fix: Correctly evaluate pl.len() with lazy validation in plan

src/polars_plugin/mod.rs

@@ -3,6 +3,7 @@ mod utils;
 mod validation_error;
 
 use polars::prelude::*;
+use polars_arrow::bitmap::Bitmap;
 use polars_core::POOL;
 use pyo3_polars::derive::polars_expr;
 use rayon::iter::{IntoParallelRefIterator, ParallelIterator};
@@ -86,13 +87,12 @@ pub fn all_rules_required(
     let failures = compute_rule_failures(rule_inputs, kwargs.null_is_valid)?;
 
     // If there's any failure, we know that validation failed and use the failure object for an
-    // informative error message. If no failure exists, we simply return a series with a single
-    // boolean value to filter by. When filtering on a series with a single value of `true`, polars
-    // neither actually runs the filter logic, nor does it copy any data. It's essentially a no-op
-    // that is not optimized away in a lazy frame.
+    // informative error message. If no failure exists, we return a bitmap with the "unset bits"
+    // flags set to 0 to trigger a fast-path in the filter executed by the caller.
     if failures.is_empty() {
-        let column = Column::new_scalar(PlSmallStr::EMPTY, Scalar::from(true), 1);
-        return Ok(column.take_materialized_series());
+        let bitmap = Bitmap::new_with_value(true, inputs[0].len());
+        let ca = BooleanChunked::from_bitmap(PlSmallStr::EMPTY, bitmap);
+        return Ok(ca.into_series());
     }
 
     // Aggregate failures into a validation error

tests/benches/test_schema.py

@@ -33,8 +33,14 @@ class RowWiseValidationSchema(dy.Schema):
 
 
 @pytest.mark.benchmark(group="schema-row-wise")
-def test_row_wise_validate(benchmark: BenchmarkFixture, dataset: pl.DataFrame) -> None:
-    benchmark(RowWiseValidationSchema.validate, dataset)
+@pytest.mark.parametrize("eager", [True, False])
+def test_row_wise_validate(
+    benchmark: BenchmarkFixture, dataset: pl.DataFrame, eager: bool
+) -> None:
+    def benchmark_fn() -> None:
+        RowWiseValidationSchema.validate(dataset, eager=eager).lazy().collect()
+
+    benchmark(benchmark_fn)
 
 
 @pytest.mark.benchmark(group="schema-row-wise")

tests/schema/test_validate.py

@@ -1,6 +1,8 @@
 # Copyright (c) QuantCo 2025-2026
 # SPDX-License-Identifier: BSD-3-Clause
 
+from pathlib import Path
+
 import polars as pl
 import polars.exceptions as plexc
 import pytest
@@ -319,6 +321,25 @@ def test_multiple_unique_columns_both_invalid(
     assert not MultiUniqueSchema.is_valid(df)
 
 
+def test_lazy_validation_scan_parquet_length(tmp_path: Path) -> None:
+    # Arrange
+    schema = create_schema("test", {"a": dy.Int64(), "b": dy.String()})
+    df = schema.sample(10)
+    df.write_parquet(tmp_path / "test.parquet")
+
+    # Act
+    height = (
+        pl.scan_parquet(tmp_path / "test.parquet")
+        .pipe(schema.validate, eager=False)
+        .select(pl.len())
+        .collect()
+        .item()
+    )
+
+    # Assert
+    assert height == 10
+
+
 # ----------------------------------- PERFORMANCE ------------------------------------ #