pá 10. 2. 2023 v 21:18 odesílatel Andres Freund <andres@anarazel.de> napsal:
Hi,
On 2023-02-10 21:09:06 +0100, Pavel Stehule wrote: > Just a small note - I executed VACUUM ANALYZE on one customer's database, > and I had to cancel it after a few hours, because it had more than 20GB RAM > (almost all physical RAM).
Just to make sure: You're certain this was an actual memory leak, not just vacuum ending up having referenced all of shared_buffers? Unless you use huge pages, RSS increases over time, as a process touched more and more pages in shared memory. Of course that couldn't explain rising above shared_buffers + overhead.
> The memory leak is probably not too big. This database is a little bit > unusual. This one database has more than 1 800 000 tables. and the same > number of indexes.
If you have 1.8 million tables in a single database, what you saw might just have been the size of the relation and catalog caches.