Agent Skills: Polars Knowledge Patch

Polars Knowledge Patch

Use this skill when writing, reviewing, migrating, or debugging Polars code whose behavior may depend on current Python or Rust APIs. It is especially useful for constructor behavior, expression semantics, lazy queries, streaming execution, SQL, serialization, Arrow interchange, and storage integrations.

How to use this skill

1. Identify whether the task concerns migration, expressions, I/O, query execution, SQL, serialization, or runtime interoperability.
2. Read the matching reference file before proposing API names or defaults.
3. Prefer explicit options where defaults have changed or are deprecated.
4. Treat persisted lazy plans, expressions, UDFs, and provider-bearing objects as compatibility-sensitive artifacts.
5. Preserve the exception specificity documented here when writing recovery code.
6. Verify schema and dtype assumptions before and after nested casts, Arrow conversion, file scans, or database reads.

Reference index

| Reference | Topics | | --- | --- | | Migration and Core API | Constructors, renamed arguments, exceptions, deprecations, core frame and series behavior | | Expressions and Data Types | Nested values, null behavior, temporal operations, categoricals, decimals, and numeric expressions | | I/O, Cloud Storage, and Databases | Parquet, CSV, spreadsheets, Delta, Iceberg, cloud credentials, and database I/O | | Lazy Execution, Streaming, and SQL | Schema planning, optimizer controls, streaming, joins, grouping, sorted merges, and SQL | | Serialization, Runtime, and Arrow | Serialization formats, Python/runtime compatibility, NumPy, Arrow import/export, and interchange |

Breaking changes: constructors and core operations

Make row orientation explicit

Pass orient="row" for heterogeneous row records. Frame construction infers orientation from dimensions and schema shape and warns when it infers rows.

Choose strict Series construction deliberately

Mixed incompatible values raise under strict construction, even when the dtype is inferred. Use strict=False only when common-dtype inference or casting is intended.

Expect datetime values to be converted to the declared zone

Constructing with a zoned datetime dtype converts values to that zone. Do not assume that construction merely replaces timezone metadata; wall-clock values can shift.

Use replace_strict for dtype-changing mappings

replace preserves the existing dtype. Use replace_strict when a mapping may change dtype, and provide default when unmapped non-null values should survive.

Update pivot calls and generated names

Use on instead of columns. It is the first positional argument, and omitted index or values inputs are inferred. Multi-value output names no longer repeat the pivot-column name.

Treat two-dimensional values as fixed-size arrays

Reshaping a series or constructing one from a two-dimensional NumPy array creates an Array, not a List. Convert with .arr.to_list() where list semantics are required.

Handle bounds explicitly

get and gather operations raise for out-of-bounds indices by default. Pass null_on_oob=True only when a missing result is the intended contract.

Collect lazy schemas explicitly

Call collect_schema() before inspecting lazy columns, dtypes, width, or schema. Direct property access can trigger expensive resolution and a performance warning.

Breaking changes: execution and persistence

Separate binary serialization from JSON I/O

Frame and expression serialization defaults to binary. Use byte streams for the default format or request format="json"; use deserialize for serialized frames, not read_json.

Recreate non-serializable runtime state

Credential-provider objects are not embedded in serialized objects. Reattach credentials after loading, and reject incompatible DSL representations rather than assuming persisted expressions or lazy plans remain portable.

Do not rely on implicit map_batches optimization

LazyFrame.map_batches defaults to no optimizer transformations. Enable only the optimizations whose interaction with the callback is known to be safe.

Expect stricter invalid-operation failures

Invalid temporal/non-temporal arithmetic, imploding Object values, incompatible nested casts, bad SQL expressions, and non-numeric integer-range inputs now fail instead of producing invalid or weakly typed results.

Update exception handling

Operations that formerly collapsed failures into broad compute errors may raise specific invalid-operation, schema, duplicate-name, or SQL-syntax exceptions. Catch the narrow type when recovery depends on the failure category.

