fastdataframe 0.2.0

A fast dataframe implementation with Pydantic integration

FastDataFrame

FastDataFrame bridges Pydantic models and dataframe/table backends. A FastDataFrame model owns backend-neutral column definitions, and backend modules expose stateless functions for Polars, PyArrow, and Apache Iceberg.

Supported backends:

• Polars DataFrame and LazyFrame
• PyArrow schemas
• Apache Iceberg schemas/tables through PyIceberg

Core idea

Define the schema once:

from typing import Annotated

from pydantic import BaseModel, Field

from fastdataframe import ColumnInfo, FastDataFrameModel, Int32


class User(BaseModel):
    user_id: Annotated[
        int,
        Field(validation_alias="userId", serialization_alias="user_id"),
        ColumnInfo(dtype=Int32()),
    ]
    name: str
    score: float = 0.0
    nickname: str | None = None


FastUser = FastDataFrameModel.from_base_model(User)

Then generate backend-native schemas with stateless backend functions:

import fastdataframe.polars as fpl
import fastdataframe.pyarrow as farrow
import fastdataframe.iceberg as fice

polars_schema = fpl.schema(FastUser)
arrow_schema = farrow.schema(FastUser)
iceberg_schema = fice.schema(FastUser)

Column definitions

FastDataFrameModel owns immutable, backend-neutral column definitions:

FastUser.column_definitions
FastUser.column_map

ColumnInfo is optional user-authored metadata. Fields without ColumnInfo receive default metadata.

class Trade(FastDataFrameModel):
    trade_id: str
    quantity: Annotated[int, ColumnInfo(dtype=Int32(), is_unique=False)]

Name accessors

Resolved names are available as immutable accessors keyed by Python field name:

FastUser.serialization_names.user_id  # "user_id"
FastUser.validation_names.user_id     # "userId"
FastUser.storage_names.user_id        # "user_id"
FastUser.serialization_names["user_id"]

The storage name is the canonical dataframe/table column name and defaults to the Pydantic serialization name.

Dtype refinements

ColumnInfo(dtype=...) can refine backend schema generation while the Python annotation remains the semantic type.

Initial backend-neutral scalar dtypes include:

• Boolean, String, Binary
• Int8, Int16, Int32, Int64
• Float32, Float64
• Date, Time, Timestamp
• Decimal

Unsigned integer dtypes are intentionally not included initially. Small signed integers are widened when mapped to Iceberg where necessary.

Polars

import polars as pl
import fastdataframe.polars as fpl

raw = pl.DataFrame({"user_id": ["1"], "name": ["Alice"], "score": ["1.5"], "nickname": [None]})

cast_df = fpl.cast(FastUser, raw)
errors = fpl.validate_schema(FastUser, cast_df)

fpl.string_schema(FastUser) returns a schema with all columns as strings for ingest flows.

PyArrow

import fastdataframe.pyarrow as farrow

schema = farrow.schema(FastUser)
string_schema = farrow.string_schema(FastUser)

PyArrow schemas encode nullability from Optional / None unions. Pydantic defaults do not imply nullable storage.

Iceberg

import fastdataframe.iceberg as fice

schema = fice.schema(FastUser)

Iceberg migration support is additive-only by default:

fice.apply_additive_migration(FastUser, table)

Destructive deletes are intentionally not automatic.

For Polars-to-Iceberg persistence, data is written through the FastDataFrame-generated PyArrow schema boundary:

fice.append_polars(FastUser, table, cast_df)

This is important because Polars schemas do not encode column nullability in the same way as PyArrow and Iceberg.

Column lifecycle

Deprecated fields remain model fields, remain in schemas, and must be nullable:

class UserV2(FastDataFrameModel):
    old_score: Annotated[float | None, ColumnInfo(deprecated=True)]
    score: float

Deprecated and removed column names can be reserved through model config to prevent unsafe reuse:

class UserV3(FastDataFrameModel):
    model_config = {
        "fastdataframe_deprecated_column_names": {"old_score"},
        "fastdataframe_removed_column_names": {"very_old_score"},
    }

    score: float

Removed names remain reserved even if a backend later physically deletes the column.

Installation

pip install fastdataframe
# or with optional backends
pip install 'fastdataframe[polars,pyarrow,iceberg]'

Development

uv sync --all-extras
uv run pytest tests/
uv run ruff check .
uv run ruff format .
uv run ty check

License

MIT