Notestransactions🌱 seedling
Distributed Transaction
Also: Distributed Transactions · Cross-Shard Transaction · Atomic Commit · Atomic Commitment
A transaction whose writes span several nodes (shards, databases or services) and must still be atomic: every participant commits or every participant aborts. It needs an atomic commit protocol, usually [[Two-Phase Commit]].
Why it exists
On one node, atomicity is one durable commit record. Across nodes there are several logs and nothing ties their outcomes together: commit on one shard, fail on the other, and the commit can't be taken back because others may already have read it.
How it works
- Atomic commit: no participant commits until all are known to be able to. Two-Phase Commit does this with a prepare round and a logged decision.
- Non-blocking variant: run 2PC over Consensus (Paxos or Raft groups), so the decision survives a coordinator crash.
- Avoid it: pick shard keys so related rows share a shard; coordinate services with a saga and outboxes instead.
Trade-offs
| Approach | Gains | Costs |
|---|---|---|
| 2PC | atomic and isolated | blocking on coordinator loss; locks held across round trips |
| 2PC over consensus | non-blocking | every phase is a consensus round |
| saga | no cross-service locks | no isolation; compensations to write |
| co-locate data | no protocol at all | constrains the data model and shard key |
In practice
- Spanner, CockroachDB run 2PC over Paxos/Raft; cross-shard transactions are serializable by default.
- DynamoDB
TransactWriteItems: up to 100 actions in one Region; two underlying writes per item. - Vitess defaults to best-effort
MULTI(partial commits possible); atomicTWOPCis opt-in.
Pitfalls
- Treating a Timeout as "it failed": the other side may have committed.
- A shard key that splits rows usually written together turns every write into a distributed transaction.
Connections
- Part of
- Transactions
- Contains
- Two-Phase Commit
- Learn first
- Partitioning
- In real systems
- SpannerCockroachDBDynamoDB
Recall
These cards are also in your review deck, scheduled with spaced repetition.