Distributed Workflow Consistency

distributed-workflow-consistency

Registry Trigger

Use when

• An operation spans independently committed service effects.
• Partial progress needs durable recovery or repair.

Do not use when

• Route one-store atomicity to transaction-consistency, local repetition to idempotency-retry-design, and schema-only work to its messaging or API owner.
• Do not design a workflow engine, SDK, scheduler, or administration product.

Skill Role

Define cross-service recovery contracts without replacing participant or transport owners.

High-Value Rules

• Define durable business, workflow, run, step, command, effect, attempt, tenant, and definition-version identity with persisted state before dispatch.
• Record transition authority, guards, fences, timestamps, terminal meaning, and invalid legacy state.
• Correlate commands and results to workflow, step, and effect while rejecting send or broker acknowledgement as completion.
• Preserve unknown outcomes after timeout, crash, cancellation, or lost response until participant authority or reconciliation proves the effect.
• Require forward and compensating effects to be idempotent or durably deduplicated, with compensation inputs, preconditions, pivots, ordering, and convergence recorded.
• Reconcile desired state against participant facts with bounded, attributable corrections safe under concurrent progress.
• Detect stuck or poison work and constrain repair by target, authority, precondition, dry run, evidence, audit identity, and terminal outcome.
• Define active-workflow version compatibility, replay, migration or pinning, and retirement evidence.

Anti-Patterns

• Duplicate delivery commits an effect twice.
• Lost completion leaves a committed effect unfinished.
• Wrong compensation violates current business state.
• Partial ordering advances a dependency early.
• Poison or stuck work loops or blocks progress.
• Old-version execution misreads state or commands.
• Manual repair changes state without audit evidence.

Execution Checklist

• Fault before/after dispatch and effect, before result persistence, and during compensation.
• Exercise duplicate, reordered, delayed, poison, lost-response, stuck, repair, and participant-drift cases.
• Replay representative old/new definitions against histories and mixed participant versions.

Stop Conditions

• Stop on ambiguous identity, missing durable state, unknown effect authority, unsafe compensation, or unowned stuck work.
• Stop repair without authorization, audit, convergence checks, and repeat safety.
• Route atomicity, retry, schema, provider guarantees, and engine implementation to their owners.

Output Contract

• workflow-consistency decision with identity, state, correlation, unknown outcomes, recovery, repair, versioning, limits, and owners

Targeted References