`daft.count('all')` for use in `group-by` without any column names

[MarcoGorelli]

Is your feature request related to a problem?

Related to #3985, but I'm looking for a daft.len() expression to use in group-by

Describe the solution you'd like

import daft
df = daft.from_pydict({'a': [1,1,2], 'b': [4,5,6], 'c': [7,8,9]})
df.groupby('a').agg(daft.col('b').sum(), daft.count('all').alias('count'))

Describe alternatives you've considered

Some previous discussion at #3985 , but there I was pointed to Expression.count. That works, but outside the group-by context when a column names is passed, e.g.

In [40]: df.agg(daft.col('a').count('all')).collect()
Out[40]:
╭────────╮
│ a      │
│ ---    │
│ UInt64 │
╞════════╡
│ 3      │
╰────────╯

Additional Context

This is for Daft support in Narwhals

Would you like to implement a fix?

No

[colin-ho]

Currently we don't have a standalone daft.len or daft.count expression, I'm guessing the intention here is to do something like a COUNT(*) in sql? In which case would defaulting to a count on the first column work?

Example:

df = df.groupby('a').agg(daft.col('b').sum(), df[0].count().alias('count'))

[MarcoGorelli]

i'm afraid not, I need an expression object, as it may be composed with other operations

[MarcoGorelli]

I'm currently implementing this as:

    def len(self) -> DaftExpr:
        def func(_df: DaftLazyFrame) -> list[daft.Expression]:
            if not _df.columns:
                msg = "Cannot use `nw.len()` on Daft DataFrame with zero columns"
                raise ValueError(msg)
            return [daft.col(_df.columns[0]).count(mode="all")]

        return DaftExpr(
            call=func,
            evaluate_output_names=lambda _df: ["len"],
            alias_output_names=None,
            backend_version=self._backend_version,
            version=self._version,
        )

This gives the correct results, but involves calculating .columns, which I know for some other lazy backends can have some overhead

Is the same true for Daft, or is .columns OK to call here?

[colin-ho]

It's ok to call .columns, there is no overhead here, as it internally just checks the schema.

[MarcoGorelli]

cool thanks, closing then 👍