Random holdout is used to select an important threshold parameter in the sunrise-sunset estimation algorithm. This can cause small changes in the estimated times from run-to-run. Normally, this isn't an issue because this is assumed to be an approximate value, but downstream applications have been developed that rely on these values. Having them be deterministic is preferable.

Proposed solution: Set the random seed when running this algorithm, so that the same results are returned every run.

Describe alternatives you've considered
We could also expose the random seed keyword to the run_pipeline method, but there are already a lot of keyword arguments to document, and the need to evaluate the randomness of the algorithm feels like a niche use case.