On Fri, 2023-06-23 at 15:05 +0530, Postgres all-rounder wrote:
> Context: We have faced a network isolation and ended-up with locally 
> committed data on the 
> old primary database server as one of the tools that is in-place for HA 
> decided to promote
> one of the SYNC standby servers. As the PostgreSQL won't provide a HA 
> solution as in-built,
> I would like to just confirm on the behaviour of core parameter 
> synchronous_commit= remote_apply.
> 
> As per the documentation the PRIMARY database server will NOT commit unless
> the SYNC standby acknowledges  that it  received the commit record of the 
> transaction
> and applied it, so that it has become visible to queries on the standby(s), 
> and also written
> to durable storage on the standbys.

That's not true.  The primary will commit locally, but wait for the synchronous 
standby
servers before it reports success to the client.

> However, during the network outage or few scenarios where the current primary 
> is waiting
> for the SYNC to acknowledge and when the application sends a cancel signal 
> [even control +c
> from a PSQL session which inserted data]  then we see locally committed data 
> on the primary
> database server.
> 
> "The transaction has already committed locally, but might not have been 
> replicated to the standby."
> 
> 1. It appears to be a known behaviour, however wanted to understand, is this 
> considered as an
> expected behaviour or limitation with the architecture

This is expected behavior AND a limitation of PostgreSQL.

> 2. Any known future plans in the backlog to change the behaviour in 
> such a way PRIMARY won't have the LOCALLY commit data which is NOT received 
> and acknowledged
> by a SYNC standby when  synchronous_commit= remote_apply is used?

There have been efforts to use two-phase commit, but that would require 
PostgreSQL to
have its own distributed transaction manager.

> 3. If the information is available in the document that primary database can 
> have locally
> committed data when it is waiting on SYNC and receive the cancel signal from 
> the application,
> it can be helpful.

I don't think that's anywhere in the documentation.

Yours,
Laurenz Albe


Reply via email to