Transactions and isolation

you wrap the two updates in a transaction so either both land or neither does, and nobody sees the halfway state.

ACID

the process died halfway through and you needed the first update to be rolled back rather than left standing on its own.

Atomicity

you told the user it was saved, so it has to still be there after somebody pulls the power cable out.

Durability

you pick how much of other people's in-flight work your transaction is allowed to see, usually by accepting whatever the default was.

Isolation level

you read a value another transaction had written but not committed, and then it rolled back and the value never existed.

Dirty read

you only see other transactions' work once they commit, but two reads inside your own transaction can still disagree.

Read committed

you read the row at the start of the transaction, read it again at the end, and it has a different value because somebody committed in between.

Non-repeatable read

you read the same row twice inside one transaction and are guaranteed to get the same answer both times.

Repeatable read

you run the same range query twice in one transaction and the second time there is a new row matching it.

Phantom read

the database promises the result is as if the concurrent transactions had run one after another, and aborts one of them to keep that promise.

Serialisable

your transaction sees the database frozen as of the instant it began, no matter what commits while it runs.

Snapshot isolation

an update writes a new version of the row instead of overwriting it, so readers already in flight keep seeing the old one.

MVCC

two transactions each check the same rule, each sees it satisfied, each writes a different row, and together they break the rule.

Write skew

someone opened a transaction, went off to call an HTTP API, and left locks and dead row versions pinned for eight seconds.

Idle in transaction

you ask both databases to prepare, wait for both to say yes, and only then tell both to commit.

Two-phase commit

you break the cross-service operation into local transactions and write an undo step for each one, because there is no shared rollback.

Saga

you cannot roll back the payment you already captured, so you issue a refund that puts the world approximately back.

Compensating transaction