java-affected-tests

Files (1)

• SKILL.md

SKILL.md (first lines)

# Java tests to run before a PR

| Suite | PR checks | Merge queue | Nightly | You, before the PR |
| --- | --- | --- | --- | --- |
| Unit tests | full | full | — | impacted classes |
| Integration tests (`openmetadata-integration-tests`) | skipped | full, every lane and engine | — | impacted classes |
| search-it (`tests/search/*IT`) | — | — | openmetadata-nightly | impacted classes |
| JavaUIIT (`*UIIT`), scale-it | — | — | openmetadata-nightly | reported, not run |

The merge queue is the first CI run of the ITs. An IT failure there ejects the PR and
re-tests every PR queued behind it. Search-indexing changes have no pre-merge CI at all, so
your run is their only check before merge. Run what the planner selects, every time.

## 1. Preconditions

- `git fetch origin main`, so the diff base is current. The planner diffs `origin/main...HEAD`
  plus uncommitted and untracked files.
- Docker running, with no other Testcontainers stack up:
  `docker ps --filter label=org.testcontainers=true`. Another checkout's ITs plus yours rarely
  fit in Docker's memory. Wait for them; don't stop them.
- JDK 21. No `mvn install` needed: every command builds the modules under test with `-am`.

## 2. Plan

```bash
make java_affected                # what runs, why, and the exact commands
```

| Planner output | What to do |
| --- | --- |
| `No Java tests are impacted` or `No impacted Java test runs locally` | Put the NOT NEEDED block in the PR (`make java_affected_run` writes it and names any nightly-only tests) |
| `[unit] … FULL suite` | Expected for poms and widely used classes |
| `[integration] FULL suite` | Run it locally, or in CI on your branch and record the run with `--ci-run` (section 3) |
| `[integration] Class lane engine <- reason` | Run them all; the reason names the area, entity, reference or caller |
| `Impacted, but not run locally` | List them in the PR with where they run (nightly, external cluster); don't run |
| `Changed classes no unit test references` | Add a unit test (see `test-enforcement`), or say in the PR why none applies |
| `Impact-map gaps` | Add the file's directory to the area that owns that code in `.github/java-tests/impact-map.json`, in this PR, then re-plan. Until then the full suite runs |
| `Fix .github/java-tests/impact-map.json in this branch before you raise the PR` | Do what each line says: add the named glob to an area's `sources`, a pattern for the new IT to its `tests`, or remove a pattern your deletions emptied. `git push` and `gh pr create` stay blocked until it's clean (rule `java-impact-map.md`) |

### Ground rules for adding tests

The plan is a floor: run everything it selects, and never remove a test from it. It reads the
code, so it misses links that only show at runtime. Before you run, look at what the change does,
and add tests when it:

- changes what other entities see on delete, restore, rename or move (cascades, cleanup, orphans);
- changes a field or response that search documents, lineage, change events or alerts read;
- changes a permission, policy or validation check other endpoints also call;
- changes what an endpoint accepts or returns: ITs reach it over HTTP, so no IT names your class.
  Grep the SDK and the IT tree for the endpoint's path or SDK method, and add the ITs that call it.

When unsure, add the whole area. A few extra minutes locally cost less than an ejection from the
merge queue. Add with a reason, on the run as well as the plan (nothing is saved between them);
the details show both:

```bash
make java_affected_run ARGS='--add-area lineage --add-it TableResourceIT --reason "moves the edge payload lineage reads" --update-pr'
```

`--add-it`, `--add-unit` and `--add-area` take comma-separated names; `--reason` is required with
any of them. If the link is durable, change the map instead (an area pattern, or a directory in
an area's `sources`), so the next change gets it without anyone remembering.

## 3. Run and record

```bash
make java_affected_run                         # unit step first, then one step per lane and engine
make java_affected_run ARGS="--update-pr"      # once the PR exists: summary in its body, details in a comment
```

Measured on 2026-10-07: one unit class took 0.9 min. One small IT took 2.0 min, almost all of
it the `-am` reactor build plus the MySQL + Elasticsearch Testcontainers bootstrap, which every
IT step pays. Big classes add their own time; CI's parallel lane runs 334 classes in
46 min on four workers.

- The run stops at the first failing step. `--keep-going` runs the rest.
- A step passes only with exit 0, no failures, more than zero tests, and a report for every
  selected class. A green `BUILD SUCCESS` with zero tests is a failure.
- Results: the summary `target/java-tests/local-pr-results.md` and the details
  `target/java-tests/local-pr-results-details.md`. Before the PR exists, paste the summary between
  the `local-java-test-results` markers under "Backend integration tests". Once it exists,
  `gh pr comment --body-file target/java-tests/local-pr-results-details.md` posts the details, and
  a later `--update-pr` run edits that comment. Never edit either by hand.
- The summary stays short: the verdict, a row per step, and up to ten failed or missing names.
  The details' "Tests run locally" list names every class each step ran, with its counts. A module
  run in full is given as counts only: its class list alone can pass GitHub's 65,536-character
  limit for a comment. That list is the PR's record of what ran before review, so tests you ran
  by hand outside the planner go in the PR's Tests section too.

A full-suite plan (`[integration] FULL suite`) is every merge-queue IT. Running it in CI is
usually faster than locally:

```bash
git push                                                            # CI runs the commit you push
gh workflow run "Integration Tests - MySQL + Elasticsearch" --ref <branch>
gh run watch <run-id>
make java_affected_run ARGS="--ci-run <run-id> --update-pr"
```

`--ci-run` takes only a passed run of an integration-test workflow on `HEAD`, and covers only the
lanes whose jobs ran and passed in it. Everything else still runs locally: unit steps, search-it,
other engines, and lanes the workflow doesn't run (the Redis workflow has no rdf lane). Repeat
`--ci-run` for a second engine's workflow.

## 4. When something fails

1. Re-run the failing class alone (section 5), then the failing method: `-Dit.test='Class#method'`.
2. Fails on your branch only → it's your change. Fix it.
3. Also fails on a clean `origin/main` worktree with the same command → pre-existing. Link or
   open an issue and say so in the PR. Don't call the run passed.
4. Passes on retry → a flake. Name the test in the PR, with how often it failed.

## 5. Running one IT by hand

Copy the exact line from `make java_affected` when you can. The general shape:

```bash
mvn -B verify -pl :openmetadata-integration-tests -am -Dspring-boot.repackage.skip=true \
  -Pmysql-elasticsearch -DintegrationTests.lane=parallel -Dit.test=TableResourceIT \

From open-metadata/OpenMetadata, Apache-2.0.

More from open-metadata/OpenMetadata