Skip to main content
How-to guideApplies to: Unified engine deployments
Follow the recorded execution rather than assuming the current editor represents what an earlier run used. A run connects the saved revision, parameter values, schedule trigger, task states, and attempts.

Find the run identity

Open the workload’s run history and select a run. Record its workload ID, run ID, revision, trigger, scheduled time, and effective parameters. For windowed work, also record the data window. Selective reruns and backfills add their origin or backfill identity. The editor can contain a newer revision. When using a current-definition graph or code view, compare it with the run’s recorded invocation before drawing conclusions.

Follow tasks

Use the timeline to compare phases and durations. Open a task to inspect attempts, failures, execution settings, and commit evidence. A retry adds an attempt to the existing task run; it does not create another workload run. A sensor can be Running without holding a SQL compute worker. Read its check count, next check time, deadline, and last observation.

Compare runs

Compare revision, supplied and effective parameters, context, windows, compute, and task attempts. Identical names do not imply identical execution settings. A reused task in a selective rerun should show its source run rather than a new execution attempt. For the first workload, two runs of the same revision should calculate the same six-order total. Editing the query should produce a later revision while earlier invocation records remain intact.

Control future and active work

Use Rerun tasks on a completed run to preview selective execution. Use the workload’s Backfills tab for historical windows, and Notification history on a run for delivery attempts. These actions have different identities and do not replace each other. Run cancellation is available through POST /api/v1/workloads/{workload_id}/runs/{run_id}/cancel. Cancellation requests a stop; it does not undo confirmed writes. Pausing does not cancel active work, and cancelling one run does not pause the schedule. Closing the browser does not stop engine execution.

Inspect results and diagnostics

Run and task pages focus on execution. Use retained result APIs or query the destination in Workbench for rows. Read Results and table commits for publication checks. Use Observability for instance-wide metrics, and Troubleshooting to choose the next action for a specific symptom.
Last modified on November 4, 2026