polars-map

Polars plugin providing a Map extension type and functions. Maps represent a mapping from unique keys of any type to values, and are stored as List(Struct({key, value})) columns. Most functions in the .map namespace accept either the Map extension type or the underlying List(Struct). The type-preserving methods (filter, filter_keys, filter_values, merge, intersection, difference) don't remap the key or value types, so they reuse the input's dtype instead of re-inferring it (which would lock the GIL). On the expression API that reuse requires the Map extension as input; cast a plain List(Struct) first if needed, e.g. with .map.from_entries(). The equivalent Series methods accept either.

Installation

pip install polars-map

Supported operations (.map.*)

Category	Methods
Accessors	entries, keys, values, len, get, contains_key
Filtering	filter, filter_keys, filter_values
Transform	eval, eval_keys, eval_values
Set ops	merge, intersection, difference
Conversion	from_entries
Iteration	__iter__, to_list (Series only)

Arrow conversion

Function	Description
from_arrow(table)	Arrow Table/RecordBatch to Polars DataFrame, preserving map<> as Map
from_arrow_array(array)	Arrow Array to Polars Series, preserving map<> as Map
to_arrow(frame)	Polars DataFrame to Arrow Table, converting Map back to map<>
to_arrow_array(series)	Polars Series to Arrow Array, converting Map back to map<>
scan_arrow(source)	Lazy scan from an Arrow source with Map preservation

Usage

import polars as pl
import pyarrow as pa
from polars_map import Map, from_arrow, to_arrow, scan_arrow

# Construction
ser = pl.Series(
    "m",
    [
        [{"key": "a", "value": 1}, {"key": "b", "value": 2}],
        [{"key": "x", "value": 10}],
    ],
    dtype=Map(pl.String(), pl.Int64()),
)
df = pl.DataFrame([ser])

# Accessors
df.select(pl.col("m").map.keys())    # [["a", "b"], ["x"]]
df.select(pl.col("m").map.values())  # [[1, 2], [10]]
df.select(pl.col("m").map.len())     # [2, 1]

# Lookup
df.select(pl.col("m").map.get("a"))           # [1, None]
df.select(pl.col("m").map.contains_key("a"))  # [True, False]

# Filtering
df.select(pl.col("m").map.filter(pl.element().struct["value"] > 1))
df.select(pl.col("m").map.filter_keys(pl.element() > "a"))
df.select(pl.col("m").map.filter_values(pl.element() >= 2))

# Transform keys or values
df.select(pl.col("m").map.eval_keys(pl.element().str.to_uppercase()))
df.select(pl.col("m").map.eval_values(pl.element() * 2))

# Merge (right-side wins on key conflict)
left = pl.Series("l", [[{"key": "a", "value": 1}, {"key": "b", "value": 2}]], dtype=Map(pl.String(), pl.Int64()))
right = pl.Series("r", [[{"key": "a", "value": 99}, {"key": "c", "value": 3}]], dtype=Map(pl.String(), pl.Int64()))
pl.DataFrame([left, right]).select(pl.col("l").map.merge(pl.col("r")))
# [{"a": 99, "b": 2, "c": 3}]

# Set operations
pl.DataFrame([left, right]).select(pl.col("l").map.intersection(pl.col("r")))  # keys in both
pl.DataFrame([left, right]).select(pl.col("l").map.difference(pl.col("r")))    # keys only in left

# Convert to/from plain List(Struct)
df.select(pl.col("m").map.entries())   # strip Map -> List(Struct)

# from_entries is the inverse: it wraps a raw List(Struct) column into a Map
entries = pl.Series(
    "e",
    [[{"key": "a", "value": 1}, {"key": "a", "value": 2}]],
    dtype=pl.List(pl.Struct({"key": pl.String, "value": pl.Int64})),
)
pl.DataFrame([entries]).select(pl.col("e").map.from_entries())  # Map, deduped to {"a": 1}

# Series iteration yields Python dicts
for d in ser.map:
    print(d)  # {"a": 1, "b": 2}, {"x": 10}

# Arrow table with map column → Polars DataFrame
table = pa.table({"m": pa.array([[("a", 1)]], type=pa.map_(pa.string(), pa.int64()))})
df = from_arrow(table)          # Map(String, Int64) dtype preserved
table2 = to_arrow(df)           # roundtrips back to arrow map<>

# Lazy scanning from an Arrow source
lf = scan_arrow(lambda: [table])
result = lf.collect()

Caveats

• Extension types — used to wrap the underlying List(Struct) storage with a semantic Map dtype, are not yet stabilized and may change across Polars releases.
• pl.dtype_of — used to efficiently cast to the extension type after some operations is also unstable.
• GIL - is required to automatically wrap an expression as the extension type, and so operations which could change the underlying key or value types will briefly lock the GIL to do the cast. This may also prevent the polars engine from reasoning about the type.
• Large offsets — Arrow's map<> type uses only 32-bit offsets, so exporting a Polars map backed by a LargeList whose offsets don't fit in a u32 will error. Arrow has no large-offset map type.