Update toggle test failing at random

[alecandido]

I haven't identified yet any random number involved, but it seems that this test is failing at random

qibocal/tests/test_task_options.py

Lines 103 to 125 in b0b569a

	@pytest.mark.parametrize("global_update", [True, False])
	@pytest.mark.parametrize("local_update", [True, False])
	def test_update_argument(platform, global_update, local_update, tmp_path):
	"""Test possible update combinations between global and local."""
	NEW_CARD = modify_card(
	UPDATE_CARD, local_update=local_update, global_update=global_update
	)
	# platform = deepcopy(GlobalBackend().platform)
	old_readout_frequency = platform.qubits[0].readout_frequency
	old_iq_angle = platform.qubits[1].iq_angle
	Runcard.load(NEW_CARD).run(
	tmp_path,
	mode=AUTOCALIBRATION,
	platform=platform,
	)
	if local_update and global_update:
	assert old_readout_frequency != approx(platform.qubits[0].readout_frequency)
	assert old_iq_angle != approx(platform.qubits[1].iq_angle)
	else:
	assert old_readout_frequency == approx(platform.qubits[0].readout_frequency)
	assert old_iq_angle == approx(platform.qubits[1].iq_angle)

https://github.com/qiboteam/qibocal/actions/runs/10219428483/job/28277670887
(the second attempt is passing, without having changed anything - also noticed that is just failing on win-py3.11 combination, but working on win-py3.9, linux, and darwin, so it seems to be just a chance)

[alecandido]

@andrea-pasquale it seems you wrote this test. Any clue about?

(it should not be difficult to debug, anyone can do it, so it doesn't have to be necessarily you - but maybe you already had an idea)

[andrea-pasquale]

Yes, I know.
The test is failing when the lorentzian fit fails therefore the parameters are not updated.
As a patch we could keep the test only for the single shot classification which shouldn't ever fail.

[alecandido]

The test is failing when the lorentzian fit fails therefore the parameters are not updated.

But how is it possible that the Lorentzian fit will pass just rerunning?
Is curve_fit using some sort of random numbers internally? In case: could we fix the seed for that?

[andrea-pasquale]

The output is different every time because we are testing with dummy, right?
Given that we don't fix the seed of the platform several runs will have different results.
So one option is to fix the seed of dummy

[alecandido]

That is definitely a seed that should be fixed.

Then, we may still decide that the fit is too unstable to be tested in the CI, especially with dummy. And we'd rather design a closure test for it.
But all seeds should be fixed, for reproducibility.