Read skew is a concurrency anomaly in which a transaction reads multiple rows that are individually valid but collectively inconsistent because the rows were read at different points in time. section 6.9.7
Bank-account transfer
Suppose Alice has $100 and Bob has $0. One transaction reads both balances while another transfers $100 from Alice to Bob. If the reads occur across the transfer, the first transaction can combine Alice’s value from one state with Bob’s value from another, producing an overall state that never existed consistently. section 6.9.7
Read skew matters because it exposes the difference between row-level validity and transaction-level consistency:
- Snapshot isolation prevents it by giving the transaction a consistent database snapshot from the transaction’s start, isolating its reads from concurrent changes. section 6.9.7
- Repeatable reads and snapshot isolation prevent fuzzy reads, but they use different techniques: repeatable reads lock data, whereas snapshot isolation reads from a consistent snapshot. section 6.9.7
- Serializability lists read skew alongside dirty reads, fuzzy reads, lost updates, write skews, and phantom reads, helping compare isolation guarantees with performance costs. section 6.9.9
