataframe, we'd have to sample a few example measurements. If we have some test data, I have a workflow to visualize measurements back on label images that allows us to cross-check whether they make sense. I can manually check and then we can just ensure that those measurements don't change.
Maybe the 2x2 test case is reasonably small? Or would we want to go for something smaller (e.g. smaller images maybe? Just one FOV?)

[tcompa]

There is a useful example of fractal-server-agnostic pipeline file in fractal-analytics-platform/fractal-client#150. This is useful to test a plain python workflow, i.e. a collection of tasks with their I/O definitions.

[tcompa]

Close in favor of #2 (which will be closed by #90)