On Tue, Jan 23, 2018 at 08:49:53PM +0100, Laurenz Albe wrote:
> When a primary with replication slots gets reset with pg_rewind,
> it keeps the replication slots.
>
> This does no harm per se, but when it gets promoted again,
> the replication slots are still there and are in the way.
> Won't they also hold back the xmin horizon?
>
> I think that pg_rewind should delete all replication slots.
Definitely agreed, but replications slots are not the only thing to
worry about. There has been a recent discussion on the matter:
https://www.postgresql.org/message-id/20180122201258.GT2416@tamriel.snowman.net
And the conclusion leads to the fact that we want the list of
directories and files that should be excluded from a base backup to
extracted into a header that pg_rewind could use for its own purpose as
it shares a lot in its logic with basebackup.c, and because using a
logic based on an exclusion list is more portable in the long run.
(I do remove pg_replslot's contents after running pg_rewind in my own
failover scripts by the way.)
--
Michael