Alexander Korotkov <a.korotkov@postgrespro.ru> writes: > This topic was already discussed (at least one time) in 2011. See [1] for > details. I'd like to raise that again.
I'm a bit worried about adding a toast table to pg_class, and more so about pg_database, because both of those have to be accessed in situations where it's not clear that we could successfully fetch from a toast table, because too little of the catalog access infrastructure is alive.
pg_class is probably all right as long as only the ACL field could ever get toasted, since it's unlikely that any low-level accesses would be paying attention to that field anyway.
For pg_database, you'd have to make sure that the startup-time check of database CONNECT privilege still works if the ACL's been pushed out of line.
Thank you for pointing. I'll check this.
> Also, I've notice performance degradation of GRANT statements themselves. > 1000 GRANT statements are executed in 1.5 seconds while 10000 GRANT > statements are executed in 42 seconds. In average single GRANT statements > becomes 2.8 times slower. That's significant degradation, but it doesn't > seem to be fatal degradation for me.
Seems all right, since we could just say "we don't really recommend that usage pattern".