Re: A few nuances about specifying the timeline with START_REPLICATION - Mailing list pgsql-hackers

On 18/06/2021 22:55, Jeff Davis wrote:
> On Fri, 2021-06-18 at 21:48 +0300, Heikki Linnakangas wrote:
>> On 18/06/2021 20:27, Jeff Davis wrote:
>> We could teach it to look into the timeline history to find the
>> correct
>> file, though.
> 
> That's how recovery_target_timeline behaves, and it would match my
> intuition better if START_REPLICATION behaved that way.
> 
>> If the client asks for a historic timeline, the replication will
>> stop
>> when it reaches the end of that timeline. In hindsight, I think it
>> would
>> make more sense to send a message to the client to say that it's
>> switching to a new timeline, and continue streaming from the new
>> timeline.
> 
> Why is it important for the standby to be told explicitly in the
> protocol about timeline switches?

So that it knows to write the WAL to the correctly named WAL segment. 
You could do it differently, looking at the 'xlp_tli' field in the WAL 
page headers, or watching out for checkpoint records that change the 
timeline. But currently the standby (and pg_receivewal) depends on the 
protocol for that.

> If it is important, why only for historical timelines?

Well, the latest timeline doesn't have any timeline switches, by 
definition. If you're connected to a standby server, IOW you're doing 
cascading replication, then the current timeline can become historic, if 
the standby follows a timeline switch. In that case, the replication is 
stopped when you reach the timeline switch, just like when you request a 
historic timeline.

>> Hmm, the timeline in the START_REPLICATION command is not specifying
>> a
>> recovery target timeline, so I don't think "latest" or "current"
>> make
>> much sense there. Per above, it just tells the server which timeline
>> the
>> requested starting point belongs to, so it's actually redundant.
> 
> That's not very clear from the docs: "if TIMELINE option is specified,
> streaming starts on timeline tli...".
> 
> Part of the confusion is that there's not a good distinction in
> terminology between:
>     1. a timeline ID, which is a specific segment of a timeline
>     2. a timeline made up of the given timeline ID and all its
> ancestors, terminating at the given ID
>     3. the timeline made up of the current ID, all ancestor IDs, and all
> descendent IDs that the current active primary switches to
>     4. the set of all timelines that contain a given ID

Agreed, that's a bit confusing.

> It seems you are saying that replication only concerns itself with #3,
> which does not require a timeline ID at all. That seems basically
> correct for now, but since we already document the protocol to take a
> timeline, it makes sense to me to just have the primary serve it if
> possible.
> 
> If we (continue to?) allow timelines for replication, it will start to
> treat the primary like an archive. That might not be quite what was
> intended, but could be powerful. You could imagine a special archive
> that implements the replication protocol, and have replicas directly
> off the archive, or maybe doing PITR off the archive.

True.

- Heikki