ijson is an iterative JSON parser for Python, useful for consuming large data feeds without loading an entire document at once. Its captured week contains one commit on master, with three insertions and three deletions. The November 10 runner update moves macOS wheel builds, functional tests, and memory leak tests from macos-14 to macos-15, making build verification the topic for this review.
Wheel builds retain their architecture matrix ¶
The distribution workflow pairs each operating system runner with an architecture list. The macOS pairing now selects macos-15 while keeping ARCHS_MACOS. Its existing base configuration requests x86_64; the full configuration requests x86_64, arm64, and universal2. The edit changes where those jobs run, with no change to the architecture lists themselves.
That distinction matters when assessing the next build. Ordinary activity uses the base matrix, while version tags, branches beginning with fullci_, and the scheduled Saturday run select the full matrix. A successful base build therefore leaves two macOS architecture entries outside that run’s evidence. The workflow passes each selected architecture through CIBW_ARCHS to cibuildwheel, then collects the resulting wheels as artifacts. For maintainers evaluating this runner move, the full build is the useful checkpoint for all three configured macOS targets.
For an ETL team, the practical concern is whether its next dependency installation receives a usable binary. ijson’s documented backends include a YAJL C extension, and its binary wheels spare users the local compilation step. The build host label alone does not establish a new minimum macOS version for those wheels. That requires checking the generated artifacts and their compatibility metadata. Nothing in this three line edit establishes a new deployment floor or a parser performance improvement.
Python coverage stays at six versions ¶
The functional test workflow keeps Python 3.10 through 3.15 in its main matrix. Six macOS combinations move to the new runner. Ubuntu, Ubuntu ARM, and Windows remain in the matrix, alongside the separate Ubuntu PyPy entry. This is continuity in the configured interpreter coverage; the change does not add a Python version.
There is also useful detail in how these jobs build ijson. The workflow checks out submodules and installs the project from source. On macOS it sets IJSON_EMBED_YAJL=1; the ubuntu-latest job instead installs the system YAJL development package. Unit tests and doctests then run against the installed project. The macOS runner move therefore exercises the existing embedded YAJL build path in a different host environment. It does not demonstrate the result of installing every published wheel.
That gives maintainers two separate things to inspect after the change: source build test results and distribution artifacts. A passing source build is useful evidence for that interpreter and runner combination. It cannot stand in for an installation check of a particular wheel on a downstream machine. For teams reproducing ingestion failures on Macs, record the interpreter, installed artifact, and selected backend alongside the input that failed. Those details help distinguish an installation difference from a JSON parsing problem.
Leak checks move with the functional tests ¶
The memory leak workflow makes the same runner substitution. It retains Ubuntu and macOS across Python 3.10 through 3.15, giving twelve configured combinations. The macOS half now uses macos-15. Its source build follows the same embedded YAJL choice as the functional workflow, then invokes pytest --memleaks-only -v.
Keeping these jobs on the same macOS generation makes failures easier to compare. If functional tests moved while leak checks stayed behind, the two suites would report against different operating system environments. This edit avoids that split across the three workflows it touches. It also replaces the previous macOS entry rather than adding a second one, so these matrices no longer provide a direct comparison against macos-14.
For ingestion workers that process many documents in one process, leak checks are relevant even when individual parses return correct data. But the captured activity contains workflow configuration, not memory measurements or completed test results. There is no new leak fix or test assertion in this patch. A useful downstream check would repeat a representative parse in the same process and observe memory across iterations. Treat that as workload validation; this CI edit provides no measured reduction in memory use to apply to capacity planning.
What to watch ¶
- Inspect a full macOS wheel run before concluding that all configured architectures work on the new host. The base matrix covers only
x86_64, and the captured activity includes no artifact results. - Check the next relevant distribution in the environment that actually runs the ingestion job. Successful source tests and successful installation of a published wheel answer different questions.
- Keep upgrade urgency proportional to the patch. This is CI maintenance with no parser source change. Downstream applications have no new API or setting to adopt from this activity alone.