You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The compat-matrix engine jobs run with continue-on-error, so the workflow run concludes "success" even when every engine job fails. Two consequences observed on 2026-07-06:
UpdateAllIncludeJoinSpec has been red on lucee7+mysqlsince at least 2026-06-28 (a real data-affecting bug, now updateAll with multiple includes silently updates 0 rows on MySQL #3294) — invisible for the same reason. The June-28 baseline additionally carries per-engine debt (Adobe: typedColumnDefaultsSpec, contentSpec, RouteTesterHardeningSpec, transaction specs; Lucee 6: RouteTesterHardeningSpec; BoxLang: miscellaneousSpec, AdapterIdentityTemplateSpec, testEvaluateExpressionSpec, DatabaseAdapterSpec) that nobody triaged because nothing ever turned red at the run level.
Same lesson as the box-install check (#3281): a check that can't fail visibly is noise, and real regressions hide behind it.
Proposal
Keep per-DB soft-fail for the known-flaky infrastructure (SOFT_FAIL_DBS=oracle already exists) but make a hard-DB leg failure fail the engine job, and an engine-job failure fail the run.
To get there without a permanently red banner, first burn down (or explicitly soft-fail with a linked issue) the pre-existing per-engine failures listed above — each should be a tracked issue, not silent debt.
Optional: have the Test Matrix Summary job post a compact per-engine × per-DB verdict table into the run summary so a glance shows exactly which leg regressed.
Evidence
Run 28809821005 (2026-07-06 dispatch): conclusion success, all 5 engine jobs failure.
Problem
The compat-matrix engine jobs run with
continue-on-error, so the workflow run concludes "success" even when every engine job fails. Two consequences observed on 2026-07-06:ConfigRoutesStaleDocUrlSpec, from fix(cli): repoint 15 dead v4-0-0-snapshot and 3.1.0 guide URLs at live v4 pages #3281) failed every engine × database leg uniformly for a full day before anyone noticed — the run banner was green (fixed in fix(build): retire dead 3.1.0 guide URL in docker engine settings templates #3293).UpdateAllIncludeJoinSpechas been red onlucee7+mysqlsince at least 2026-06-28 (a real data-affecting bug, now updateAll with multiple includes silently updates 0 rows on MySQL #3294) — invisible for the same reason. The June-28 baseline additionally carries per-engine debt (Adobe:typedColumnDefaultsSpec,contentSpec,RouteTesterHardeningSpec, transaction specs; Lucee 6:RouteTesterHardeningSpec; BoxLang:miscellaneousSpec,AdapterIdentityTemplateSpec,testEvaluateExpressionSpec,DatabaseAdapterSpec) that nobody triaged because nothing ever turned red at the run level.Same lesson as the box-install check (#3281): a check that can't fail visibly is noise, and real regressions hide behind it.
Proposal
SOFT_FAIL_DBS=oraclealready exists) but make a hard-DB leg failure fail the engine job, and an engine-job failure fail the run.Evidence
success, all 5 engine jobsfailure.