Saga Pattern: Distributed Transactions via Choreography and Orchestration

The Saga pattern is how microservices achieve consistency across databases without distributed ACID transactions.

• Two-phase commit locks rows across all participants until every database votes — at thousands of operations per second, that becomes a global mutex that kills throughput.
• Sagas take a different approach: each service executes a local transaction and publishes an event.
• If a later step fails, compensating transactions undo the effect.

The key insight about compensation:

• Compensation is not a database rollback — when payment succeeded but inventory is out of stock, you issue a refund, which is a new forward transaction that creates a credit record, notifies the payment gateway, and updates the ledger
• Compensating transactions must be idempotent because at-least-once delivery means a compensation command can arrive twice — "release reserved inventory" must be safe to call multiple times
• Every step must have a defined compensation that can be executed independently of the system's current state

Choreography vs orchestration:

• Choreography has no central coordinator — services react to each other's events; clean for 2–3 step flows, but once you have six services reacting to each other's success and failure events, the interaction graph becomes unmanageable and tracing what happened requires following events across dozens of topics
• Orchestration puts the flow in a dedicated state machine — it sends commands, tracks which steps completed, and drives compensation when something fails; you can query it at any moment and know exactly where the saga stands; for Java, Axon Framework provides this with persistent state and retry

Visibility and monitoring:

• Monitor saga duration and alert when one stays in an intermediate state beyond your SLA — the most common cause is a consumer crash with a message stuck in a dead-letter queue

I recommend choreography for simple linear flows with 2–3 services, and orchestration for anything with conditional logic, many participants, or business-critical flows where operational visibility matters.