FIELD NOTE / X
A timeout is not a rollback.
The short film, the complete written thought, and the evidence behind it.
The X conversation link will follow its public release.
A timeout is not a rollback.
Video caption
A timeout is not a rollback. A second move starts from the wrong belief. Hold the second action until independent state is known. #EricFieldNotes
Full written post / accessibility read
An operator agent asks to move a customer's inference service to another pool. The control plane accepts it. The reply vanishes. The agent writes failed, even though the change is already underway.
If the agent retries with a new intent, it may start a second move while the first is draining traffic. A transport exception cannot tell it which target state exists.
Persist the original operation identity. Query its acceptance record, then read placement and serving state from their owners. Failed, pending and unknown are different branches in the recovery workflow.
Put unknown between attempted and failed. Reconcile by stable operation ID and authoritative target readback before permitting another move. Do that because lost observation is not proof of lost effect.
#EricFieldNotes
Four-beat scene transcript
1. A timeout is not a rollback.
An operator agent asks to move a customer's inference service to another pool. The control plane accepts it. The reply vanishes. The agent writes failed, even though the change is already underway.
Visual: The caller lost a reply after the service accepted the move.
2. The next action can double the harm.
If the agent retries with a new intent, it may start a second move while the first is draining traffic. A transport exception cannot tell it which target state exists.
Visual: A second move starts from the wrong belief.
3. Give uncertainty its own state.
Persist the original operation identity. Query its acceptance record, then read placement and serving state from their owners. Failed, pending and unknown are different branches in the recovery workflow.
Visual: Query the operation record and target before another side effect.
4. Do not guess failed.
Put unknown between attempted and failed. Reconcile by stable operation ID and authoritative target readback before permitting another move. Do that because lost observation is not proof of lost effect.
Visual: Hold the second action until independent state is known.
Research and claim limits
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.