acoustic-carpenter-78188
05/12/2026, 8:20 AMBadRequest) or HTTP 422 (Invalid) — typically a validating admission webhook saying no, or an invalid field caught by the apiserver — the failure path falls through to the default branch and surfaces as a retryable system error. The workflow then burns through its retry budget making the exact same request, with the exact same outcome.
These errors are intrinsic to the request payload. Retrying with the same input produces the same rejection. Neighboring EntityTooLarge is already handled this way.
The right transition was already written but commented out since #1 (2019) with no documented rationale. Uncommenting and adding a regression test.
What changed
• `flytepropeller/pkg/controller/nodes/task/k8s/plugin_manager.go`: uncomment the PhaseInfoFailure("BadTaskFormat", ...) transition so BadRequest / Invalid produce a permanent failure instead of falling through to the system-error path. Added a short comment so this doesn't get re-commented in the future.
How was this patch tested?
Added two sub-tests to TestK8sTaskExecutor_Handle_LaunchResource — jobBadRequest (using k8serrors.NewBadRequest) and jobInvalid (using k8serrors.NewInvalid). Both assert pluginsCore.PhasePermanentFailure with error code BadTaskFormat. Both fail on master and pass with this PR; the full plugin_manager_test.go suite stays green.
ok github.com/flyteorg/flyte/flytepropeller/pkg/controller/nodes/task/k8s 1.724s
Check all the applicable boxes
• I updated the documentation accordingly.
• All new and existing tests passed.
• All commits are signed-off.
flyteorg/flyteacoustic-carpenter-78188
05/15/2026, 3:25 AM