[weiishann]

Why

PR #15 lost a whole unit-test matrix to a transient sbt-launch.jar download: one leg died in 70 ms, and fail-fast cancelled the other. Digging into that, a passing run turned out to spend most of its time doing work twice.

Measured from run 35453363517, per matrix leg (~22.5 min):

Step	Time	Share
integrationTests/test:compile	69s	5%
clean package publishM2	52s	4%
+jacoco	952s	70%
licenseCheck	265s	20%

Only spark and integrationTests read sparkVersion, but +jacoco ran the whole root aggregate in both legs — so ~7 min of server and cli tests, plus a ~4.4 min licenseCheck, ran twice per pull request, identically.

What changed

unit-tests.yml splits into two jobs. core runs the eight Spark-independent modules once; spark matrixes only the modules that actually depend on the Spark version. Together they cover exactly root's .aggregate(...) — nothing dropped, nothing run twice.

licenseCheck moves to its own workflow, gated on build.sbt, version.sbt and project/**, since dependency licences can only change when those do.

The launcher download retries. --retry covers 5xx and timeouts, --retry-all-errors extends that to connection resets, --show-error stops --silent hiding the HTTP status. dev/test-sbt-launch-lib.sh covers both failure modes against a local server, and was confirmed to fail before the change and pass after.

Spark's test logging is quieted. 226,580 of 240,773 lines in a passing job were Spark INFO messages on stderr, which sbt tags [error]. All of it came from the connector tests. Measured on SchemaOperationsTest: −77% log lines, same 5/5 result.

Also: dependency caching, fail-fast: false, concurrency groups, timeout-minutes on every job, jacoco reports uploaded instead of discarded, format and UI checks extended to branch-*, and a Dependabot config.

Projected from the measured step times: runner time ~45 → ~31 min, wall-clock ~22.5 → ~12.5 min.

Please note before merging

test (Spark 4.0.0) and test (Spark 4.1.0) keep their exact names, because both are required status checks — renaming them would leave every PR waiting on checks that never report.

core and license-check are new contexts and are not in branch protection. Until they're added, the server and CLI suites and the licence check all run and report but do not block a merge. They gate today via test (Spark X), so this is a real gap that should be closed right after merge.

For the same reason, path-filtering unit-tests.yml is not included here: a required check skipped by a path filter never reports, and the pull request hangs forever.

Verification

• Both required contexts still produced, verified by expanding the matrix from the parsed YAML
• core ∪ spark == root's .aggregate(...), verified against build.sbt
• All 10 sbt task keys resolve, sbt exit 0
• Launcher tests red before the fix, green after; real Maven Central fetch still works
• shellcheck: 13 findings before, 13 after — identical, all pre-existing
• Spark connector test passes 5/5 with the new logging config

🤖 Generated with Claude Code