Обсуждение: max_wal_senders
I've been thinking about the max_wal_senders parameter lately and wondering if there is any harm in setting it too high. I'm wondering if I should try to shave a few senders off, perhaps to match my logical replicas + 1, instead of just leaving it at the default of 10. Or vice-versa, can clients use more than one sender if they are available? Would increasing it result in lower latency? The documentation is a little vague.
The documentation mentions an orphaned connection slot that may take a while to time out. How can I tell if I have any of those? I was looking for a `pg_wal_slots` table similar to the `pg_replication_slots` table, but don't see anything obvious in the catalog.
On Wed, 2023-02-08 at 18:07 -0500, Rick Otten wrote: > I've been thinking about the max_wal_senders parameter lately and wondering if there > is any harm in setting it too high. No, there isn't, except that if you end up having too many *actual* WAL senders, it will cause load. A high limit is no problem as such. > The documentation mentions an orphaned connection slot that may take a while to time out. > How can I tell if I have any of those? I was looking for a `pg_wal_slots` table > similar to the `pg_replication_slots` table, but don't see anything obvious in the catalog. The view is "pg_stat_replication", but you won't see there if an entry is abandoned before PostgreSQL does and terminates it. You can set "tcp_keepalived_idle" low enough so that the kernel will detect broken connections early on. Yours, Laurenz Albe
Hi, On 2023-02-09 06:59:53 +0100, Laurenz Albe wrote: > On Wed, 2023-02-08 at 18:07 -0500, Rick Otten wrote: > > I've been thinking about the max_wal_senders parameter lately and wondering if there > > is any harm in setting it too high. > > No, there isn't, except that if you end up having too many *actual* WAL senders, it > will cause load. A high limit is no problem as such. That's not *quite* true. The downsides are basically the same as for max_connections (It's basically treated the same, see InitializeMaxBackends()): You need more shared memory. There's a small degradation of performance due to the increased size of some shared datastructures, most prominently the lock table for heavyweight locks. Greetings, Andres Freund