[tae898]

Native parity on every supported Python, cheaply, in the selective-matrix design of test-native.yml. The native wheel is py3-none (a ctypes loader, no per-interpreter extension), so one built wheel can be installed into every Python without a rebuild. This is the redo of humemai/arcadedb-embedded-python#540 for this repository (that PR stays as a reference).

What changes

scripts/select_native_matrix.py gets a fourth output, pyver=<json>: the supported Pythons that the full parity run of the event leaves out, plus the free-threaded 3.14t. test-native.yml gets a pyver job that runs those legs. So nothing is tested twice and the load stays small:

event	full parity (test job, every platform)	pyver legs (linux-x86_64 wheel, subset)
pull request	3.12	3.11, 3.13, 3.14, 3.14t
push to main, nightly, dispatch	3.12, 3.14	3.11, 3.13, 3.14t
release (release-native.yml calls the workflow with release: true)	3.11, 3.12, 3.13, 3.14	3.14t

The pyver job has needs: [plan, build], runs on ubuntu-latest with fail-fast: false and a 20 minute timeout, downloads the native-wheel-linux-x86_64 artifact of the build job (no rebuild) and runs uv run --no-project --python <ver> --with <wheel> ... native/wheel/parity_wheel.py --files <subset>. The gate is a copy of the test job's awk (no failure, error, crash, timeout or uncounted skip; both copies carry a keep-in-sync comment). Actions are pinned by the SHAs the file already uses. A step prints sys._is_gil_enabled() after loading the wheel and numpy, pandas, pyarrow and jpype1, and warns on the 3.14t leg if the GIL got re-enabled.

It is informational: release-native.yml (verify and publish) reads only the native-passed-* markers and the native-wheel-* wheels and no job waits for pyver (a unit test pins both), and the workflow header and the job comment say so. A red leg is visible but never blocks a release. Promote it once it has been green for a while.

The subset (13 files, 14 where polars installs)

test_core, test_native_udf (ctypes layer, database lifecycle, callbacks), test_resultset (queries and results), test_parameter_binding, test_type_conversion, test_transaction_config (transactions), test_async_executor, test_concurrency and test_free_threading (threads; the second file is written for a free-threaded CPython and also passes with the GIL), test_vector, test_numpy_support, test_resultset_arrow, test_columnar_readers (numpy, pandas, arrow), and test_dataframe_ingest (pandas, polars and arrow ingest) on the legs that have polars. The files are self-contained, fast, and fail early on an interpreter-specific problem. The 70-file suite stays in the test job.

3.14t: included, without polars

All dependencies of the subset have cp314t wheels (numpy 2.5.4, pandas 3.1.0, pyarrow 26.0.0, jpype1 1.7.1, requests, pytest), the wheel and its ctypes layer load and run on the free-threaded build, and the GIL stays off after the wheel and those packages are loaded, so the leg tests free threading for real. The one exception is polars: polars and polars-runtime-32 ship abi3 wheels only, which cannot be installed on a free-threaded build, and test_dataframe_ingest.py imports polars at module level (pytest.importorskip; this repository accepts no skips). The 3.14t leg therefore installs no polars and leaves that file out; the matrix carries polars: false for it, and a unit test pins polars == not freethreaded. If you prefer the strict reading (any missing cp314t wheel means no 3.14t leg), drop FREE_THREADED_PYTHONS in the script.

3.14t is not a classifier of the wheel, so ALL_PYTHONS (checked against native/wheel/pyproject.toml) is unchanged, and I did not add a "Free Threading" trove classifier (a support claim; the optional polars extra cannot be installed there).

Merge order: this lands after #69

tests/test_async_executor.py::test_per_record_callback_runs_before_the_batch_commits_and_on_ok_after aborts the whole process (Fatal Python error: Aborted, in wait_completion through lib.call) against every native image built before the glue of #69 (the nbatch image, the two-dist image and the whl-pre wheel alike). It is the same on 3.11, 3.12, 3.13, 3.14 and 3.14t, so it is not a Python-version problem: it is the glue/image mismatch that #69 fixes. I kept test_async_executor.py in the subset (async is part of what this checks), so the pyver legs show test_async_executor.py(CRASH) and go red on an image without #69. Therefore #68 lands after #69; the other 12 or 13 files pass on every Python.

Local results

Rebased onto native-two-dist after the workflow rename (native.yml is now test-native.yml, the release jobs live in release-native.yml): the pyver job moved into test-native.yml unchanged in behaviour, and tests/test_select_native_matrix.py, tests/test_select_matrix.py and tests/test_check_version.py pass on the rebased head, as do actionlint and pre-commit on the touched files. The pyver job needs no --with setuptools (no file of the subset builds with setup.py). The measurements below are from before the rebase.

Run from the repository root against the existing nbatch image (ARCADEDB_BACKEND=native, ARCADEDB_NATIVE_LIB=.../nbatch/libarcadedb.so, PYTHONPATH=src, taskset -c 8-15), one pytest process per file, uv run --no-project --python <ver> --with pytest --with numpy --with requests --with jpype1 --with pyarrow --with pandas [--with polars]. All files other than the aborting one passed on every version; test_async_executor.py with that one test deselected passes 20 of 20 on every version.

Python	files passed (of 14, or 13 on 3.14t)	tests passed	skipped	wall (without async)
3.11.15	13 of 14	268	32	25 s
3.12.13 (reference)	13 of 14	268	32	26 s
3.13.13	13 of 14	268	32	25 s
3.14.5	13 of 14	268	32	27 s
3.14.5t	12 of 13	199	32	22 s

The skips are the NOTYET and NA-JVM categories of tests/conftest.py, which the parity runner counts separately. I also replayed the job's run steps verbatim (uv invocation, parity_wheel.py --files, the GIL step, the awk gate) against an already-built wheel (arcadedb-26.11.1.dev0-py3-none-manylinux_2_34_x86_64.whl from another worktree, no rebuild) on 3.14t (199 passed, 0 failures, GIL stays off) and 3.11 (268 passed, 0 failures); both show the gate failing on test_async_executor.py(CRASH) as described above, which also confirms that the gate catches a crash. tests/test_select_native_matrix.py (9 tests, 3 new) passes, actionlint, pre-commit and scripts/check_no_embedded_name.py pass on the touched files. docs/modules/ROOT/pages/development/ci.adoc describes the job.

Not verified

The workflow itself cannot be run locally: the artifact download, the 3.14t interpreter download with the pinned uv 0.12.0 (I used uv 0.13.4 locally), the fromJSON matrix from the plan job, the runner's core count and the real CI timing are untested until the first run on this branch. I expect roughly 3 to 5 minutes per leg, mostly downloads (numpy, pandas and pyarrow are about 80 MB); that is a guess.

🤖 Generated with Claude Code