Hibernate ORM × CUBRID dialect verification

Run the hibernate-core suite against CUBRID, then quantify a dialect change: total failures, recovered/regressed vs a baseline, and failure families. Requires the Hibernate repo (default ~/Devel/hibernate), Docker, and python3 (stdlib). The full suite is ~16,900 tests and takes ~30 min: use --tests for fast iteration.

Step 1: Decide version, scope, baseline

• Version: 10.2 (dialect minimum) or 11.4 (current). Read from $ARGUMENTS; ask if unset (semantics differ).
• Scope: full suite (default) or a --tests filter (e.g. --tests 'org.hibernate.orm.test.manytomany.*') for a quick check.
• Baseline: the failing-test set to diff against: either a saved baseline.json from a prior run, or capture one now by running Step 4–5 on the un-changed dialect first.

Step 2: Start CUBRID

cd ~/Devel/hibernate
CUBRID_IMAGE=docker.io/cubrid/cubrid:<version> ./db.sh cubrid   # waits for the csql healthcheck

Step 3: Apply the change under test

The change lives in the working tree: e.g. hibernate-community-dialects/.../CUBRIDDialect.java, or a config toggle like hibernate.auto_quote_keyword. For a before/after diff, capture the baseline (Steps 4–5) with the change reverted first, then re-apply it.

Step 4: Run the suite

./gradlew :hibernate-core:test -Pdb=cubrid -Plog-test-progress=true --stacktrace --no-daemon 2>&1 | tee <run>.log
# fast iteration: add  --tests '<filter>' ;  slow deps:  prefix  MAVEN_MIRROR="https://maven-central.storage-download.googleapis.com/maven2/"

The Gradle task exits non-zero when tests fail: that is expected; continue to Step 5. Profile -Pdb=cubrid selects the CUBRID dialect/JDBC from local-build-plugins/.../local.databases.gradle. Results: hibernate-core/target/test-results/test/TEST-*.xml.

Step 5: Parse results → failure set

python3 ${CLAUDE_SKILL_DIR}/assets/parse_results.py parse \
  ~/Devel/hibernate/hibernate-core/target/test-results/test  <run>.json

Prints total / passed / failed / skipped and a per-family failure breakdown, and saves the failing-test set (with each failure's exception signature + family) to <run>.json.

Step 6: Diff vs baseline → recovered / regressed

python3 ${CLAUDE_SKILL_DIR}/assets/parse_results.py diff  baseline.json  <run>.json

Prints net (baseline − after), +N recovered (failed in baseline, not now), −N regressed (new failures), each broken down by family, and lists the regressed tests.

Step 7: (optional) separate flaky failures

Load-dependent driver NPEs often pass in isolation. Re-run a suspect/regressed subset alone:

./gradlew :hibernate-core:test -Pdb=cubrid --tests '<Class or pattern>' --stacktrace --no-daemon

Failures that pass on isolated rerun = flaky → exclude from the regression count and note them.

Step 8: Summarize / hand to the report skill

State: total OFF→ON, +N recovered / −N regressed, net, and the top failure families. To produce the Word report, feed these to the report skill as a 비교 분석: a 전/후 comparison table + a bar chart (failures before→after) + an hbar of recovered-by-family.

Step 9 - Tear down (free disk)

The CUBRID container holds a ~2.7 GB anonymous data volume, and db.sh's built-in compose_down is a no-op unless REMOVE_ORPHANS is set: so re-running Step 2 for a new image/version orphans the old volume and the disk fills up over repeated runs. Always tear down when finished:

docker compose -p hibernate_orm \
  -f ~/Devel/hibernate/docker-compose/latest/cubrid/docker-compose.yaml down -v
# sweep stragglers from earlier runs (anonymous = 64-hex; named volumes like jenkins are kept):
docker volume ls -qf dangling=true | grep -E '^[0-9a-f]{64}$' | xargs -r docker volume rm

(The hhh-dialect-verify-report orchestrator does this automatically after every version.)