ci: Remove duckdb nightly from cov

[dangotbanned]

What type of PR is this? (check all applicable)

• 💾 Refactor
• ✨ Feature
• 🐛 Bug Fix
• 🔧 Optimization
• 📝 Documentation
• ✅ Test
• 🐳 Other

Related issues

• Change arrow() to export record batch reader duckdb/duckdb#18642
• https://github.com/narwhals-dev/narwhals/actions/runs/17096107509/job/48483947375
• https://discord.com/channels/1235257048170762310/1272835922924273694/1407676349925036052

Checklist

• Code follows style guide (ruff)
• Tests added
• Documented the changes

If you have comments or can explain your changes, please do so below

Important

Originally named fix: Handle pa.RecordBatchReader in SparkLikeLazyFrame.collect

[dangotbanned]

Not really sure what to do about this new failure
(https://github.com/narwhals-dev/narwhals/actions/runs/17098212889/job/48487645716?pr=3015)

FAILED tests/expr_and_series/reduction_test.py::test_empty_scalar_reduction_with_columns[sqlframe] - 
 NotImplementedError: This engine does not support schema inference likely since it does not have an active connection.

I've chased down (e6b21ca) already, but this is all stems from a duckdb nightly issue that hasn't been handled yet in sqlframe

@MarcoGorelli do we still need to run duckdb nightly for coverage?
Every other package seems to have the nightly install only in the nightly workflow 🤔

[MarcoGorelli]

@MarcoGorelli do we still need to run duckdb nightly for coverage?

we don't need it anymore 👍

[MarcoGorelli]

[dangotbanned]

@MarcoGorelli do we still need to run duckdb nightly for coverage?

we don't need it anymore 👍

Ahhh, in that case should we just take the simpler route and edit the workflows instead? 😂

Edit

And raise an issue in sqlframe to let them know their API may break soon

[MarcoGorelli]

seems fine to merge, we'll need this fix in place eventually anyway, no?

[dangotbanned]

seems fine to merge, we'll need this fix in place eventually anyway, no?

We might need it, but I'm not 100% sure

https://github.com/eakmanrq/sqlframe/blob/0245f44bbe118067777a5cef3c904666db4993ff/sqlframe/duckdb/dataframe.py#L55-L65

Unless sqlframe also wants to break backwards-compatibility here:

    @t.overload
    def toArrow(self) -> ArrowTable: ...

    @t.overload
    def toArrow(self, batch_size: int) -> RecordBatchReader: ...

    def toArrow(self, batch_size: t.Optional[int] = None) -> t.Union[ArrowTable, RecordBatchReader]:
        self._collect(skip_rows=True)
        if not batch_size:
            return self.session._last_result.arrow()
        return self.session._last_result.fetch_record_batch(batch_size)

They still have the option mentioned in the duckdb PR:

We also have a dedicated function to explicitly return an arrow table with fetch_arrow_table(), and a record batch reader with fetch_record_batch.

I feel like there's a reasonable chance that arrow() call will end up being replaced, WDYT?

[MarcoGorelli]

sure, could be

let's just remove the nightly install from this job then?

[dangotbanned]

sure, could be

let's just remove the nightly install from this job then?

Sounds good, also pinging @eakmanrq!

Edit

Following this duckdb PR, toArrow always returns RecordBatchReader.

• Change arrow() to export record batch reader duckdb/duckdb#18642

[MarcoGorelli]

feeel free to merge once coverage is addressed