Backend Development › Transactions & Concurrency Control

Non-Repeatable Read

Reading the same row twice and getting different values.

Backend Engineer · Senior Data Engineer · Senior

Also known as: non-repeatable read, unrepeatable read, non repeatable reads

A non-repeatable read happens when a transaction reads a row, another transaction commits an update to it, and the first transaction reads it again and sees a different value — within the same transaction.

T1: SELECT balance → 100
T2: UPDATE balance = 50; COMMIT
T1: SELECT balance → 50     ← same transaction, different value

The data was committed (unlike a dirty read), but the transaction’s view isn’t stable. Any logic that reads, decides, and reads again can produce inconsistent results.

Isolation levels address it: Read Committed allows non-repeatable reads; Repeatable Read (and snapshot-based levels) prevent them by giving the transaction a stable view of the rows it reads (see MVCC).

The classic mistakes:

Doing a read-decide-read in one transaction at Read Committed. The value can change between the two reads, so your decision is based on stale data. Use a higher isolation level, or lock the row.
Confusing it with a lost update. A non-repeatable read is about reading an inconsistent value; a lost update is about overwriting one. They can happen together but are distinct.
Assuming Repeatable Read is free. Higher isolation can increase aborts (the transaction may need to fail and retry if its snapshot can’t be reconciled). It’s a trade, not a strict upgrade.
Expecting it from a single query. Within one query, the statement sees a consistent snapshot even at Read Committed in many databases; non-repeatable reads concern multiple statements across time.
Forgetting that your own transaction’s current views can shift. Even with snapshot isolation, note that a transaction may or may not see effects of writes it made — details vary by database.

How to handle it: if a transaction must see a stable value across several statements, use a snapshot-based isolation level (Repeatable Read / Snapshot) or lock the relevant rows. If you only need the current value once, Read Committed is fine. Know which reads must be repeatable, and set isolation accordingly — see isolation levels and phantom reads.