Deprecations to remove from new code

• Do not rely on dataframe interchange integration for a long-lived boundary.
• Avoid StringCache; prepare existing uses for removal.
• Pass plan_stage to show_graph() and empty_as_null to .explode().
• Remove cache-related arguments from scan_ipc calls.
• Stop supplying ddof to rolling_corr; it is ignored.
• Give struct.rename_fields() exactly one name per field.
• Replace categorical-to-integer and numeric-to-categorical casts with explicit categorical conversion expressions.
• Replace string-to-temporal, scalar-to-List, and mixed integer/Boolean bitwise casts or operators with explicit, typed conversions.
• Do not depend on cat.get_categories() or cat.to_local().

High-value expression and dtype updates

Null handling is configurable or preserved more consistently

Nested contains, any, and all operations expose explicit null controls. Rolling-by results preserve null keys, EWM outputs preserve null positions, and unbiased EWM variance and standard deviation start with null rather than zero.

Nested types support more native operations

Fixed-size arrays can be grouping keys. Lists and arrays support missing-aware comparisons, uniqueness checks, null-aware membership, Boolean reductions, and scalar broadcasting for list.slice.

Decimal and half-precision behavior is stronger

Arrow decimals stay decimal, decimal sums widen their result precision and raise on overflow, and stable Float16 values participate in Parquet reads and group-by aggregation.

Categorical conversion is explicit

Use Expr.cat.to for categorical conversion and Expr.cat.physical for physical representation access. Align Enum categories before append, and remember that lexical categorical order controls extrema where configured.

High-value query and SQL updates

Treat streaming as a supported engine

Streaming covers a wider set of aggregations, grouped as-of joins, interpolation, automatic datetime parsing, covariance and correlation, PyArrow datasets, and Parquet output.

Use centralized optimizer controls

Pass a QueryOptFlags object when customizing lazy optimization. Keep callback semantics in mind when combining flags with map_batches.

Use the top-level SQL context for multiple frames

Frame-local SQL sees only its own frame. Use pl.sql(...) when a query must resolve multiple named frames.

Account for current SQL arithmetic and aggregates

Division uses true-division semantics. Aggregate filters, STRING_AGG, TOTAL, computed grouping keys, implicit joins, and subquery membership predicates are available. All-null SUM and CORR results are null.

Preserve sortedness contracts

Each set_sorted call annotates one independently sorted column. Use merge_sorted for already-sorted frames, including multiple frames or multiple lexicographic keys where applicable.

High-value I/O updates

Be explicit about Hive partition discovery

Directory Parquet reads enable Hive partitioning by default; individual paths, globs, and file lists do not. Pass hive_partitioning=True when partition columns must be recovered outside directory input.

Use current spreadsheet engines and inputs

Excel reads default to Calamine. Select xlsx2csv when engine options are needed. Spreadsheet readers accept raw bytes, can drop empty rows, and can select named Excel tables; Excel writers accept file-like outputs.

Carry cloud and schema settings deliberately

Credential providers, certificate options, endpoint discovery, Parquet cast controls, multi-file CSV inference, explicit schemas, metadata, and field IDs all affect planning or output. Do not assume these settings survive serialization.

Prefer stable sink APIs for large outputs

Use the sink interfaces for Parquet, Delta, Iceberg, callback, and remote output. Configure the out-of-core disk budget when execution can spill to disk.

Review checklist

• Are constructor orientation, strictness, timezone conversion, and byte-scalar broadcasting intentional?
• Are nested dtypes (Array, List, Struct, Enum, Decimal, Float16) handled without relying on legacy coercions?
• Are null and out-of-bounds policies explicit where they affect results?
• Are lazy schema collection and optimizer controls explicit?
• Does SQL code account for frame scope, true division, null aggregates, and syntax validation?
• Does I/O code choose its engine, partition behavior, schema controls, and credential lifecycle deliberately?
• Are serialized artifacts read with the matching serializer and supplied with fresh runtime-only state?
• Are Arrow, NumPy, and database boundaries checked for duplicate names, ordered dictionaries, integer width, and null preservation?