Submit and monitor a runΒΆ
Validate the complete request and review the estimate before submission. Keep one idempotency key for one intended run; use it to reconcile a lost response. A new key can start additional compute.
The central service returns a durable job ID. Follow get_job or GET /v1/jobs/{job_id}, using the returned queue message and progress information. Queue time, execution time and result preparation are different stages. A disconnected client or expired local wait does not cancel the job.
Read incremental events and saved histories when available. Report observed residuals and forces with their qualification rather than inferring their absence from a missing field export. Never automatically rerun a failed or numerically unconverged calculation.
If the state is submission_unknown, retain the ID and request and seek reconciliation instead of submitting a replacement. If cancellation is requested, continue until a terminal state confirms the outcome.
On completion, assess convergence and open the saved result viewer. Viewer preparation can have its own durable operation. A visualization failure does not invalidate an otherwise saved simulation; recover the view without rerunning the solver.
See job states and recovery, convergence and post-processing.