There seems to be a problem with polars being called via ctest: in my venv it works locally (using run.sh), but ctest fails for me in the same bash session.

A problem is that we currently overwrite the PATH at configuration time. So any further modifications are not picked up.

I just saw that CMake 3.22 (our new baseline) provide ENV modification support for exact this use-case. https://cmake.org/cmake/help/v3.22/prop_test/ENVIRONMENT_MODIFICATION.html

I am wondering why it worked or works for the precice-profiling script (also used in the mapping tester), whereas it fails for the compare script. The precice-profiliing script is called through the gathering script os.subprocess, maybe that's the simple answer.

In my venv, the tests pass now. The CI doesn't seem to be happy

https://github.com/precice/aste/actions/runs/11552451140/job/32151545870?pr=212#step:10:447

I am not sure if the actual solution here would be more simple.

+ python3 /Users/runner/work/aste/aste/examples/mapping_tester/../../tools/mapping-tester//compare.py reference-statistics.csv test-statistics.csv
The test results differ in both files
DataFrames are different (value mismatch for column 'median(abs)')
[left]:  [0.00565710724966223, 0.0004046891096261551]
[right]: [0.00567422622697189, 0.00040468910964214233]

Isn't this a property from the mapping accuracy that should be unchanged?

The polars comparision uses rtol of 1e-05 and an atol of 1e-08.

https://docs.pola.rs/api/python/stable/reference/api/polars.testing.assert_frame_equal.html#polars.testing.assert_frame_equal

ASTE generates, exports, partitions to VTU with binary data and uses the VTK-internal calculator.

The only moving targets are the VTK version and the numpy version, which we use to calculate the percentiles/median.

Maybe they changed the interpolation method? Still uses linear interpolation

The assert checks col by col and the first column is the median. So it is possible that other columns don't match too.

Do you have the complete result table for both runs?

Could it be a matter of the column order or something?

After a morning of debugging we realised that the partitioning wasn't deterministic, which lead to major differences in NP and even tiny run to run differences in RBF TPS.
We never faced this issue before as we are now using separate cases, compared to
separate runs. Using topology-based partitioning solves the issue and leads to reproducible cases.

Would require a documentation update.

This is primarily to get the CI working. I'll revert later

But it will remain a requirement to execute the tests, right? In the past, we also had tests for the mapping-tester here, which has optional dependencies. Then, however, users install the software, run ctest and see that tests are failing.
How does this behave here then?

Currently, the mapping-tester-test needs precice-profiling analyze which requires polars anyway. So the dependency is already there.

If we merge #210, then the timings and thus the dependency on polars becomes optional.
In that case, it may be an idea to implement the compare without polars.

On the flip side, this then becomes a bad test, as it tests two different things depending on the environment. Maybe a warning in CMake is enough here.