Functional Transforms¶

All transforms return a new FrozenFrame. The original is never modified.

Schema-preserving operations¶

These operations change the rows but not the schema. They return the same concrete subtype — not a generic FrozenFrame base class.

filter¶

top: UserMetrics = df.filter(df.score > 8.0)
active: UserMetrics = df.filter(df.active == True)

Filter expressions are built from FrozenSeries comparisons. They produce a lazy predicate evaluated against the underlying Arrow array — not a Python list of booleans.

sort¶

ranked: UserMetrics = df.sort("score", descending=True)
multi:  UserMetrics = df.sort("active", "score", descending=[False, True])

slice¶

first_ten: UserMetrics = df.slice(0, 10)

Schema-altering operations¶

These operations change the columns and therefore the schema. They return a FrozenFrame instance (full static typing of the output schema is a Phase 3 goal — see the roadmap).

select¶

slim = df.select("user_id", "score")

rename¶

renamed = df.rename({"user_id": "id", "score": "metric"})

drop¶

without_active = df.drop("active")

with_column¶

Add a new column from a FrozenSeries or Arrow array:

import pyarrow as pa
labels = pa.array(["high", "low", "high"])
labelled = df.with_column("label", labels)

join¶

result = df.join(other_df, on="user_id", how="inner")

map_column¶

map_column applies a typed function to a single column and returns a new frame with that column replaced. It is the intended replacement for Pandas' apply — and unlike apply, the function signature is type-checked:

def normalize(s: FrozenSeries[float]) -> FrozenSeries[float]:
    ...

normed: UserMetrics = df.map_column("score", normalize)

Note

There is no apply or map escape hatch in freezeFrame. map_column with a typed signature is the only supported UDF path, by design.