transaction-consistency-designer

Transaction Consistency Designer

Source mapping: Tier 1 critical skill derived from Kotlin_Spring_Developer_Pipeline.md (SK-09).

Mission

Place transaction boundaries where business invariants are actually enforced, not where annotations are easiest to type. Prevent data loss, duplicate side effects, and hidden consistency bugs.

Gather These Inputs

• The business workflow step by step.
• The repositories and tables touched by each step.
• Current @Transactional annotations, propagation, isolation, and exception handling.
• Any external HTTP, message broker, scheduler, or file I/O inside the workflow.
• Idempotency, retry, and concurrency requirements.

Model The Workflow Explicitly

• Break the use case into state-changing steps and side effects.
• Mark which steps must be atomic together and which can be asynchronous.
• Mark where external systems are called.
• Mark where retries may happen.
• Mark the business invariant that must not be violated.

Decision Rules

• Keep one database transaction focused on one consistency boundary.
• Avoid holding a database transaction open across external network calls.
• Prefer idempotency keys plus unique constraints for duplicate-request safety.
• Prefer the outbox pattern or post-commit publication for messages that must reflect committed state.
• Choose locking strategy based on contention and correctness needs:
  • optimistic locking for low-contention update races
  • unique constraints for duplicate prevention
  • pessimistic locking only when contention and correctness justify it

Spring-Specific Checks

• Verify whether @Transactional is on a proxied public entry point.
• Verify whether self-invocation bypasses the transaction boundary.
• Verify rollback rules. By default, unchecked exceptions roll back, checked exceptions may not.
• Verify whether readOnly = true is used only where appropriate.
• Verify whether REQUIRES_NEW is truly required or is masking a design issue.
• Treat NESTED as database- and platform-dependent, not a universal escape hatch.

Anti-Patterns To Flag

• @Transactional on controllers by default.
• One transaction that does database writes and then performs slow HTTP calls.
• Catching exceptions inside the transaction and converting them to success-like flows.
• Publishing irreversible side effects before commit.
• Assuming retries are safe without idempotency.
• Using a bigger propagation setting to hide unclear boundaries.

Advanced Consistency Nuances

• Distinguish duplicate prevention from concurrency control. A unique constraint solves one class of race, not lost updates or write skew.
• Remember that many integrity failures surface on flush or commit, not at the line that changed the entity. Design tests and exception handling accordingly.
• UnexpectedRollbackException often means an inner operation marked the transaction rollback-only even though the outer layer tried to return success.
• Isolation levels are database-specific in effect. The same setting on Postgres, MySQL, and SQL Server may protect different anomalies.
• Deadlock and serialization-failure retries belong at a carefully chosen outer boundary. Retrying a half-executed workflow with external side effects is dangerous.
• @TransactionalEventListener and TransactionSynchronization are phase-sensitive. Choose before-commit, after-commit, or after-rollback behavior deliberately.
• If the workflow crosses service boundaries, distinguish local transaction design from saga or orchestration design. Do not pretend one local transaction can guarantee distributed consistency.
• In reactive or coroutine transaction flows, verify which transaction manager and context propagation model is actually in use. Imperative assumptions often fail there.

Expert Heuristics

• Start from the invariant, not from the annotation. Ask what must never be observably false to users or downstream systems.
• Prefer database-enforced invariants for uniqueness and impossible states, then use application logic to make violations rare and understandable.
• If a workflow mixes command and query steps, decide whether read-your-write guarantees are required immediately or whether eventual consistency is acceptable.
• If the team wants REQUIRES_NEW, ask whether they are isolating audit logging, masking rollback behavior, or compensating for a larger design problem.

Output Contract

Return these sections:

• Consistency goal: the business invariant being protected.
• Recommended boundary: where the main transaction starts and ends.
• Propagation and isolation: only the settings that matter and why.
• Idempotency and concurrency: duplicate handling, locking, and retry safety.
• External side effects: what must happen outside the transaction or through outbox-style patterns.
• Verification: tests or scenarios that prove rollback, duplicate handling, and conflict behavior.

Guardrails

• Do not put @Transactional on every service method by default.
• Do not recommend distributed 2PC or XA unless the project already uses it and truly requires it.
• Do not ignore the cost of holding database connections during external calls.
• Do not treat duplicate prevention as an application-only concern when the database can enforce it.

Quality Bar

A good run of this skill turns a vague workflow into explicit consistency boundaries and testable invariants. A bad run decorates methods with @Transactional without modeling failure paths, retries, and side effects.