Two-Phase Commit — Promises, Commit Point, In Doubt Diagram

Voting yes is the first point of no return: the participant promises it can commit under any circumstance — the right to abort is surrendered
The coordinator's decision, written to its own log, is the commit point; after it, commit must be enforced however many retries it takes
Single-node atomic commit packs both points into one event: writing the commit record to the log
If a participant crashes after voting yes, it must commit on recovery — a promise cannot be taken back

A single commit request to all nodes can succeed on some and fail on others; committed data cannot be retracted once others have read it
2PC is a blocking protocol: after voting yes, a participant can only wait for the coordinator to recover
Three-phase commit needs bounded network delay and bounded response times, which real systems do not have, so it cannot guarantee atomicity
The practical fix is a replicated coordinator built on a consensus protocol instead of a single-node log

While in doubt, the transaction holds its row locks; a coordinator that takes 20 minutes to recover blocks that data for 20 minutes
Orphaned in-doubt transactions can sit forever if the coordinator log is lost — rebooting the database will not free the locks
Heuristic decisions let a participant guess; the name is a euphemism for probably breaking atomicity
XA makes the application process part of the durable state and the single point of failure; internal distributed transactions replicate the coordinator instead