pl-row-encode 0.3.1

Row-level, type-preserving encode/decode for Polars columns

pl-row-encode

A Polars plugin for row-level, type-preserving encode/decode.

encode(*cols) packs a set of columns into a single Binary column where each value is an opaque, self-describing token: the polars-row encoding of the row, prefixed with an embedded schema header. decode_series(...) reverses it back into a Struct, recovering the original dtypes without needing any external schema.

DataFrame
  -> encode(*cols)
  -> opaque bytes
  -> decode(...)   # (row bytes -> Struct -> original typed columns)
  -> DataFrame

The type information rides with the token and can be decoded on the spot at some later date.

Token layout

Each Binary value is:

[ u32 header_len (LE) ][ header bytes ][ row bytes ]

header is a bincode-serialized Vec<Field> (logical schema); row bytes is the unordered polars-row encoding of that single row. Embedding the header per value makes every token independently decodable.

Usage

import polars as pl
from pl_row_encode import encode, decode_series

df = pl.DataFrame({"id": [1, 2], "name": ["alice", "bob"]})

tokens = df.select(tok=encode("id", "name"))["tok"]   # dtype: Binary
# ... hand `tokens` to a vendor, get them back ...

decoded = decode_series(tokens).struct.unnest()        # back to id / name with dtypes

For the lazy engine, the output Struct dtype must be known up front, so pass a token's header explicitly:

from pl_row_encode import decode
header = ...  # the [u32 len][header] prefix of any token
lf.select(decode("tok", schema_header=header)).collect()

Development

make develop   # build the Rust extension into the venv (uv run maturin develop)
make test      # build + run pytest
make lint      # ruff + ty

The first make develop compiles the full Polars Rust workspace and takes a few minutes; subsequent builds are incremental and fast.

Notes / limitations

• Built on polars-row, the same machinery Polars uses internally for sort/group-by row encoding — lossless for primitive, string, boolean, temporal, and nested types.
• decode_series infers the schema from the first non-null token, so an all-null/empty Series needs the explicit decode(schema_header=...) form.