Testing
Running the suite
pip install -e ".[dev]"
pytest tests/ # everything
pytest tests/unit/ # object model and emitted script text
pytest tests/parser/ # parsing, recovery, fixed point
pytest tests/integration/ # output after GLE has compiled it
The whole suite takes well under a minute. Useful variations:
pytest tests/unit/test_plotting.py::TestBasicPlotting # one class
pytest tests/ -k errorbar # by keyword
pytest tests/ --durations=10 # slowest tests
pytest --cov=gleplot.writer --cov-report=term-missing tests/
Coverage is a way to find untested branches in the module you changed, not a target; there is no minimum.
Layers
Directory |
What it asserts |
Needs GLE |
|---|---|---|
|
The object model and the |
No |
|
Lexing, parsing, recovery warnings, and the writer/parser fixed point |
No |
|
What GLE actually renders: files exist, ink lands where expected |
Yes |
Put a test in the lowest layer that can prove the point. An integration test
that only repeats a script-text assertion from tests/unit adds runtime and
no information.
Tests that need GLE
GLE is optional, and the suite must pass without it. A test that shells out to
the gle binary carries the gle marker:
import pytest
@pytest.mark.gle
def test_step_histogram_compiles(tmp_path):
...
# or for a whole module
pytestmark = pytest.mark.gle
tests/conftest.py skips marked tests when the binary is missing, so do not
write your own availability check. This applies in every directory: a unit or
parser test that compiles needs the marker too.
Two CI jobs keep this honest. test runs everything against a pinned GLE on
Ubuntu, macOS and Windows: it builds GLE on Ubuntu and macOS, and uses
upstream’s prebuilt static build on Windows (on Ubuntu it also runs the
examples and a strict Sphinx build);
test-without-gle runs the suite on a machine with no GLE and fails if an
unmarked test tries to compile.
To install GLE locally, see the
getting started guide.
gle -info confirms it is on PATH; set GLE_PATH to pick a specific binary.
The fixed point
Saving a figure, parsing the .gle back with parse_gle_figure, and saving
again must produce byte-identical output, including the .dat sidecars.
tests/parser/test_fixed_point.py runs this over every figure builder in
tests/parser/_golden_battery.py and tests/integration/test_project_io.py.
Its exemption set is empty and a test asserts that it stays empty. If your change makes a builder fail, the writer or the recognizer has a bug; fix that rather than exempting the builder.
When you add something the writer emits (a series field, a new option), add a
builder that uses it to _golden_battery.py.
Batteries
Several behaviour areas keep their cases in a shared helper so the unit and integration layers exercise the same figures:
tests/parser/_golden_battery.py— figures for the fixed pointtests/unit/_degenerate_battery.py— degenerate autoscale rangestests/unit/_refline_battery.py— reference lines and spanstests/unit/_single_bar_battery.py— single-bar width handling
Add a case to the relevant battery before creating a new test file. A new file is for a new behaviour area.
Checking compiled output
tests/integration/graphics_analysis.py has small analyzers for compiled
files:
from tests.integration.graphics_analysis import (
EPSAnalyzer, PDFAnalyzer, PNGAnalyzer, validate_graphics_file,
)
assert PDFAnalyzer(path).is_valid_pdf()
width, height = PNGAnalyzer(path).get_image_dimensions()
Tests that need to know what GLE drew, not just that a file exists, avoid decoding images where they can:
tests/integration/test_ink_on_page.pywraps the script in abegin nameblock and reads back its corner points, which gives the true bounding box of the ink, including anything that fell off the page.tests/integration/test_line_style_rendering.pyreads the dash pattern GLE writes into its EPS output.tests/integration/test_guide_marker_compiles.pycompares rasters with Pillow (in thedevextra) when two scripts must render identically.
Writing a test
Use pytest functions and the
tmp_pathfixture.Call
glp.close()before and after tests that use the module-levelglp.plot(...)API, so figures do not leak between tests.Run
black,isortandflake8overtests; the CIlintjob checks them.
Known platform issue
GLE 4.3.9 on the macOS ARM CI runners crashes rendering begin contour
blocks. Those tests carry the non-strict darwin_gle_contour_crash xfail from
tests/integration/conftest.py.