[ozenyunus35]

Which issue does this PR close?

• Closes [Bug] unbase64 rejects input that Spark's MIME decoder accepts #26340.

Rationale for this change

Spark's unbase64 decodes with Java's MIME Base64 decoder, which accepts missing padding and skips characters outside the Base64 alphabet. datafusion-spark rewrote unbase64 to the core decode(..., 'base64pad'), so inputs like '123', 'YQ' or 'YW Jj' failed here while Spark decodes them.

What changes are included in this PR?

unbase64 no longer simplifies to decode; it decodes each value with a small mime_decode function:

• characters outside the standard alphabet (A-Z a-z 0-9 + /) are skipped, and padding is optional;
• a final unit of one character is an error ("last unit does not have enough valid bits");
• = must complete the current unit (xx== or xxx=), and only non-alphabet characters may follow it; anything else is an "invalid padding" error.

Binary, LargeBinary and BinaryView inputs are handled; the return type is unchanged (LargeBinary for LargeBinary, Binary otherwise). base64 is unchanged.

I took the expected results from running java.util.Base64.getMimeDecoder().decode(...) on JDK 21 for each input in the tests, including the edge cases around padding (YQ=, YQ= =, YQ===, YQ==x, YWJj=, YQ==YQ==).

What is the testing strategy for this PR?

• base64.slt: unbase64('123') (string and binary) now returns d76d instead of an error, as the issue asks, plus cases for missing padding, skipped characters, an input with only skipped characters, xxx=, a LargeBinary input, and the three error cases. Six of these fail on main.
• Unit tests for mime_decode cover the accepted inputs and the error cases above.
• cargo test -p datafusion-spark --lib string::base64, the spark/string/base64 sqllogictests and cargo clippy -p datafusion-spark --all-targets -- -D warnings pass.

Are there any user-facing changes?

Yes: unbase64 now accepts the inputs Spark accepts (missing padding, characters outside the alphabet), instead of failing.

I used an AI assistant to help draft parts of the code and tests; I reviewed the changes and ran the tests myself.

[alamb]

Thanks for the contribution! I am closing this under DataFusion's AI Dump policy.

https://datafusion.apache.org/contributor-guide/ai-policy.html#what-is-an-ai-dump-and-why-it-is-not-helpful

Review capacity is very limited and this is one of three Spark compatibility PRs opened the same day across unrelated functions, with no engagement since. The policy asks new contributors to work with maintainers on one PR at a time rather than opening several at once.

If this fix matters to you, please pick the single change you care most about, reopen or resubmit it with a short description in your own words, and we are happy to help with review.