Skip to content

Atomic Commit — Junior

At junior level, focus on this question:

How does Two-Phase Commit (2PC) get multiple independent databases to commit (or abort) together, as one atomic unit?

The two phases

sequenceDiagram participant Coordinator participant DB1 as Participant 1 participant DB2 as Participant 2 Note over Coordinator,DB2: PHASE 1: PREPARE (voting) Coordinator->>DB1: PREPARE Coordinator->>DB2: PREPARE DB1->>DB1: locks resources, writes to its\nown log, ready to commit OR abort DB1-->>Coordinator: VOTE: yes (ready) DB2-->>Coordinator: VOTE: yes (ready) Note over Coordinator,DB2: PHASE 2: COMMIT (only if ALL voted yes) Coordinator->>DB1: COMMIT Coordinator->>DB2: COMMIT DB1->>DB1: actually commits, releases locks DB2->>DB2: actually commits, releases locks

Phase 1 (Prepare): the coordinator asks every participant "can you commit this?" Each participant does everything needed to guarantee it can commit if told to (acquiring locks, writing to its own durable log) without actually committing yet, and votes yes or no.

Phase 2 (Commit): if every participant voted yes, the coordinator tells everyone to commit for real. If any participant voted no, the coordinator tells everyone to abort instead — guaranteeing all-or-nothing across every participant.

Why "prepared" means "guaranteed to be able to commit"

A participant that votes "yes" in phase 1 is making a binding promise: it must be able to commit later, no matter what happens to it in the meantime (even a crash and restart) — this is why it must durably persist its prepared state to its own log before voting yes, using the same write-ahead-log durability guarantee from the Transactions & ACID professional page.

🎓 Takeaway: 2PC's core trick is splitting "can I commit" from "actually commit" into two separate steps, so the coordinator only issues the final commit once it knows everyone is guaranteed able to follow through — this is exactly how 2PC achieves atomicity across multiple independent systems that have no other shared coordination mechanism.

Test yourself

  1. Why must a participant durably persist its "prepared" state before voting yes, rather than just voting yes based on its current in-memory state?
  2. What happens if participant 1 votes yes but participant 2 votes no — walk through phase 2 in that case.
  3. Why can't a participant simply "change its mind" and refuse to commit after it has already voted yes in phase 1?

Continue to middle.md.