[R-pkg-devel] check_equal() test fails on ATLAS

[Noah Greifer]

Hello all,

My package *WeightIt* is failing
<https://cran.r-project.org/web/checks/check_results_WeightIt.html> the
CRAN test on ATLAS. The failure is due
<https://www.stats.ox.ac.uk/pub/bdr/Rblas/ATLAS/WeightIt.out> to a failed
*testthat* test comparing two covariance matrices using check_equal(). A
few things are confusing to me, so I'm not sure how to solve it.

First, the covariance matrices should be identical; there is no random
component in their computation and the the only difference in the calls
used to generate them are whether an argument is omitted or whether it is
explicitly set to its default (i.e., the purpose of the test is to ensure
the default argument corresponds to the supplied argument). I am willing to
accept that there is some non-deterministic process in ATLAS, e.g., where
convergence is measured based on time rather than number of iterations,
though I don't know if this is actually the case.

Second, the package tests have passed on previous CRAN checks (this is
weeks after this version was accepted to CRAN), on all other versions of R,
and in the rhub test
<https://github.com/ngreifer/WeightIt/actions/runs/11489721542/job/31979056981#step:6:124>
with ATLAS.

Does anyone have any advice on how to solve this? I know I can change the
tolerance argument to check_equal() but this is a case where the output
should be identical. I cannot reproduce the problem with rhub, so I can't
see where the problem is. Could this be a false alarm due to caching?
Thanks for any insights.

Noah

	[[alternative HTML version deleted]]