Hi,
On 2021-10-06 08:22:08 +0000, Dimitri Fontaine wrote:
> I believe we have found another bug in Postgres when using pg_auto_failover. The details can be seen at
https://github.com/citusdata/pg_auto_failover/issues/814; and the Postgres warning message to consider is the
following:
>
> WARNING: oldest xmin is far in the past
>
> When a replication slot switches from active to inactive, whatever xmin
> value that is registered on the replication slot is then kept.
That's required - otherwise the slot would e.g. stop keeping
hot_standby_feedback across the replication connection breaking.
> It seems to me that we should either document that a replication slot that
> has been active (used in streaming replication) can not be maintained
> through calls to pg_replication_slot_advance later; or better yet that this
> should be made to work, somehow.
You encountered this on a physical slot, by the sound of this? For a logical
slot we cannot just safely change xmin, but
pg_physical_replication_slot_advance() should update it.
I wonder if we optionally should do something similar in
pg_physical_replication_slot_advance(). I.e. read the WAL between the current
position and the "moveto" LSN, and see what xmin should be updated to. If we
see WAL records that would cause conflicts for an older xmin, we can update
xmin to that.
Greetings,
Andres Freund