lfoppiano / grobid-quantities / 32657569994

39%

push

github

Fix integration tests (#202)

* test: make the integration tests find grobid-home, and stop them sharing a JVM

`./gradlew integration` failed 18 of 252 tests, in
QuantityParserIntegrationTest and ValueParserIntegrationTest. Two
independent causes, both of which had to go.

**The tests shared one JVM.** Grobid keeps its state in statics -
GrobidProperties, the resolved grobid-home, the wapiti model registry -
so whichever test class initialises first decides for every class after
it. Both failing classes pass when run on their own and fail in the
full suite, which is the signature. Setting forkEvery = 1 on the
integration task turns the suite green: 252 tests, 0 failures. It costs
about two minutes instead of twenty seconds, acceptable for a task that
is not part of `build`.

This is also what CI was hitting. The manual integration workflow
reported exactly the same 18 failures in the same two classes, on a
runner whose log shows it had loaded values/model.wapiti - so its
grobid-home was fine and the path was never the problem there.

**The configured grobid-home was ignored.** initEngineForTests() read
grobidHome from config-test.yml and then called
GrobidProperties.getInstance() with no GrobidHomeFinder, so Grobid fell
back to its own default and the setting did nothing. Demonstrated:
pointing config-test.yml at a complete grobid-home changed nothing (6
failures before, 6 after), while passing the same path through a
GrobidHomeFinder made those tests pass.

findGrobidHome() now resolves it, taking the first candidate that
contains config/grobid.yaml rather than the first that merely exists:

  -Dorg.grobid.home=...   config-test.yml   ../grobid/grobid-home   ../grobid-home

The usable-versus-exists distinction is the point. A partially
populated ../grobid-home - models but no config - satisfies
GrobidHomeFinder's own existence check and shadows a complete
installation beside it, which is exactly the layout on hand here. When
nothing is usable the fa... (continued)

1019 of 3002 branches covered (33.94%)

Branch coverage included in aggregate %.

2670 of 6463 relevant lines covered (41.31%)

1.65 hits per line