States, phases and evidence
Run state summarizes durable public work. Backend observations and progress phases add context but do not independently establish a terminal outcome.
- QueuedAccepted and waiting for assignment
- RunningPreparing → waiting → executing → collecting → uploading
- TerminalSucceeded, failed or cancelled
Public run state
Section titled “Public run state”| State | Interpretation |
|---|---|
queued |
Accepted operation waiting for an eligible assignment; may carry a blocked reason |
running |
Delegated preparation, backend waiting, execution or result delivery is underway |
succeeded |
Required results are validated and published |
failed |
A confirmed failure has been retained, including applicable execution or delivery failures |
cancelled |
Cancellation is confirmed rather than merely requested |
Running phases are preparing, waiting_for_backend, executing, collecting_results and uploading_results. They describe progress within an assignment. A scheduler saying COMPLETED does not independently establish public success; worker/result evidence and required output checks still matter.
Queue diagnostics
Section titled “Queue diagnostics”blocked_reason |
Investigate |
|---|---|
publication_revoked |
The accepted publication was revoked |
installation_revoked |
Its exact installation snapshot was revoked |
no_approved_execution_path |
Binding/installation is not approved for dispatch |
model_access_revoked |
Submitter’s dispatch authorization changed |
compute_access_revoked |
Selected compute grant changed |
compute_unavailable |
The connector or its context is unavailable under its lifecycle policy |
connector_maintenance |
Connector maintenance retains accepted queued work until resume |
context_cutover |
Context cutover closed queued dispatch; inspect its accepted-work policy and cancellation receipt |
Capacity exhaustion and temporary disconnection can be ordinary waiting without a blocked reason. Closing new admission does not itself invalidate accepted work. Inspect configuration lag, connector health and admission limits alongside the queue.
Connector maintenance and delivery
Section titled “Connector maintenance and delivery”Maintenance (draining) stops new acceptance and claims, retains queued runs for
resume and permits existing assigned preparation, pending allocations, execution
and delivery to finish subject to independent authority. Reversible connector
authority withdrawal is separate: it blocks prelaunch work and requests cancellation
of pending allocations while already running execution and delivery can drain.
Resume and authority restoration change only their respective controls; permanent
revocations, user access, grants and credential validity remain independent.
Connector diagnostics separate retained queued runs, assigned phases, unconfirmed
assignments, terminal assignments without durable delivery receipts, active uploads
and pending log archives. Scientific terminal state does not imply that all delivery
obligations finished. connector wait-maintenance completes only while maintenance
remains enabled at the requested generation and assigned work plus delivery have
settled. Retained queued runs and configuration acknowledgement lag alone do not
block completion. Stale contact never proves execution or delivery ended.
Cancellation
Section titled “Cancellation”Cancellation status is separate from state: requested, confirmed or completed if another terminal outcome won the race. Assigned work remains running while termination is unconfirmed. Repeated cancellation preserves the original intent/time. A wait deadline, disconnected stream or stopped connector process does not confirm model termination.
Context cutover with cancel_queued records connector-initiated cancellation
with the actual actor and a context-cutover reason. It applies only to runs still
queued at cutover commit. Creator cancellation continues to use its ordinary
authority and assigned-work cancellation path.
Context replacement and retirement
Section titled “Context replacement and retirement”Context lifecycle is active → draining → retired; only draining can reopen.
Execution definitions and accepted destinations stay immutable. Cutover requires
all live current publications moved or explicitly withdrawn, then closes fresh
grants, registration, activation, standard submissions and candidate trials on the
source. Accepted-work drain continues eligible queued claims; cancel_queued
cancels only queued work. Assigned work and delivery continue under their original
identity and independent authority, maintenance, capacity and revocation controls.
Configuration release explicitly requests retirement of obsolete publications,
including abandoned candidates. Context retirement_ready requires no live current
publication, queued work, unfinished or uncertain assignment, pending delivery,
active upload or pending archive, plus all source publication configuration retired
and connector acknowledgement of final-reference removal. Final retirement retains
history, provenance, grants and permanent revocations. Reopening never reverses
cancellations, configuration-retirement requests or author publication decisions.
Readiness concerns the source context. Whole-connector work and local cleanup have separate evidence; shared connectors, journals and protected releases need their existing inspection and retention procedures. See context replacement.
Publication and archives
Section titled “Publication and archives”Candidates progress through configuration validation before ready permits activation. Inspect the exact candidate status and validation_message; a failed candidate leaves the current publication intact. Publication updates preserve accepted runs’ pinned publication, scientific definition, execution destination and installation. Superseding a publication does not revoke accepted work.
Log archive attachment status is capturing, pending, ready, unavailable or expired, independent of the scientific outcome. Completeness describes retained source ranges rather than publication success. A ready archive may be partial; an unavailable archive cannot turn a successful model into a failed one.
Use execution troubleshooting to resolve observed states. Preserve backend/job ownership and exact request identity when recovering uncertain work.
Documentation 205dcb3116ea · public sources 8c70356d340f · built .Download documentation identity.