Re: Native Logical Replication Initial Import Qs

On 6/7/23 2:12 PM, Don Seiler wrote:
> On the logical replication front, the concern is with the initial data
> import that happens when the subscription is created (by default). I
> know that you can tell the subscription to not copy data and instead use
> pg_dump and a replication slot snapshot to achieve this manually.
> However I'm unable to explain (to myself) why this is better than just
> having the subscription do it upon creation. Given that I can create
> pub/sub sets for individual tables for parallel operations, I'm curious
> what advantages there are in using pg_dump to do this import.

FWIW, I think the place this feature shines the most is when you can
safely leverage things like storage-level snapshots. Sometimes that
means you can get a copy of a multi-TB database almost instantly if the
storage or filesystem does copy-on-write, for example database lab
(postgres.ai) which uses ZFS.

Another thing I can think of is that while it's true you can create
multiple pub/sub sets, I'm not sure you can reduce the number of sets
later. So if you were concerned about having too many slots doing
decoding on the source, then you might want the flexibility of pg_dump
(or perhaps restoring a backup) to get more parallelism while having
more control over how many slots will be used later.

In your case, the whole setup is hopefully temporary, so maybe these
particular concerns aren't as relevant to you.

This is just what comes to mind... probably there's a few more things
I'm not thinking of and hopefully others will chime in. :)

-Jeremy

--
http://about.me/jeremy_schneider