Polars Groups Columns So One Pass Feeds Many Aggregations
Grouping in Polars is a separate stage: keys are encoded (single key, key rows, row encoding, or binview), hashed into group ids by a Grouper, and every aggregation expression then …
Arrow
Polars Evaluates Expressions in an Arena
Polars lowers the user-facing Expr tree into a flat Arena of nodes addressed by integer Node handles, which the physical planner walks to build Arc operators.
Nulls Are Not Values: How DataFusion Skips Work Using Arrow's Validity Bitmaps
Nulls live in a side-channel validity bitmap, not in the value buffer, which lets DataFusion branch once on ‘has any nulls’ and skip whole loops, filters, and comparisons.
Polars Shares Columns Instead of Copying Them: Arc, Bytes and Reference Counting
Column bytes live in refcounted buffers; sharing a column clones handles, and the Arrow FFI imports foreign memory by keeping the C array alive as the buffer’s owner.
Polars Stores Short Strings Inline in 16 Bytes: The View Type
The View type packs a 4-byte length, 4-byte prefix, buffer index and offset into 16 bytes, and inlines any string of 12 bytes or less directly into those bytes.