Re: deferred writing of two-phase state files adds fragility

On Wed, Dec 4, 2024 at 12:19 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> > You can't execute COMMIT
> > TRANSACTION or ROLLBACK TRANSACTION, so there's now way to resolve the
> > prepared transaction.
>
> Could we fix it so ROLLBACK TRANSACTION removes the GID from the
> list of prepared xacts that need to be written out?  Then we'd
> no longer have a pending requirement to read the broken WAL record.

That would be nice, but I'm not sure that it's possible. As currently
implemented, FinishPreparedTransaction() always reads the two-phase
state data either from the two-phase file or the WAL, whether it's
committing or rolling back. One might expect the commit or rollback to
proceed purely on the basis of in-memory state, but I think that does
not work because nsubxacts might be greater than
PGPROC_MAX_CACHED_SUBXIDS. Even when it isn't, we have no shared
memory record of abortrels or abortstats.

--
Robert Haas
EDB: http://www.enterprisedb.com