Workflow retries and recovery
A retry can continue a failed Step, repair one Loop iteration, or start an entirely new run. Before choosing a control, inspect the failed input, saved version, and result in the target app. A timeout does not prove the provider made no change.
Check automatic work first
Open Run history, select the run, and inspect Status, Attempts, Version, and the first failure. Temporary failures may already be waiting for an automatic retry. Queued can also mean a Delay is waiting.
The worker allows up to 10 queue retry decisions after the initial execution attempt for retryable failures. Provider Retry-After information can influence the wait within runtime limits. Validation errors, missing permissions, and deleted resources usually require a fix rather than repeating the same call.
Retry handling precedes a connected Error route. A final handled failure may let the run succeed; an unhandled failure ends it as Failed. Queue delivery exhaustion is recorded as workflow_delivery_exhausted. See Error handling.
Refresh and Retry loading runs only reload history. They do not retry execution.
Choose the recovery scope
For ordinary failed Steps, the available buttons distinguish the original run from a new execution:
| Control | Settings and scope | What repeats |
|---|---|---|
| Retry and continue, or Retry and continue on original vN | Resumes the existing run on its original version. | The failed Step and following work. Earlier successful results are reused. |
| Retry and continue on current vM | Creates a new run from the failed Step with current settings, when the version differs. Requires an active Workflow. | The failed Step and following work. Earlier saved successful results are reused. |
| Run again on vM | Creates a new run from the recorded Trigger input using the current Workflow version. Requires an active Workflow. | Actions start again from the beginning, including previously successful work. |
Read the version number on the button. If you fixed a mapping in the current graph, an original-version retry still uses the old mapping. Conversely, current-version continuation does not recompute the earlier successful Steps whose outputs it reuses.
Buttons depend on run status and available evidence. Missing saved output, an older run without an exact graph snapshot, or an incompatible graph change can make recovery unavailable. Do not substitute an empty payload or current sample for missing historical data without understanding that it changes the operation.
Repair one failed Loop iteration
In a run's Loop details, use Search iterations and Iteration status to find the failed item. Expand it and inspect the failed child Step. Open iteration link and Copy link help identify the exact nested item.
Where offered, Retry failed iteration applies the latest saved Step settings, starts at the failed Step, and reuses completed results. The Loop items and routes must remain unchanged. Other failed iterations remain unchanged, and completed Steps after the Loop are not run again.
If the original run stopped, unfinished work can resume after the repair. This operation differs from ordinary original-version Step retry: read the iteration-specific explanation in the UI. Missing evidence or incompatible item and route changes can prevent it.
Inspect the provider for partial success before repairing the failed child Action. Reusing completed results avoids deliberately repeating those Steps, but the failed request itself may already have produced an external effect. See Loops.
Recover after changing access or settings
- Inspect the first failure and its actual mapped input.
- Repair account access, resource selection, or invalid inputs as appropriate.
- Check the provider for an existing or partial result.
- Choose original-version continuation, current-version continuation, iteration repair, or a full new run according to the work that should repeat.
- Inspect the resulting run and provider result, including other selected Paths.
An unhandled failure in one Path does not undo or necessarily stop the other selected Paths. A failed run can therefore contain successful external changes.
Limit duplicate effects
Use stable business IDs, the provider's idempotency support where available, and an appropriate Only new items configuration. An existing-record lookup can help, but a lookup followed by a create is not automatically atomic across simultaneous runs.
Do not assume a generic webhook delivery replay has the same scope as Workflow recovery. Retries and replay covers Connection delivery; use Runs & errors for Workflow evidence.