(See: https://github.com/python/cpython/pull/28107#issuecomment-915627148)

The output from marshal (e.g. PyMarshal_WriteObjectToString(), marshal.dump()) may be different depending on if it is a debug or non-debug build.  I found this while working on freezing stdlib modules.

FYI, I came up with a fix (for frozen modules, at least) in  https://github.com/python/cpython/pull/28107.

One consequence of this is that frozen module .h files can be different for debug vs. non-debug, which causes CI (and Windows builds) to fail.

New changeset cbeb81971057d6c382f45ecce92df2b204d4106a by Eric Snow in branch 'main':
bpo-45020: Freeze some of the modules imported during startup. (gh-28335)
https://github.com/python/cpython/commit/cbeb81971057d6c382f45ecce92df2b204d4106a

I would propose that marshal internally make an extra pass over its input in order to determine which objects are referenced multiple times. This will speed up reading marshalled data (in addition to addressing the reproducibility issue with debug builds) at the cost of slowing down writing it, so there may need to be a way for 3rd party users to turn this off (or a way for importlib and compileall to turn it on).

That's a good idea.  It's certainly cleaner than the approach I took (optionally pass in to marshal.dumps() the list of "before" object/refcount pairs to compare in w_ref()).

Adding a flag to marshal.dumps() to opt out shouldn't be too big a deal.  (I expect all users of marshal will want the improvement by default.)

FYI, This issue is duplicate of https://bugs.python.org/issue34093, and I had made two pull requests to solve the issue.

Thanks, Inada-san.  That's super helpful.

I'm closing this in favor of bpo-34093.