On Sun, Jul 28, 2024 at 8:40 AM Laurenz Albe <laurenz.albe@xxxxxxxxxxx> wrote:
On Sat, 2024-07-27 at 14:05 -0600, Scott Ribe wrote:
> Similar argument applies to turning off fsync, which I have found to sometimes make a
> significant difference (depending on hardware).
That's bad advice. Very bad advice.
That is, unless you are ready to delete the cluster and run a new "initdb" after an OS crash.
Which I am, if there's only one database in the cluster.
But why risk that, if you can get virtually the same positive effect by disabling
"synchronous_commit". But all that shouldn't have a big effect on "pg_restore".
To tune "pg_restore", increate "max_wal_size", "checkpoint_timeout" and "maintenance_work_mem".
I do that too.
> The other argument I've seen, that if there's a crash during restore you'll have a
> corrupted database, is bogus. What are you going to try to do with a database if there's
> a crash during restore???
Drop it?
You are wrong: it is not the database that is broken after a crash, but the entire cluster.
Maybe I'm spoiled by high-quality hardware and SANs, plus VMware, but crashes are damned rare in my environment.
I'll take that risk to restore a database faster in a single-database cluster.
