To achieve that it is enough to hash inputs, and cache resulting outputs. Repeating a build from scratch with an emtpy cache would not necessarily have to yield the same hashes all they way down to the last artifact, but that's actually a simplification of the whole process, and not a bad thing per se.
> To achieve that it is enough to hash inputs, and cache resulting outputs.
Thing is, inputs can be nondeterministic too - some programs (used to) embed the current git commit hash into the final binary so that a `./foo --version` gives a quick and easy way for bug triage to check if the user isn't using a version from years ago.
Adding the Git hash is reproducible, assuming you build from a clean tree (which the build script can check). Embedding the current date and time is the canonical cause of non-reproducibility, but that can be worked around in most cases by embedding the commit and/or author date of the commit instead.
This is only a problem if those nondeterministic inputs are actually included in the hash. This is often not the case, because the values are included implicitly in the build rather than explicitly.
