2PC Participant States — Working, Prepared, In Doubt, Decided Diagram

Two Points of No Return

Voting yes is the first: the participant writes everything to disk and gives up the right to abort
The coordinator's decision, written to its own log, is the second — the commit point of the whole protocol
Before those points anything can still abort safely; after them, commit must be enforced however long it takes
A single-node commit folds both points into one event: writing the commit record

Life While In Doubt

A prepared participant cannot decide anything on its own — timeouts do not make unilateral abort safe
Every lock the transaction holds stays held: other transactions wanting that data are blocked
A coordinator that takes twenty minutes to recover blocks that data for twenty minutes
In principle participants could talk among themselves; in 2PC that is simply not part of the protocol

The Worst Case

If the coordinator's log is lost, no automatic recovery exists — an administrator must decide by hand
Orphaned in-doubt transactions survive even database restarts, because the locks must be preserved
Heuristic decisions let a participant guess — a polite word for probably breaking atomicity
The structural fix: replicate the coordinator with a consensus protocol instead of one node's disk