| 29 Feb 2024 |
|  kip93 joined the room. | 14:33:52 |
| 1 Mar 2024 |
|  patka joined the room. | 08:10:16 |
|  fgaz joined the room. | 09:27:14 |
| 2 Mar 2024 |
|  nh2 joined the room. | 01:42:07 |
| nh2 | raitobezarius: OK I joined. Copying my message from the other channel, for context for others:
You were looking at dedicated-hosting the binary cache regarding the AWS cost sink. I run 1 PB of CephFS clusters on Hetzner, and can set that up quite easily with NixOps. Do you want to team up on this topic?
My feeling is we could Host the 500 TB of binary cache on 6x SX134 servers (960 TB raw), with EC 4+2, which provides 580 TB HA storage. With 10 Gbit/s Internet.
For 6 * 245 = 1470 EUR/month.
To transfer out, with AWS Snowball: If we can content-deduplicate the 500 TB by factor 2x to 250 GB (using attic or bupstash as shown on https://github.com/NixOS/nixpkgs/issues/89380) the one-off cost to ship to Germany would be ~12k EUR or so.
| 01:46:11 |
 raitobezarius | Yes, so the problem is that the Foundation doesn't have 1470EUR/mo | 01:46:27 |
| raitobezarius | Even if we offsetted the Snowball, it's further unclear | 01:46:48 |
| raitobezarius | (plus, discussing with Hetzner to do the operations) | 01:47:01 |
| raitobezarius | (BTW, infra people just removed NixOps — finally — :D — but I guess anything modern as deployment framework can work) | 01:47:28 |
| nh2 | raitobezarius: what is the latest discourse post that shows the current spending on hosting and how much of that AWS gives us for free? | 01:47:38 |
| raitobezarius | In reply to @nh2:matrix.org
raitobezarius: what is the latest discourse post that shows the current spending on hosting and how much of that AWS gives us for free? cache is 15K USD/mo | 01:47:51 |
| raitobezarius | 9K of it is free credit from AWS | 01:47:57 |
| raitobezarius | 6K of it is out of our own pocket | 01:48:02 |
| nh2 | So I guess "the Foundation doesn't have 1470EUR/mo" is referring to "... on top of the AWS spending, vs instead of it"? | 01:48:48 |
| raitobezarius | I'd need to jump again in the finances to look at all of that | 01:49:37 |
| raitobezarius | but I don't think we can support it on the long term that easily | 01:49:44 |
| raitobezarius | even out of AWS spending | 01:49:50 |
| raitobezarius | the big problem also is that running the operations to have a self hosted cache is currently difficult | 01:50:02 |
| raitobezarius | so it would have to be on the top of AWS spending to avoid disasters | 01:50:10 |
| raitobezarius | there's a double problem that is | 01:50:32 |
| raitobezarius | (a) AWS offers a "nice" durability metric that cannot be reproduced with a Hetzner setup | 01:50:45 |
| raitobezarius | (b) Infra people may not have the bandwidth to touch this in the next months | 01:50:55 |
| raitobezarius | (c) Foundation may not have the money/energy, when it will come to it, to go for that solution | 01:51:11 |
| raitobezarius | (actually a triple problem) | 01:51:16 |
| raitobezarius | Another solution I am personally pursuing is just to get 1PB myself and put it in a non-super-available colo but almost free colo and duplicate the data | 01:52:46 |
| nh2 | I mean for the 5k/month difference you can literally hire Ceph experts full time to run the cluster for you | 01:54:25 |
| raitobezarius | yeah but 6K/mo not being sustainable means that even by throwing 5k/mo to hire Ceph experts full time, this won't be a viable solution, no? | 01:54:48 |
| raitobezarius | Right now, the trajectory the Foundation is taking is to launch garbage collection to reduce those 6K/mo | 01:55:19 |
| nh2 | In reply to @raitobezarius:matrix.org
yeah but 6K/mo not being sustainable means that even by throwing 5k/mo to hire Ceph experts full time, this won't be a viable solution, no? Until the data grows another 2 years, when it's 10k/mo | 01:55:23 |
| raitobezarius | In reply to @nh2:matrix.org
Until the data grows another 2 years, when it's 10k/mo Well, given the current growth pattern, it won't need 2 years | 01:55:41 |
