Roman Gershman

Thank you for bringing this issue to our attention! We're always looking for ways to improve Dragonfly. If you're interested in helping out, we'd greatly appreciate contributions in the following...

Another thing, Dragonfly does not have `maxmemory-policy` flag, instead it has [cache_mode flag](https://www.dragonflydb.io/docs/managing-dragonfly/flags#--cache_mode) for that.

@skrlance would you like to send a PR?

> Also, I forgot to mention, we should be removing: > > #Restart=always > > From: /usr/lib/systemd/system/dragonfly.service > > Or, dragonfly will crash and won't start. why is that?

@kissingtiger have you had a chance to update the version ? we won't be able to help with fixing v1.20.1

How long should should I wait? We had a case in production where server had hundreds of thousands of such sockets which translated to gigabytes of memory `Slab` pages.

I added sleep after the loop but the sockstat still shows me over 100K sockets that do not go down. 

@axboe please observe the run on 6.8 kernel (ubuntu 24.04). https://github.com/user-attachments/assets/075f8385-d8fd-4933-ade3-23c70e0cbb8f

strace shows that the calls are consistent with my pseudo code: ``` .... [pid 1808295] io_uring_register(4, IORING_REGISTER_FILES_UPDATE, {offset=1, fds=[6]}, 1) = 1 [pid 1808295] io_uring_register(4, IORING_REGISTER_FILES_UPDATE, {offset=1, fds=[-1]}, 1) =...

@axboe I can confirm that your code does not reproduce the problematic behavior. I will continue investigating until I find the differentiator that triggers it.