Fix pathologically slow assertion diffs for large inputs (#8998)

AUTHORS

@@ -267,6 +267,7 @@ Kevin J. Foley
 Kian Eliasi
 Kian-Meng Ang
 Kim Soo
+Kiril Klein
 Kodi B. Arfer
 Kojo Idrissa
 Kostis Anagnostopoulos

changelog/8998.bugfix.rst

@@ -0,0 +1,3 @@
+Assertion failures comparing very large strings, lists, or dataclasses no longer hang for a long time (sometimes minutes) while building the diff.
+
+When the inputs are large enough that :func:`difflib.ndiff` would be pathologically slow, pytest now runs it over a bounded prefix of the input instead, so the detailed (character-level) diff is kept for the part shown while the rest is truncated with a note.

src/_pytest/assertion/_compare_sequence.py

@@ -6,6 +6,8 @@
 
 from _pytest._io.pprint import PrettyPrinter
 from _pytest._io.saferepr import saferepr
+from _pytest.assertion._diff import ndiff_too_slow_for_lines
+from _pytest.assertion._diff import truncated_ndiff
 from _pytest.assertion._typing import _HighlightFunc
 from _pytest.compat import running_on_ci
 
@@ -27,6 +29,9 @@ def _compare_eq_iterable(
 
     yield ""
     yield "Full diff:"
+    if ndiff_too_slow_for_lines(left_formatting, right_formatting):
+        yield from truncated_ndiff(left_formatting, right_formatting, highlighter)
+        return
     # "right" is the expected base against which we compare "left",
     # see https://github.com/pytest-dev/pytest/issues/3333
     yield from highlighter(

src/_pytest/assertion/_diff.py

@@ -0,0 +1,94 @@
+from __future__ import annotations
+
+from collections.abc import Iterator
+from collections.abc import Sequence
+from itertools import chain
+
+from _pytest.assertion._typing import _HighlightFunc
+
+
+# Past these limits ``difflib.ndiff`` becomes pathologically slow: its
+# character-level "fancy replace" step compares every pair of similar lines in a
+# differing block, so its cost grows with the *product* of the line count and
+# the character count. A few hundred similar lines can already take seconds, and
+# the pretty-printed form of a large list/dataclass takes minutes (see issue
+# #8998). The limits below keep ``ndiff`` under roughly a second in the worst
+# case. Above them we still run ``ndiff`` -- so the detailed diff is kept -- but
+# only over a bounded prefix of the input.
+NDIFF_MAX_INPUT_SIZE = 10_000  # characters (left + right)
+DIFF_MAX_LINES = 100  # lines (left + right)
+
+
+def ndiff_too_slow_for_text(left: str, right: str) -> bool:
+    """Whether ``ndiff`` would be pathologically slow for these strings.
+
+    Counts line separators instead of splitting into lines, so the check stays
+    cheap even for huge inputs.
+    """
+    if left.count("\n") + right.count("\n") > DIFF_MAX_LINES:
+        return True
+    return len(left) + len(right) > NDIFF_MAX_INPUT_SIZE
+
+
+def ndiff_too_slow_for_lines(
+    left_lines: Sequence[str], right_lines: Sequence[str]
+) -> bool:
+    """Whether ``ndiff`` would be pathologically slow for these lines.
+
+    Exits as soon as a limit is exceeded instead of measuring the whole input.
+    """
+    if len(left_lines) + len(right_lines) > DIFF_MAX_LINES:
+        return True
+    size = 0
+    for line in chain(left_lines, right_lines):
+        size += len(line)
+        if size > NDIFF_MAX_INPUT_SIZE:
+            return True
+    return False
+
+
+def truncated_ndiff(
+    left_lines: Sequence[str],
+    right_lines: Sequence[str],
+    highlighter: _HighlightFunc,
+) -> Iterator[str]:
+    """Yield an ``ndiff`` over a bounded prefix of the input (issue #8998).
+
+    The character-level diff is kept, but only for a slice small enough to
+    compute quickly; the rest of the input is dropped.
+    """
+    from difflib import ndiff
+
+    left = _bounded_prefix(left_lines, DIFF_MAX_LINES // 2, NDIFF_MAX_INPUT_SIZE // 2)
+    right = _bounded_prefix(right_lines, DIFF_MAX_LINES // 2, NDIFF_MAX_INPUT_SIZE // 2)
+    yield (
+        f"Diff too large to show in full (over {NDIFF_MAX_INPUT_SIZE} characters "
+        f"or {DIFF_MAX_LINES} lines); showing a truncated diff:"
+    )
+    # "right" is the expected base against which we compare "left",
+    # see https://github.com/pytest-dev/pytest/issues/3333
+    yield from highlighter(
+        "\n".join(line.rstrip("\n") for line in ndiff(right, left)),
+        lexer="diff",
+    ).splitlines()
+
+
+def _bounded_prefix(lines: Sequence[str], max_lines: int, max_chars: int) -> list[str]:
+    """Return the longest prefix of ``lines`` within both limits.
+
+    The line that would cross the character limit is included truncated, so a
+    single huge line still yields some (bounded) output.
+    """
+    kept: list[str] = []
+    chars = 0
+    for line in lines:
+        if len(kept) >= max_lines:
+            break
+        room = max_chars - chars
+        if len(line) > room:
+            if room > 0:
+                kept.append(line[:room])
+            break
+        kept.append(line)
+        chars += len(line)
+    return kept

