One request can contain several commit boundaries
A partial commit occurs when a business operation changes several systems but only some of those changes become durable. One button or API request does not place a database, queue, payment service and notification provider inside one atomic boundary. Rolling back a local database transaction does not automatically reverse an external action that already completed. List each destination’s success condition, persistence boundary, query capability and cancellation semantics. Separate the business wish for all-or-nothing behavior from the guarantees available in the actual components. This makes intermediate states explicit and allows the design to state which states require waiting, recovery or compensation.
Trace database-plus-event dual writes
For a hypothetical dual write, save an order in a database and then publish its event. Failure after database commit can leave a durable order with no downstream notification. Reversing the order leaves a different possibility: an event exists although the database change later fails. This demonstrates the gap rather than reports a production incident. Reordering two independent commits does not combine them into one atomic operation. Preserve the business-operation identifier and the status of each stage so that confirmed, attempted and pending work can be distinguished. The useful question is not only which call returned an error, but which effects remain after that error.
Keep uncertain outcomes distinct from failure
A timeout does not prove that the remote system did nothing. The remote effect may have succeeded while its reply was lost, leaving the caller uncertain. Retrying unconditionally under a new identifier can duplicate a charge or notification. Use stable operation identity, supported idempotent retry or status queries according to the destination’s contract. Distinguish a transport failure from a definitive business rejection in logs and responses. Hiding every partial success behind one generic failure leaves the user and operator without a safe recovery strategy. Keep uncertain outcomes as a separate state until evidence resolves them rather than converting uncertainty into a convenient assumption of rollback.
Choose an outbox, transaction protocol or compensation
A transactional outbox stores the business update and the intention to publish within one database transaction, addressing the gap between those two records. Actual delivery remains asynchronous and consumers still need duplicate and ordering handling. A distributed transaction protocol instead requires participating resources and a coordinator whose persistence and failure behavior must be understood. Saga compensation is a business action, not erasure of history by an atomic rollback. A refund or reservation release may not perfectly restore every earlier consequence. Choose according to allowed intermediate visibility, recovery time, participant support and the reversibility of external effects. No pattern name supplies those business decisions automatically.
Include retries and reconciliation in completion
Test failures immediately before and after each commit, lost replies, sender restart, duplicate delivery and failed compensation. Verify what the user sees as success, pending or failure. During recovery, reconcile destinations by operation identity and either resume missing work or apply the specified compensation. Monitor both the count of partial operations and the age of the oldest unresolved case rather than only overall error rate. Retries need limits, isolation and an operator path when automatic progress cannot continue. Completion should mean confirmed business effects or an explicit recoverable state, not merely that the original request was accepted by the first component in the chain.