On Mon, Jan 10, 2022 at 10:29 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
"David G. Johnston" <david.g.johnston@gmail.com> writes: > On Mon, Jan 10, 2022 at 12:09 PM Dominique Devienne <ddevienne@gmail.com> > wrote: >> Tom wrote "relation" for the number of locks necessary for DROP OWNED BY. >> What does it mean in this context? relation = table?
> I'm confused here a bit as well. The items being talked about are tables > and indexes, both of which manifest as files on the filesystem. But not > all relations do (e.g., views). But if this isn't tied to the filesystem > then I would expect that other object types, especially functions, would > require locking as well, but those are decidedly not relations.
I was wrong actually --- I wrote that thinking that we acquire exclusive lock when dropping a relation (where relation may be defined as "something with a pg_class entry"). That's true, but these days we acquire a lock when deleting *any* cataloged database object. So you'd also need a lock for each schema, function, etc that was due to get dropped. This is basically to avoid problems in case of concurrent drop commands.
It's still true that the size of a relation in columns or rows is not relevant here.
Given that Tom mentions max_locks_per_transaction can be safely increased,
and given the stats I mentioned in this thread, what would a "reasonable" max_locks_per_transaction
be in my case? By reasonable, I mean "as large as possible w/o being too large"...
Obviously 64*100 is not quite large enough to be safe in this case. I'd appreciate some advise. TIA, --DD
