After an administrator stops a long-running workflow, the operating table can still show its last waiting step. That does not necessarily mean the stop failed. The record may describe the execution before it ended, with no later function invocation available to tell the reporting plugin what happened. An apparently live state can therefore be an old observation.

AWS's October 28 explanation of the Lambda durable execution telemetry exporter identifies this boundary directly. The plugin runs inside the function and emits records when the function is invoked. Service-originated STOPPED and TIMED_OUT states that occur without an invocation do not reach it. Complete lifecycle coverage needs the corresponding state events as an additional observation path.

Separate the last observation from the execution's status

An international sales or approval process may span several invocations, waits and resumes. A successful step does not establish completion of the entire task. A last snapshot showing a wait does not establish that a resumable execution still exists now. The report should express what has been observed rather than extend the observation indefinitely into the present.

Keep the observation time and source accessible beside the state. When a technical operator sees an old waiting record, the next question is whether a later service-side state exists. It should not automatically become a request for the business owner to approve again. Otherwise a monitoring gap can create additional human actions or repeated handling of the same request.

The reconciliation view need not expose the buyer's complete message or documents to every operator. Execution identity, observed state and timing can support routine investigation, while sensitive task content remains available only through the appropriately restricted business record.

The opposite inference is also unsafe. No fresh record does not prove that the workflow stopped. A durable process may legitimately wait for a long period. Distinguish a confirmed active state, a confirmed ending and a case that needs reconciliation. A color based on record age can flag attention without pretending to be authoritative evidence of termination.

For a Chinese export team whose overseas colleagues depend on these records, this distinction belongs in the handoff language too. A status message should not say the buyer's request is still waiting for approval merely because that was the last available snapshot. Explain uncertainty where it affects the next person's action, rather than making them infer it from an old timestamp.

Give service-side endings a route into the record

AWS describes subscribing to durable execution lifecycle events on the default EventBridge bus and updating the destination through a handler. An implementation still needs to correlate those events with the same execution. A familiar workflow or step name is not enough to distinguish different buyers or runs. Preserve the execution identity rather than overwrite whichever row looks similar.

Snapshots and lifecycle events may arrive through different paths. Retaining the time and identity of what they describe helps prevent an older snapshot arriving later from reverting a terminal state to waiting. The actual conflict rule needs design and testing in the selected system. Adding another event source is not a guarantee that ordering and duplicate handling are already correct.

Use reconciled states for business notifications. A confirmed stop can establish that the execution will not continue, but it does not establish that earlier external actions were undone. An email sent before the stop remains a separate fact. Inspect completed steps and their effects before describing the result as a successful business rollback.

This is different from letting a workflow wait for human confirmation. Durable waiting preserves the basis for later execution. Terminal-state observation establishes whether that basis still exists. Both may look inactive in a simple interface, yet they require different operational responses. The distinction matters when someone is deciding whether to resume, investigate or create a separate task.

Test several endings in a controlled environment: normal completion, a failed step, a manual stop and execution timeout. Check the resulting record, retained evidence and notification target for each. A demonstration of normal completion alone cannot establish lifecycle coverage. The useful acceptance question is whether the team can determine how this execution actually ended without repeating business work to find out.

Sources