Isolation governs how concurrent transactions interact and which changes they can observe. Full isolation makes their outcome equivalent to some serial execution. Databases can relax isolation guarantees to improve throughput, so the isolation level matters when business logic depends on a consistent view of data. section 9.1.1 section 9.3
Queries need isolation even when they do not modify data. Statement-level isolation gives an individual SQL statement a consistent snapshot; transaction-level isolation can give multiple SELECTs a consistent snapshot across the transaction. Without an appropriate guarantee, a query can combine old and new values that never existed together. section 12.1.3
A saga spans multiple local ACID transactions but has no global isolation. Other requests can observe changes between saga steps, and compensation can overwrite a concurrent update. Preventing local anomalies within each service is not enough to prevent anomalies across services. section 9.3 section 9.3.1
Concurrent sagas can produce several anomalies:
- Lost updates: one saga overwrites another's update without detecting the conflict. section 9.3.1
- Dirty reads: a transaction or saga reads changes from a saga that has not completed. section 9.3.1
- Nonrepeatable reads and phantom reads: repeated reads return different values or matching rows after concurrent changes. section 9.3.1
- Write skew: sagas read overlapping data and update different data in a way that violates a business invariant. section 9.3.1
Concurrent order sagas
A create-order saga creates an order, reserves credit, and schedules delivery; a cancel-order saga cancels the order, releases credit, and cancels delivery. An unlucky interleaving can overwrite the order state and release credit that was never reserved. Another can temporarily increase available credit, allowing an order to succeed when the customer has insufficient credit. section 9.3.1
Use local ACID transactions for operations confined to one service. For sagas, identify shared data and invariants that concurrent steps can affect, then choose application-level countermeasures. Options include semantic locks, commutative updates, and reordering steps to minimize business risk. Sagas reduce runtime coupling compared with classic distributed transactions, but require explicit compensation and careful isolation design. section 9.4 section 9.2.4














