diff --git a/pgtricks/pg_dump_splitsort.py b/pgtricks/pg_dump_splitsort.py
index 0b9eee7..38ec5fc 100755
--- a/pgtricks/pg_dump_splitsort.py
+++ b/pgtricks/pg_dump_splitsort.py
@@ -16,24 +16,23 @@
 MEMORY_UNITS = {"": 1, "k": KIBIBYTE, "m": MEBIBYTE, "g": GIBIBYTE}
 
 
-def try_float(s1: str, s2: str) -> tuple[str, str] | tuple[float, float]:
+def try_float(s1: str, s2: str) -> tuple[float, float]:
     """Convert two strings to floats. Return original ones on conversion error."""
     if not s1 or not s2 or s1[0] not in '0123456789.-' or s2[0] not in '0123456789.-':
         # optimization
-        return s1, s2
-    try:
-        return float(s1), float(s2)
-    except ValueError:
-        return s1, s2
+        raise ValueError
+    return float(s1), float(s2)
 
 
 def linecomp(l1: str, l2: str) -> int:
     p1 = l1.split('\t', 1)
     p2 = l2.split('\t', 1)
     # TODO: unquote cast after support for Python 3.8 is dropped
-    v1, v2 = cast("tuple[float, float]", try_float(p1[0], p2[0]))
-    result = (v1 > v2) - (v1 < v2)
-    # modifying a line to see whether Darker works:
+    try:
+        v1, v2 = try_float(p1[0], p2[0])
+        result = (v1 > v2) - (v1 < v2)
+    except ValueError:
+        result = (p1[0] > p2[0]) - (p1[0] < p2[0])
     if not result and len(p1) == len(p2) == 2:
         return linecomp(p1[1], p2[1])
     return result
diff --git a/pgtricks/tests/test_pg_dump_splitsort.py b/pgtricks/tests/test_pg_dump_splitsort.py
index e38e8d3..e885984 100644
--- a/pgtricks/tests/test_pg_dump_splitsort.py
+++ b/pgtricks/tests/test_pg_dump_splitsort.py
@@ -1,3 +1,4 @@
+from contextlib import nullcontext
 from functools import cmp_to_key
 from textwrap import dedent
 
@@ -36,29 +37,32 @@ def test_sql_copy_regular_expression(test_input, expected):
 @pytest.mark.parametrize(
     's1, s2, expect',
     [
-        ('', '', ('', '')),
-        ('foo', '', ('foo', '')),
-        ('foo', 'bar', ('foo', 'bar')),
+        ('', '', ValueError),
+        ('foo', '', ValueError),
+        ('foo', 'bar', ValueError),
         ('0', '1', (0.0, 1.0)),
-        ('0', 'one', ('0', 'one')),
+        ('0', 'one', ValueError),
         ('0.0', '0.0', (0.0, 0.0)),
-        ('0.0', 'one point zero', ('0.0', 'one point zero')),
+        ('0.0', 'one point zero', ValueError),
         ('0.', '1.', (0.0, 1.0)),
-        ('0.', 'one', ('0.', 'one')),
+        ('0.', 'one', ValueError),
         ('4.2', '0.42', (4.2, 0.42)),
-        ('4.2', 'four point two', ('4.2', 'four point two')),
+        ('4.2', 'four point two', ValueError),
         ('-.42', '-0.042', (-0.42, -0.042)),
-        ('-.42', 'minus something', ('-.42', 'minus something')),
-        (r'\N', r'\N', (r'\N', r'\N')),
-        ('foo', r'\N', ('foo', r'\N')),
-        ('-4.2', r'\N', ('-4.2', r'\N')),
+        ('-.42', 'minus something', ValueError),
+        (r'\N', r'\N', ValueError),
+        ('foo', r'\N', ValueError),
+        ('-4.2', r'\N', ValueError),
     ],
 )
 def test_try_float(s1, s2, expect):
-    result1, result2 = try_float(s1, s2)
-    assert type(result1) is type(expect[0])
-    assert type(result2) is type(expect[1])
-    assert (result1, result2) == expect
+    with pytest.raises(expect) if expect is ValueError else nullcontext():
+
+        result1, result2 = try_float(s1, s2)
+
+        assert type(result1) is type(expect[0])
+        assert type(result2) is type(expect[1])
+        assert (result1, result2) == expect
 
 
 @pytest.mark.parametrize(