How do transactions work in NoSQL databases?

A single-document update is usually atomic because the document is one storage unit. Multi-document transactions require coordination across partitions and are therefore more expensive. Some systems support them fully, some only within a partition, and some offer compare-and-set instead.

Mechanisms you will meet

• MongoDB multi-document transactions with sessions and snapshot isolation
• DynamoDB TransactWriteItems and TransactGetItems for all-or-nothing operations across items
• Cassandra lightweight transactions using Paxos for compare-and-set on a partition
• Conditional writes, which are often enough to prevent lost updates without a full transaction

const session = client.startSession();
try {
  await session.withTransaction(async () => {
    await accounts.updateOne(
      { _id: "acct_a", balance: { $gte: 100 } },
      { $inc: { balance: -100 } },
      { session }
    );
    await accounts.updateOne(
      { _id: "acct_b" },
      { $inc: { balance: 100 } },
      { session }
    );
  });
} finally {
  await session.endSession();
}

Transactions are not free. They hold locks or reservations, increase latency and can fail under contention, so retries must be idempotent. Keep them short and touch as few items as possible. If a workflow spans minutes or external systems, a saga or outbox is usually a better fit than a long transaction.