src/_pytest/assertion/compare_text.py

@@ -3,6 +3,8 @@
 from collections.abc import Iterator
 
 from _pytest._io.saferepr import saferepr
+from _pytest.assertion._diff import ndiff_too_slow_for_text
+from _pytest.assertion._diff import truncated_ndiff
 from _pytest.assertion._typing import _AssertionTextDiffStyle
 from _pytest.assertion._typing import _HighlightFunc
 from _pytest.assertion.highlight import dummy_highlighter
@@ -75,6 +77,11 @@ def _diff_text(
         left = repr(str(left))
         right = repr(str(right))
         yield "Strings contain only whitespace, escaping them using repr()"
+    if ndiff_too_slow_for_text(left, right):
+        yield from truncated_ndiff(
+            left.splitlines(keepends), right.splitlines(keepends), highlighter
+        )
+        return
     # "right" is the expected base against which we compare "left",
     # see https://github.com/pytest-dev/pytest/issues/3333
     yield from highlighter(

testing/test_assertion.py

@@ -14,9 +14,12 @@
 
 from _pytest import outcomes
 import _pytest.assertion as plugin
+from _pytest.assertion import _diff
 from _pytest.assertion import truncate
 from _pytest.assertion import util
 from _pytest.assertion._compare_any import _compare_eq_cls
+from _pytest.assertion._diff import ndiff_too_slow_for_lines
+from _pytest.assertion._diff import ndiff_too_slow_for_text
 from _pytest.assertion.compare_text import _compare_eq_text
 from _pytest.config import Config as _Config
 from _pytest.monkeypatch import MonkeyPatch
@@ -459,6 +462,45 @@ def callequal(
     )
 
