Skip to content

Research: Python warning root causes (2026-09-21)

  • Status: Complete
  • Scope: six warning-heavy classic-harness test modules
  • Decision: ADR-1278

Question

Which operations generated the Python 3.14/SciPy warning batch, and what changes remove their causes without suppressing warnings, changing Netflix golden-score assertions, or allowing duplicate assets to execute concurrently?

Reproducer and baseline

PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -m pytest -q -W always \
  python/test/cuda_default_model_test.py \
  python/test/executor_test.py \
  python/test/local_explainer_test.py \
  python/test/sycl_motion_parity_test.py \
  python/test/train_test_model_test.py \
  python/test/vmaf_v1_quality_runner_test.py

The isolated-worktree baseline reported 334 warnings: one deprecated scipy.misc import, 330 unsafe-fork warnings from multiprocessing/popen_fork.py, two numpy.exp overflow warnings, and one SciPy covariance OptimizeWarning. The integration checkout's preceding run reported 333; both inventories contained the same four root causes and the same remediation scope.

Findings

The import warning was a dead dependency

tools/sigproc.py imported scipy.misc but never referenced it. Removing the import removes the warning and changes no execution path.

The process warning identified a real unsafe operation

tools/misc.py forced multiprocessing to use fork globally, then created a manager and process pool while the test process had live threads. Python's documentation states that safely forking a multithreaded process is problematic; Python 3.12 and newer may emit a DeprecationWarning, and Python 3.14 no longer defaults to fork on any platform. Filtering that warning would leave the unsafe inherited thread state in place.

Joblib's loky backend starts isolated worker processes and transports local functions with cloudpickle. A keyed grouping layer preserves the executor's second invariant: equal asset strings run sequentially while different groups can run concurrently, and indexed reconstruction retains caller order. FIFO producer helpers use an explicit spawn context because they do not pass arbitrary local callbacks.

The 5PL warning exposed the wrong equation

The implementation was effectively b1 + sigmoid(x) + b4*x + b5. Because both b1 and b5 were additive intercepts, the optimizer could not identify them independently and reported an indeterminate covariance matrix. Equation 3 of Sheikh, Sabir, and Bovik uses b1 as the sigmoid amplitude:

Q(x) = b1 * (expit(b2 * (x - b3)) - 0.5) + b4 * x + b5

scipy.special.expit evaluates the same sigmoid without materializing the overflowing exponential. A synthetic curve test now proves that all five parameters contribute, and predictions at -10000 and 10000 remain finite.

The fatal-warning gate exposed a temporary-file leak

The broader tools_test.py run found that import_python_file(..., override=...) kept its NamedTemporaryFile wrapper open while reopening the path. Python 3.14 reported the implicit garbage-collection cleanup as a ResourceWarning. The override path now closes the wrapper before rewriting the file and removes the temporary module in a finally block, including when the generated module fails to import.

Validation contract

  • importing vmaf.tools.sigproc under -W error succeeds;
  • parallel_map under a live thread emits no warning and retains input order;
  • equal asset keys never overlap even when distinct keys run in parallel;
  • the established 5PL RMSE assertion remains unchanged and passes under warnings-as-errors;
  • synthetic 5PL fitting recovers the reference curve without overflow or covariance warnings;
  • overridden Python-file imports close and remove their temporary file without ResourceWarning;
  • the original six-module census reports zero warnings.

The final CUDA-enabled Python 3.14 run used -W error and completed with 31 passed, 7 skipped, and zero warnings. The seven skips were three unavailable SYCL-device cases and four pre-existing HFR-model ENOTSUP cases.

Regression gate

The root pytest configuration now contains only filterwarnings = ["error"]; there are no ignore entries. The legacy tox configuration previously passed -p no:warnings, disabling pytest's warning plugin completely, so GPU CI runs that select it with -c python/tox.ini bypassed the root policy. That switch is removed and the tox config independently sets filterwarnings = error.

The pytest-asyncio 1.4 configuration warning is fixed by declaring asyncio_default_fixture_loop_scope = "function" in each config that loads the plugin. This makes the intended existing per-test fixture isolation explicit; it is not a warning filter.

Package-local pytest configurations override discovery from the repository root when tests run from their package directories. The AI, dev-LLM, MCP, vmaf-tune, and ROI-score configurations therefore each carry the same fatal warning policy; MCP also carries the explicit asyncio fixture scope.

Primary sources