Conversation

Hannu Krosing
@PostSQL
Replying to
@samokhvalov
Congrats! Very thorough and clear article. As to why subtransactions use own txids consider that they are used for A (and indirectly D) from ACID as well. Your surprise at the end of the article considers only I.
1
1
Nikolay Samokhvalov 🇺🇦
@samokhvalov
Replying to
@PostSQL
Thanks. Yeah, I see what you're saying: if we inserted a row in a subtrans, the did ROLLBACK TO SAVEPOINT, then we need to consider such tuple as dead - but with "main XID" it would be not directly possible without extra actions.
1
1
Nikolay Samokhvalov 🇺🇦
@samokhvalov
Replying to
@samokhvalov
and
@PostSQL
This resonates with my general understanding of Postgres MVCC approach as "commit-pessimistic" (or "rollback-optimistic") -- I would better prefer having UNDO log here too, and more overhead for cases of [partial] ROLLBACK than constant overhead for committed transactions.
1
1
Peter Geoghegan
@petervgeoghegan
Replying to
@samokhvalov
and
@PostSQL
I agree with your conclusion (I think), but I still don't think that Postgres MVCC is rollback optimistic 🙂. Turns out that rollbacks used by the TPC-C benchmark contribute significantly to the disproportionate problems we see with free space management:
postgresql.org
Re: The Free Space Map: Problems and Opportunities
On Mon, Aug 16, 2021 at 5:15 PM Peter Geoghegan wrote: > This is actually quite simple -- the …
Twitter Web App
5
Likes
Hannu Krosing
@PostSQL
Replying to
@petervgeoghegan
and
@samokhvalov
Getting free space back immediately after Rollback would be nice, but you really do not want to make that part of Rollback cmd itself. (I just discovered the thread on pghackers, probabli this is already there). But also CHECKPOINT should do cleanup, freeze, etc.
1