#7093 feat(executor): in-place task retry with scoped output paths per attempt Pull request opened by AdilFayyaz Why are the changes needed? When a task pod fails with a recoverable error, the executor relied on the runs service to create a new TaskAction CR for each retry. This external orchestration was broken, leaving retryable tasks stuck in a non-terminal RetryableFailure loop indefinitely. What changes were proposed in this pull request? • Intercepts PhaseRetryableFailure in TaskActionReconciler.Reconcile() before updating status • If retries remain (currentAttempts < maxAttempts): calls p.Abort() to delete the failed pod, increments Status.Attempts, clears PluginState, and overrides the transition to PhaseQueued — causing the next reconcile to launch a fresh pod under the same TaskAction • If retries are exhausted: converts PhaseRetryableFailure → PhasePermanentFailure, making the TaskAction terminal (Failed=True) • Pod naming is naturally collision-free: buildGeneratedName encodes the attempt number (-0, -1, ...), so each restart creates a distinct pod How was this patch tested? • Deploy a task with retries: 2 (maxAttempts=3) that always exits non-zero • Verify pods

[name]-0, [name]-1, [name]-2

are each created and deleted in sequence:

kubectl get pods -n [namespace] | grep [taskaction-name]

• Verify Status.Attempts increments on each failure:

kubectl get taskaction [name] -n [namespace] -o jsonpath='{.status.attempts}'

• After the third failure, verify the TaskAction is terminal:

kubectl get taskaction [name] -n [namespace] -o jsonpath='{.status.conditions}'

--- Failed=True, Reason=PermanentFailure • Deploy a task with retries: 0 — confirm first failure immediately produces PermanentFailure This is important to improve the readability of release notes. Setup process Screenshots Check all the applicable boxes • I updated the documentation accordingly. • All new and existing tests passed. • All commits are signed-off. Related PRs Docs link •

main

• #6583 • feat(executor): in-place task retry with scoped output paths per attempt 👈 flyteorg/flyte