- We're running 12.2.12 with civit web.
- 3 RGW's with haproxy round robin
- 32 GiB RAM (handles = 4, thread pool = 512)
- We run mon+mgr+rgw on the same hardware.
Looking at our grafana dashboards, I don't see us running out of free memory, they hover around 5 GB free (but sometimes lower).
From my quick tests (a fair while ago) I found making more handles eventually lead to OOMs.
Currently, we've been getting some "connection reset by peer" (errno 104)... from client reads, which I suspect is due to the low handles setting.
We plan to move the rgws onto 125 GiB beefier machines.