Presumed abort
Presumed abort is a two-phase commit optimization in which a recovery coordinator with no transaction state answers an inquiring prepared participant with abort.
In R*, an abort decision can be forgotten without awaiting acknowledgements. Abort records need not be forced to stable storage; commit still requires durable decision tracking until updating participants acknowledge it. This is not a no-logging protocol. [1]
A read-only leaf can report READ VOTE and leave without a prepare log record or a second-phase decision message. No writes at that participant is different from a yes vote promising to make writes durable. [1]
The missing-state rule is a response from the responsible recovery coordinator, not permission for a prepared participant to abort after a timeout. R* explicitly retains blocking when that coordinator cannot be reached. [1]
In historical JBoss EAP 5 WS-AT, after restarting, the coordinator distinguishes logged decisions and active transactions from unknown ones. A waiting participant resends its prepared message; an unknown transaction receives an abort instruction. Commit records are retained until every participant confirms completion. [2]
The documentation treats a participant's unilateral timeout decision as a separate heuristic path. A conflicting outcome must be reported and reconciled, not reported as atomic success. Recovery callbacks can also run more than once across a crash, so completion handling needs replay-safe behavior. [2]
Do not infer this decision from a missing PostgreSQL pg_twophase file. PostgreSQL 16 may retain short-lived prepared transactions in shared memory and WAL, with files added when they span checkpoints. pg_prepared_xacts exposes prepared transactions, not the external coordinator's complete outcome history. [3]
Relations
- related
- Two-phase commit, Stable storage
Limitations
- The protocol assumes recoverable coordinator state. Deleting or losing its decision log is not equivalent to the legitimate absence of an entry. [1, 2]
- When R* chooses between presumed abort and presumed commit per transaction, the chosen protocol must accompany recovery inquiries. The same missing-state observation has different meanings. [1]
Citations
- Mohan, Lindsay and Obermarck. Transaction Management in the R* Distributed Database Management System. TODS 11(4), 1986, sections 2.2, 3, 5.
- Red Hat JBoss Enterprise Application Platform 5: Transactions Development Guide, chapter 23, WS-AT coordinator and participant crash recovery.
- PostgreSQL 16: Two-Phase Transactions; WAL, checkpoint files and prepared-transaction inspection.
Sources & caveats
- Primary-source synthesis, not an executed recovery experiment or a throughput measurement. JBoss EAP 5 is a historical implementation example, not a claim about every current transaction manager.
- Operational inference: distinguish authoritative absence, unavailable communication and lost recovery state before interpreting a missing entry.