 
+class TestNdiffTooSlow:
+    """Heuristic guarding against pathologically slow diffs (#8998)."""
+
+    def test_small_input_is_not_too_slow(self) -> None:
+        assert ndiff_too_slow_for_text("spam", "eggs") is False
+        assert ndiff_too_slow_for_lines(["spam"], ["eggs"]) is False
+
+    def test_too_many_characters(self, monkeypatch: MonkeyPatch) -> None:
+        monkeypatch.setattr(_diff, "NDIFF_MAX_INPUT_SIZE", 5)
+        assert ndiff_too_slow_for_text("abc", "abcd") is True
+        assert ndiff_too_slow_for_lines(["abc"], ["abcd"]) is True
+
+    def test_too_many_lines(self, monkeypatch: MonkeyPatch) -> None:
+        monkeypatch.setattr(_diff, "DIFF_MAX_LINES", 3)
+        assert ndiff_too_slow_for_text("a\nb\nc\nd\ne", "f") is True
+        assert ndiff_too_slow_for_lines(["a", "b", "c", "d"], ["e"]) is True
+
+    def test_bounded_prefix(self) -> None:
+        # All lines fit within both limits: everything is kept.
+        assert _diff._bounded_prefix(["a", "b"], max_lines=10, max_chars=100) == [
+            "a",
+            "b",
+        ]
+        # The line limit stops collection.
+        assert _diff._bounded_prefix(["a", "b", "c"], max_lines=2, max_chars=100) == [
+            "a",
+            "b",
+        ]
+        # The line crossing the character limit is kept truncated.
+        assert _diff._bounded_prefix(["abc", "defgh"], max_lines=10, max_chars=4) == [
+            "abc",
+            "d",
+        ]
+        # When the character limit is exactly full, the next line is dropped.
+        assert _diff._bounded_prefix(["abcd", "e"], max_lines=10, max_chars=4) == [
+            "abcd"
+        ]
+
+
 class TestAssert_reprcompare:
     def test_different_types(self) -> None:
         assert callequal([0, 1], "foo") is None
@@ -513,6 +555,34 @@ def test_text_skipping_verbose(self) -> None:
         assert "- " + "a" * 50 + "eggs" in lines
         assert "+ " + "a" * 50 + "spam" in lines
 
+    def test_text_diff_large_input_is_truncated(self, monkeypatch: MonkeyPatch) -> None:
+        # Inputs over the character limit show a fast, truncated diff instead
+        # of the pathologically slow full ndiff (#8998).
+        monkeypatch.setattr(_diff, "NDIFF_MAX_INPUT_SIZE", 40)
+        left = "the answer is 41\nand a tail" + "x" * 100
+        right = "the answer is 42\nand a tail" + "x" * 100
+        lines = callequal(left, right, verbose=1)
+        assert lines is not None
+        assert any("Diff too large to show in full" in line for line in lines)
+        # ndiff is still used, so the character-level detail is kept.
+        assert any(line.startswith("? ") for line in lines)
+
+    def test_text_diff_many_lines_is_truncated(self, monkeypatch: MonkeyPatch) -> None:
+        # Inputs over the line limit are diffed over a bounded prefix only, so
+        # far fewer than all the lines are shown (#8998).
+        monkeypatch.setattr(_diff, "DIFF_MAX_LINES", 4)
+        left = "\n".join(f"left line {i}" for i in range(50))
+        right = "\n".join(f"right line {i}" for i in range(50))
+        lines = callequal(left, right, verbose=1)
+        assert lines is not None
+        assert any("Diff too large to show in full" in line for line in lines)
+        # The fallback still shows which of the first lines differ.
+        assert "- right line 0" in lines
+        assert "+ left line 0" in lines
+        # Only a bounded prefix is diffed, not all 50 lines.
+        differing = [line for line in lines if line.startswith(("- ", "+ "))]
+        assert 0 < len(differing) < 50
+
     def test_multiline_text_diff(self) -> None:
         left = "foo\nspam\nbar"
         right = "foo\neggs\nbar"
@@ -673,6 +743,20 @@ def test_iterable_quiet(self) -> None:
             "Use -v to get more diff",
         ]
 
+    def test_iterable_large_input_is_truncated(self, monkeypatch: MonkeyPatch) -> None:
+        # Large iterables show a truncated diff over a bounded prefix of their
+        # pprint output instead of the pathologically slow full ndiff (#8998).
+        monkeypatch.setattr(_diff, "DIFF_MAX_LINES", 6)
+        left = [f"item-{i}" for i in range(50)]
+        right = [f"other-{i}" for i in range(50)]
+        lines = callequal(left, right, verbose=1)
+        assert lines is not None
+        assert "Full diff:" in lines
+        assert any("Diff too large to show in full" in line for line in lines)
+        # Only a bounded prefix is diffed, not all 50+ pprint lines.
+        differing = [line for line in lines if line.startswith(("- ", "+ "))]
+        assert 0 < len(differing) < 50
+
     def test_iterable_full_diff_ci(
         self, monkeypatch: MonkeyPatch, pytester: Pytester
     ) -> None: