On Fri, Dec 20, 2019 at 11:07 AM Andrey Borodin <x4mmm@yandex-team.ru> wrote:
> I think changing synchronous_standby_names to some available standbys will resume all backends waiting for
synchronousreplication.
> Do we need to check necessity of synchronous replication in any other case?
The GUCs are not re-checked in the main loop in SyncRepWaitForLSN, so
backends will remain stuck there even if synchronous replication has
been (temporarily) disabled while they were waiting.
I do agree with the general sentiment that terminating the connection
is preferable over sending a response to the client (except when
synchronous replication was already disabled). Synchronous replication
does not guarantee that a committed write is actually on any replica,
but it does in general guarantee that a commit has been replicated
before sending a response to the client. That's arguably more
important because the rest of what the application might depend on the
transaction completing and replicating successfully. I don't know of
cases other than cancellation in which a response is sent to the
client without replication when synchronous replication is enabled.
The error level should be FATAL instead of PANIC, since PANIC restarts
the database and I don't think there is a reason to do that.
Marco