pandera.backends.narwhals.base.NarwhalsSchemaBackend

class pandera.backends.narwhals.base.NarwhalsSchemaBackend

Base schema backend for Narwhals-backed DataFrames.

Provides shared helpers used by ColumnBackend (components.py) and DataFrameSchemaBackend (container.py).

Methods

drop_invalid_rows(check_obj, error_handler)

Remove invalid rows — pure Narwhals, no backend delegation.

Builds a pass-mask boolean column per check_output, combines with nw.all_horizontal, filters, and drops the temporary columns. Works identically for polars lazy frames and ibis tables.

Two check_output conventions are handled: - nw.Expr (DATAFRAME_CHECK path): True=row passes check - nw.LazyFrame/DataFrame with CHECK_OUTPUT_KEY=True meaning “failed”

(SERIES_CONTAINS_NULLS / check_nullable): True=row has null (failing). Reconstructed as ~col.is_null() from err.schema.selector.

Parameters:

• check_obj – The frame to filter.

• error_handler – ErrorHandler whose schema_errors carry check_output.

Returns:

Filtered frame with only rows where all checks passed.

failure_cases_metadata(schema_name, schema_errors)

Create failure cases metadata required for SchemaErrors exception.

Backend-agnostic: returns native ibis.Table for ibis inputs and pl.LazyFrame/pl.DataFrame for polars inputs — no forced polars conversion, no Arrow roundtrip for lazy/SQL backends.

Return type:

FailureCaseMetadata

is_float_dtype(check_obj, col_name)

Return True if the column col_name has a float dtype.

Uses collect_schema() so it works on both LazyFrame and DataFrame without triggering full materialization.

Parameters:

• check_obj – Narwhals LazyFrame or DataFrame.

• col_name (str) – Name of the column to inspect.

Return type:

bool

Returns:

True if the column dtype is a floating-point type.

run_check(check_obj, schema, check, check_index, *args)

Execute a single Check object and return a CoreCheckResult.

Single unified code path — no _is_ibis_result bifurcation. Materializes only the scalar passed bool via _materialize(check_passed). failure_cases and check_output stay as Narwhals wrappers in the returned CoreCheckResult; callers (failure_cases_metadata) materialize as needed.

subsample(check_obj, head=None, tail=None, sample=None, random_state=None)

Return a (possibly subsampled) version of check_obj.

Never materializes check_obj — delegates directly to .head()/.tail() so the result stays lazy (nw.LazyFrame) for Polars inputs.

Parameters:

• head (UnionType[int, None]) – Number of rows to take from the head.

• tail (UnionType[int, None]) – Number of rows to take from the tail.

• sample (UnionType[int, None]) – Not supported — raises NotImplementedError.

• random_state (UnionType[int, None]) – Ignored (no random sampling supported).

Raises:

NotImplementedError – If sample is not None, or if tail= is requested on a SQL-lazy backend (ibis.Table) that does not support TAIL without forced full ordering.