On 2023-03-28 17:08:38 +0200, Alexander Kukushkin wrote:
> On Tue, 28 Mar 2023 at 16:55, Peter J. Holzer <hjp-pg...@hjp.at> wrote:
> 
> 
>     However, when we took down one node for about two hours for some tests
>     recently (with some moderate traffic on the remaining node), the replica
>     didn't catch up after being restarted and inspection of the logs showed
>     that it was trying to get WALs which had already been deleted.
> 
> 
> 
> It is impossible to know whether it is down temporarily or gone forever,
> therefore the slot is removed as soon as the member key expires (after TTL).
>  

TTL is 30 seconds by default. So that doesn't seem very useful.


>     So apparently, `use_slots: true` isn't enough. What else do I have to
>     configure? (I know about wal_keep_size, but it was my understanding that
>     this isn't needed when slots are used)
> 
> 
> The best option is to configure continuous archiving and PITR. Backups are
> always important.

Yeah, but for the given application a daily dump is sufficient for
backups, so I'd like to keep that simple.

> The second option - you can put all member names into permanent slots
> configuration (using patronictl edit-config):
> slots:
>   nodename1:
>     type: physical
>   nodename2:
>     type: physical
>   nodename3:
>     type: physical
> 
> This way slots representing these members will not be removed.

That seems to work (at least the slot didn't disappear within a few
minutes). I'll do some more testing.

        hp

-- 
   _  | Peter J. Holzer    | Story must make more sense than reality.
|_|_) |                    |
| |   | h...@hjp.at         |    -- Charles Stross, "Creative writing
__/   | http://www.hjp.at/ |       challenge!"

Attachment: signature.asc
Description: PGP signature

Reply via email to