But what if I just forbid to change recent_global_xmin? If it is stalled at FirstNormalTransactionId and never changed? Will it protect all versions from been deleted?
That's totally impractical, you'd have unbounded bloat and a nonfunctional system in no time.
You'd need a mechanism - akin to what we have with replication slots - to set a threshold for age.
> Then there's another issue that logical replication has had to deal > with -- catalog changes. You can't start looking at tuples that have a > different structure than the current catalog unless you can figure out > how to use the logical replication infrastructure to use the old > catalogs. That's a huge problem to bite off and probably can just be > left for another day if you can find a way to reliably detect the > problem and raise an error if the schema is inconsistent.
Yes, catalog changes this is another problem of time travel. I do not know any suitable way to handle several different catalog snapshots in one query.
I doubt it's practical unless you can extract it to subplans that can be materialized separately. Even then, UDTs, rowtype results, etc...
The question is how we should handle such catalog changes if them are happen. Ideally we should not allow to move back beyond this point. Unfortunately it is not so easy to implement.
I think you can learn a lot from studying logical decoding here.
