Hello pg hackers,
This is the definition of the function:
SyncRepWaitForLSN(XLogRecPtr lsn, bool commit)
1. In the code, it emits ereport(WARNING) for the
ProcDiePending/QueryCancelPending case like this:
ereport(WARNING,
(errcode(ERRCODE_ADMIN_SHUTDOWN),
errmsg("canceling the wait for synchronous
replication and terminating connection due to administrator command"),
errdetail("The transaction has already committed
locally, but might not have been replicated to the standby.")));
The message "The transaction has already committed locally" is wrong
for non-commit waiting e.g. 2PC Prepare or AbortPrepare, right? so maybe we
just give the errdtail for the commit==true case.
2. I'm curious how the client should proceed for the ProcDiePending corner
case in the function (assuming synchronous_commit as remote_write or
above). In this scenario, a transaction has been committed locally on
master but we are not sure if the commit is replicated to standby or not if
ProcDiePending happens. The commit is not in a safe status from the
perspective of HA, for example if further when auto-failover happens, we
may or may not lose the transaction commit on the standby and client just
gets (and even can not get) a warning of unknown commit replication status.
Thanks.
