@zacchiro @GuilloteauQ @hannes @Zimm_i48 @luj My impression is that most people who believe in Docker as a reproducibility silver bullet have a very superficial understanding of reproducibility and why / when / how it matters. Pointing out to them what precisely does *not* work with Docker should help with explaining the issues. So... thanks for this work!

@civodul

@zacchiro very interesting indeed! How deoes this lack of reproducibility translate in reproducibility of the experiments the docker images were set up for, i.e. did you find a correlation between a higher discrepancy in packages and a lower statistical significance of the difference in downstream experimental results? I have been recently inspired by https://www.sfscon.it/talks/why-you-cannot-apt-install-an-llm-yet/ and would like to find a sound, principled way to make data+code+exp available for reproducibility of ML-based packages