Nick Crews

hmm, you are definitely right, `pg.create_view("my_view", duckdb_con.table("my_table"))` should error. Limiting to `.create_table()` would be totally adequate.

See https://github.com/ibis-project/ibis/issues/7139, I asked for this same thing there originally, and then we restricted the scope to Table.sample(). Are there other use cases besides sampling that you need this for?

I think you can do ```python t = t.mutate(_id=ibis.row_number()) test = t.sample(fraction=fraction) train = t[~t._id.isin(rest._id)] test = test.drop("_id") train = train.drop("_id") ``` ?

No idea re speed, I think benchmarks sound like a great idea! Perhaps once you play around there you will find a way that works with Table.sql() and then you...

@cpcloud relatedly, int64 columns with nulls are turned into floats when executed into pandas dataframes. Really those should be pd.Int64(). And similar for all other nullable int dtypes.

Yeah sgtm. One tricky case: consider an int64 column: if it has nulls then it should be turned into an pd.Int64(), but what if there are no nulls? Should it...

Cool, always going to nullable works for me

I can make a new issue for this if you ask, but I think I ran across another flavor of this today. Is this what we want the behavior to...

Oops, indeed you are right!

I don't actually need an exact amount, but somewhere in the ballpark, so I just use fraction and the result is usually within 10% of my goal. So I don't...