Stabilize tests with seeded randomness and retry logic

src/psd/log_analyzer.py

@@ -4,6 +4,8 @@
 from dataclasses import dataclass
 from pathlib import Path
 
+from .utils import retry
+
 
 @dataclass
 class LogStats:
@@ -14,13 +16,16 @@ class LogStats:
     latency_count: int
 
 
+@retry(OSError, tries=3, delay=0.01)
 def analyze_log(path: str | Path) -> LogStats:
     """Summarise errors and latency metrics in a structured log file.
 
     Parameters
     ----------
     path:
-        Path to a log file containing one JSON object per line.
+        Path to a log file containing one JSON object per line.  The function
+        tolerates transient ``OSError`` issues (for example when a log file is
+        being rotated) by retrying with a short exponential backoff.
     """
 
     p = Path(path)

tests/conftest.py

@@ -0,0 +1,36 @@
+"""Pytest configuration for deterministic test runs.
+
+This file seeds all common sources of randomness before each test so that
+results are reproducible across runs and on CI.  Hypothesis based property
+tests are additionally configured with ``derandomize=True`` in the individual
+modules.
+"""
+
+from __future__ import annotations
+
+import os
+import random
+
+import numpy as np
+import pytest
+
+try:  # pragma: no cover - optional dependency
+    import torch
+except Exception:  # pragma: no cover - torch might not be installed
+    torch = None  # type: ignore[assignment]
+
+
+@pytest.fixture(autouse=True)
+def _seed_everything() -> None:
+    """Seed RNGs for ``random``, ``numpy`` and ``torch`` if available.
+
+    Seeding happens automatically for every test via the ``autouse`` fixture
+    mechanism.  ``PYTHONHASHSEED`` is also set to ensure deterministic hashing
+    behaviour in dictionaries and other hash based collections.
+    """
+
+    os.environ.setdefault("PYTHONHASHSEED", "0")
+    random.seed(0)
+    np.random.seed(0)
+    if torch is not None:
+        torch.manual_seed(0)

tests/test_algorithms_property.py

@@ -9,7 +9,7 @@
 from psd.config import PSDConfig
 
 
-@settings(max_examples=50, deadline=None)
+@settings(max_examples=50, deadline=None, derandomize=True)
 @given(
     st.lists(
         st.floats(min_value=-10, max_value=10, allow_nan=False, allow_infinity=False),

tests/test_functions_numerical.py

@@ -41,7 +41,7 @@ def _finite_diff_hess(g: callable, x: np.ndarray, eps: float = 1e-6) -> np.ndarr
     return hess
 
 
-@settings(max_examples=25, deadline=None)
+@settings(max_examples=25, deadline=None, derandomize=True)
 @given(_vector_1d)
 @pytest.mark.fast
 def test_separable_quartic_grad_hess_match(x: np.ndarray) -> None:
@@ -54,7 +54,7 @@ def test_separable_quartic_grad_hess_match(x: np.ndarray) -> None:
     np.testing.assert_allclose(h(x), num_hess, rtol=1e-4, atol=1e-4)
 
 
-@settings(max_examples=20, deadline=None)
+@settings(max_examples=20, deadline=None, derandomize=True)
 @given(_vector_1d_ge2)
 @pytest.mark.slow
 def test_rosenbrock_grad_hess_match(x: np.ndarray) -> None:
@@ -67,7 +67,7 @@ def test_rosenbrock_grad_hess_match(x: np.ndarray) -> None:
     np.testing.assert_allclose(h(x), num_hess, rtol=1e-4, atol=1e-4)
 
 
-@settings(max_examples=20, deadline=None)
+@settings(max_examples=20, deadline=None, derandomize=True)
 @given(_vector_1d, st.integers(min_value=0, max_value=2**32 - 1))
 @pytest.mark.fast
 def test_random_quadratic_determinism_and_derivatives(x: np.ndarray, seed: int) -> None:
@@ -94,7 +94,7 @@ def g(z: np.ndarray) -> np.ndarray:
     np.testing.assert_allclose(functions.random_quadratic_hess(A1), num_hess, rtol=1e-5, atol=1e-6)
 
 
-@settings(max_examples=5, deadline=None)
+@settings(max_examples=5, deadline=None, derandomize=True)
 @given(st.integers(min_value=0, max_value=2**32 - 1))
 @pytest.mark.slow
 def test_psd_deterministic_given_seed(seed: int) -> None:

tests/test_graph_properties.py

@@ -44,7 +44,7 @@ def dfs(node: Hashable, weight: float, path: list[Hashable]) -> None:
     return best
 
 
-@settings(max_examples=50, deadline=None)
+@settings(max_examples=50, deadline=None, derandomize=True)
 @given(dag_graphs())
 def test_find_optimal_path_matches_bruteforce(data: tuple[Graph, str, str]) -> None:
     graph, start, end = data

tests/test_log_analyzer.py

@@ -20,3 +20,25 @@ def test_analyze_log(tmp_path: Path) -> None:
     assert stats.error_count == 1
     assert stats.latency_count == 2
     assert stats.avg_latency == pytest.approx(0.2)
+
+
+def test_analyze_log_retries(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    """Ensure transient ``OSError`` during file access are retried."""
+
+    log_file = tmp_path / "psd.log"
+    log_file.write_text("{}\n")
+
+    calls = {"count": 0}
+    path_cls = type(log_file)
+    real_open = path_cls.open
+
+    def flaky_open(self: Path, *args, **kwargs):
+        if self == log_file and calls["count"] == 0:
+            calls["count"] += 1
+            raise OSError("transient")
+        return real_open(self, *args, **kwargs)
+
+    monkeypatch.setattr(path_cls, "open", flaky_open)
+    stats = analyze_log(log_file)
+    assert stats.error_count == 0
+    assert calls["count"] == 1