Data in pytest tests

[davidorme]

I'm starting to play around with data loading code, and that involves providing data to load in tests.

One option is provide data in the test files themselves - and for simple text/JSON objects that is probably easiest. However, for more complex files such as NetCDF, creating ByteIO objects within test files so the data only ever exists in memory just seems like it adds a lot of complex code to the test file and having the actual files seems easier?

I've had a look at a couple of options:

• pytest-datafiles
• pytest-datadir

Both of those are currently 2.7 and then 3.8 or lower, but pytest-datadir - which I think has the better API - looks to be on the verge of a new release that drops 2.7 and goes up to 3.10. As in, we could probably add a poetry requirement of their master branch as a stop gap right now.

But - any other options or thoughts?

[dalonsoa]

I'm not sure I understand the issue here. Is there any problem on having real files for tests, JSON, NetCDF or anything else? Unless they have sensitive data or are huge files (>50 Mb), that's what we have done in other projects, just putting them within a subfolder data inside tests, and create an associated feature in conftest.py so they can be used easily within any test.

[davidorme]

I may be over-engineering again! I've seen comments about making sure the tests work on independent copies of files. The python-datadirs package provides a couple of built-in fixture that essentially removes the need to adding a feature in conftest.py.

What kind of feature do you add to conftest.py? Do you mean create a fixture that points to particular sets of files?

[dalonsoa]

It depends of what you need. It might be just the path to the real file. It might be the path to a temporal copy to the file in a temp directory, so it is always a fresh file with every use (this might be what that package does - I've never used it). It might be the file already open and provided as a stream... Up to you, to be honest. But all cases start with a real file in the directory tree.

[davidorme]

OK - in which case, I think python-datadirs is just a plugin that automatically create fixtures that handle that (and uses temporary files automatically). Which seems useful - I suspect a lot of tests may require data, so having those fixtures defined for us seems attractive.

[dalonsoa]

That sounds ok, but just a note of caution here: the more dependencies yo have , the higher the chance for they to break at some point. If the only advantage is to save you creating a simple fixture without adding anything else to the table, it might not be worth it.

[davidorme]

I completely agree that we want to keep dependencies to a minimum. I'll hang on to it for the moment, but it is at least easy to reverse - not a complex API that we'd need to reinvent, simply a shortcut for what may be a common use case.