#7044 fix(plugins): classify CreateContainerError as system-retryable instead of user error Pull request opened by themavik Summary When a pod fails with

CreateContainerError

or

CreateContainerConfigError

and the configured grace period is exceeded, the error is classified as a USER permanent failure (

PhaseInfoFailureWithCleanup

). This prevents system retries from triggering — the task fails immediately even when

retries

is configured. These errors are typically transient infrastructure issues (e.g. "failed to reserve container name", "container name already in use") that should be retried by the system. Root Cause In

pod_helper.go

, both

CreateContainerError

and

CreateContainerConfigError

use

PhaseInfoFailureWithCleanup

after the grace period, which creates: • Phase:

PhasePermanentFailure

• Error kind:

ExecutionError_USER

USER errors do not trigger system retries (

max-node-retries-system-failures

), so the task fails immediately. Fix Changed both error paths to use

PhaseInfoSystemRetryableFailureWithCleanup

, which creates: • Phase:

PhaseRetryableFailure

• Error kind:

ExecutionError_SYSTEM

This allows the system retry mechanism to handle transient container creation failures. Testing • Updated existing

CreateContainerErrorOutsideGracePeriod

test to assert

PhaseRetryableFailure

instead of

PhasePermanentFailure

• Updated existing

CreateContainerConfigErrorOutsideGracePeriod

test similarly • Both tests verify

CleanupOnFailure()

remains

true

• Within-grace-period tests remain unchanged (still

PhaseInitializing

) Closes #6786 flyteorg/flyte