Search ARuntime.com

Find runtime definitions and implementation guidance

Search page titles, summaries, headings, glossary terms, use cases, and runtime-directory entries.

Enter at least two characters.

Build and Operate

Failure Recovery

Detection, retry eligibility, idempotency, rollback, compensation, escalation, and evidence for runtime failures.

Audience: Platform engineers; application developers; incident responders Reading time: 2 minutes Status: Production guidance Last reviewed:

A production runtime must define what happens after timeouts, invalid output, denied tools, partial side effects, duplicate execution, budget exhaustion, and evidence failures.

Failure and recovery state machine

Failure and recovery state machineA state machine begins at admitted, then executing, validating, and completed. Failures branch to classified. A transient and idempotent failure can enter delayed retry and return to executing. An authorization or policy denial ends as denied. A partial side effect enters reconcile, then either compensated, completed with warning, or escalated. Budget exhaustion and human-review timeout end as suspended. Every terminal state writes evidence, and duplicate requests return the prior result when the idempotency record is complete.AdmittedExecutingValidatingCompletedClassified failureSafe retryDeniedReconcileCompensateEscalateAll terminal states persist classified evidence; retries require idempotency.
Retries are allowed only at classified safe points; partial side effects require reconciliation or compensation.
Text description

A state machine begins at admitted, then executing, validating, and completed. Failures branch to classified. A transient and idempotent failure can enter delayed retry and return to executing. An authorization or policy denial ends as denied. A partial side effect enters reconcile, then either compensated, completed with warning, or escalated. Budget exhaustion and human-review timeout end as suspended. Every terminal state writes evidence, and duplicate requests return the prior result when the idempotency record is complete.

Recovery rules

  1. Detect and classify the failure.
  2. Determine whether retry is safe and whether an idempotency key is required.
  3. Preserve or restore known-good state.
  4. Compensate completed side effects where rollback is impossible.
  5. Escalate to a human or incident path when authority or evidence is insufficient.
  6. Persist minimized evidence of decisions, attempts, effects, and recovery.

Use the reference architecture failure matrix for failure-specific behavior.

Maintenance record

Found an error, outdated capability, or unclear category boundary? Submit a correction with a supporting source.