[cyberspace-cs]

What

LazyFileStorage.write calls DataFrame.applymap, which was removed in pandas 3.0. Any operator that writes a DataFrame through LazyFileStorage now fails with:

AttributeError: 'DataFrame' object has no attribute 'applymap'

Why this is live, not a deprecation warning

requirements.txt does not constrain pandas, so a fresh install today resolves to 3.x and hits this on the first write. The bundled example pipelines never reach it because they write through FileStorage, not LazyFileStorage — which is exactly why the failure is invisible until someone uses the lazy path.

This is the site #371 missed

#371 ("switch df.applymap() to df.map") changed one line in this file — FileStorage.write — and left LazyFileStorage.write on the removed API. The two classes are siblings with the same shape:

class	method	on main
FileStorage	write	data.map(clean_surrogates) — fixed by #371
LazyFileStorage	write	data.applymap(clean_surrogates) — this PR

Fix

DataFrame.map is the replacement and has existed since pandas 2.1. Keep a fallback so environments still on an older pandas are unaffected:

if hasattr(data, "map"):
    dataframe = data.map(clean_surrogates)
else:
    dataframe = data.applymap(clean_surrogates)

Evidence

Bidirectional check against 05a7d9e (current main), Python 3.13.14 / pandas 3.0.6:

# test only, fix reverted
$ pytest test/cpu_only/test_lazy_storage_dataframe_write.py
1 failed, 1 passed
E   AttributeError: 'DataFrame' object has no attribute 'applymap'

# with fix
$ pytest test/cpu_only/test_lazy_storage_dataframe_write.py
2 passed

The second test does not just assert "no exception": it checks that surrogate cleaning still happens and that the written file round-trips, so the fix cannot silently change behaviour.

grep -rn applymap dataflow/ now returns only the fallback branch above.

Happy to drop the fallback if you consider pandas < 2.1 out of scope.

[SunnyHaze]

LGTM, thanks. PR will be merged after CI check.