Support and dependency policy

Platforms

pstrain supports macOS, Linux, and Windows. CI builds wheels on all three platforms, and release builds publish those wheels to PyPI.

Windows support does not extend to training. Running any stage of the training pipeline on Windows raises RuntimeError: pipeline provenance locking requires the POSIX fcntl module; pipeline execution is unavailable on this platform, because provenance replacement locks its critical section with fcntl.flock and Windows has no fcntl. What does work there is the package itself, the native library and command-line programs, and the native worker that carries requests to them. Training a model requires macOS or Linux.

On Windows, CI builds the native library and command-line programs with both MSVC and clang-cl, then runs a Windows-compatible subset of the Python tests against an MSVC build on Python 3.13. The full Python test suite is run on macOS and Linux; Windows CI does not currently claim that broader coverage.

Windows Python coverage

This has been measured. The Windows Python 3.13 job requested 146 test units and reported 101 passed, 45 skipped. Those 45 are the 39 cases and the six module-level collection skips below, and 101 plus 45 is the job’s 146, so the breakdown here and the CI summary line describe the same run.

Six POSIX resource-accounting modules skip during collection, leaving 140 collected items. Of those 140, 39 skip and 101 pass:

  • Six whole modules skip because POSIX training resource accounting uses resource: test_bw_sharding.py, test_dtree.py, test_e2e_training.py, test_feat_params.py, test_train_convergence.py, and test_train_retry.py.

  • 39 cases skip because pipeline provenance replacement locks its critical section with fcntl.flock, which Windows does not have. In tests/test_numeric_harness.py that is every case built on the flat_project or full_project fixtures, since creating either one runs a pipeline (30 cases). In tests/test_pipeline_tasks.py it is test_meaningful_feature_config_change_rebuilds_features, test_reverting_feature_config_rebuilds_features, test_irrelevant_config_edit_does_not_rebuild_features, test_model_and_package_copy_build_provenance, test_split_runs_end_to_end_and_partitions, test_lm_target_succeeds_on_setup_project_layout, test_editing_persistent_split_revalidates_and_changes_membership, test_configured_bw_parameters_reach_training_call, and test_configured_untied_schedule_and_variance_reach_training_call.

  • The remaining 101 cases pass: everything selected from tests/test_paths.py, tests/test_commands.py, tests/test_cffi_abi.py, tests/test_lib_structure.py, tests/test_pstrainc.py and tests/test_features.py; all of tests/test_pipeline_tasks.py outside the nine provenance-lock cases above; and the four cases in tests/test_numeric_harness.py that do not build a pipeline fixture.

No case is skipped any longer for native-worker request transport, but that buys less than the count suggests, so it is worth stating what it does buy. Sixty-nine cases carried the transport skip. Thirty-nine of them now skip for the provenance lock instead, and thirty run. Twenty-seven of those thirty never reach the native worker at all: they build and inspect pipeline plans, or check provenance and fingerprint payloads (the one that executes a task, test_linear_default_untied_stage_builds_occurrence_inventory, reaches the C library in this process rather than through the worker). The three that genuinely drive the worker are the test_m4_real_utterances_reach_shared_final_state cases, whose per-utterance feature payload is larger than the pipe buffer and so takes the pending-write path rather than completing at once. One further case, tests/test_native_worker.py::test_windows_write_request_gives_up_at_the_deadline_on_a_full_pipe, has since been added to the Windows selection to cover the request deadline directly; the job now requests 147 units, and that case is not part of the measurement above.

The transport skip existed because the transport did not work on Windows at all: the spawned child reported ready, and the parent then failed before sending its first request with OSError: [Errno 9] Bad file descriptor, raised by os.set_blocking() on the connection handle. On Windows a multiprocessing duplex pipe is an overlapped, message-mode named pipe and Connection.fileno() returns a Win32 HANDLE, which os.set_blocking, select.select and os.write all reject: they take a C-runtime file descriptor or a socket. The request write is now expressed in the terms each platform actually offers – a non-blocking descriptor and select on POSIX, a bounded wait on an overlapped completion event on Windows – and the deadline that a full pipe must not outlive is preserved on both.

What that fix moved is where Windows stops, not whether it stops. The wall is now fcntl, one layer past the transport, and no amount of further work on the request pipe reaches it.

Dependencies

Runtime dependencies declare tested minimum versions and may float within their compatible major releases; lock files are intentionally not used for library consumers. Development, documentation, and CI tools are constrained in pyproject.toml, while GitHub Actions and pre-commit hook revisions are pinned in their workflow files and advanced through reviewed dependency updates. Security and compatibility fixes may raise a minimum version; unnecessary runtime dependencies should not be added.