Do not push down inf/nan float literals in the polars lazy filter

duckdb/polars_io.py

@@ -313,6 +313,10 @@ def _pl_tree_to_sql(tree: _ExpressionTree) -> str:
             "Float64",
             "Boolean",
         ):
+            if value[dtype] is None:
+                # JSON has no inf/nan, polars serializes them as null
+                msg = f"Non-finite {dtype} literals cannot be pushed down"
+                raise NotImplementedError(msg)
             return str(value[dtype])
 
         # Time type

tests/fast/arrow/test_polars.py

@@ -264,6 +264,23 @@ def test_polars_lazy_pushdown_numeric(self, data_type, duckdb_cursor):
         valid_filter((pl.col("a") == 100) & (pl.col("b") == 10) & (pl.col("c") == 100))
         valid_filter((pl.col("a") == 100) | (pl.col("b") == 1))
 
+    @pytest.mark.parametrize("data_type", ["FLOAT", "DOUBLE"])
+    def test_polars_lazy_pushdown_non_finite_float(self, data_type, duckdb_cursor):
+        duckdb_cursor.execute(f"CREATE TABLE test_non_finite (a {data_type})")
+        duckdb_cursor.execute("INSERT INTO test_non_finite VALUES (1), ('inf'), ('-inf'), ('nan'), (NULL)")
+        lazy_df = duckdb_cursor.table("test_non_finite").pl(lazy=True)
+        eager_df = lazy_df.collect()
+
+        # polars serializes inf/nan literals as JSON null, so they cannot be pushed down
+        for predicate in (
+            pl.col("a") < float("inf"),
+            pl.col("a") > float("-inf"),
+            pl.col("a") == float("inf"),
+            pl.col("a") >= float("nan"),
+        ):
+            invalid_filter(predicate)
+            pl_testing.assert_frame_equal(lazy_df.filter(predicate).collect(), eager_df.filter(predicate))
+
     def test_polars_lazy_pushdown_bool(self, duckdb_cursor):
         duckdb_cursor.execute(
             """