| 7 Mar 2024 |
 raitobezarius | In reply to @whentze:matrix.org
oh, so the plan is not to self host the CDN in some very distant long term super future where Fastly may drop us | 00:27:22 |
 edef | In reply to @whentze:matrix.org
oh, so the plan is not to self host the CDN yeah, not in the short term anyway | 00:27:29 |
| edef | i'd like to have a more worked out story for "what if Fastly dies / quits on us" but i'm broadly satisfied with our risk exposures there | 00:28:01 |
 @delroth:delroth.net | In reply to @linus:schreibt.jetzt
we also have the advantage of the cache not being super critical for nix/nixos to be usable I think it's part of the prerequisite work for self hosting to actually determine answers to those questions (at least to some level of consensus). My personal opinion is different from yours there - I think that (recent, not historical) cache unavailability is very visible and very impactful for nixpkgs users. | 00:28:28 |
| edef | doesn't mean we can't diversify if it's easy or cheap | 00:28:33 |
| edef | In reply to @delroth:delroth.net
I think it's part of the prerequisite work for self hosting to actually determine answers to those questions (at least to some level of consensus). My personal opinion is different from yours there - I think that (recent, not historical) cache unavailability is very visible and very impactful for nixpkgs users. i mean, we can always turn it off for a bit if we want to find out | 00:28:51 |
 Wanja Hentze | the ol' yank 'n' yell | 00:29:51 |
| edef | in part by how reliable it has been, we've created this dependence | 00:30:41 |
| edef | like, if we exceed our SLOs by far, we ought to blink it | 00:31:26 |
| edef | Hyrum's law holds always and everywhere, even for availability | 00:33:08 |
| Wanja Hentze | so the thing to self host would be large, robust, highly available but not necessarily screaming fast (object?) storage? | 00:33:31 |
| edef | yeah | 00:33:41 |
| edef | we can always slap on more cache, our hot set is relatively small | 00:33:59 |
| Wanja Hentze | slaps roof of fastly | 00:34:15 |
| raitobezarius | actually as cache stakeholders are around here | 00:34:22 |
| raitobezarius | what is the process / is it fine / is it bad to download the last 2-3 releases of each channels directly? | 00:34:36 |
| raitobezarius | over Fastly | 00:34:38 |
| raitobezarius | or does it thrash the CDN to your opinion? | 00:34:44 |
| raitobezarius | I don't have the data | 00:34:45 |
| raitobezarius | should I do S3 Requester Pays stuff for that | 00:35:04 |
| edef | Fastly is fairly okay about scans, but it would be nice if you hit the bucket with requester pays | 00:35:08 |
| raitobezarius | okie | 00:35:29 |
| raitobezarius | i might bother you later on with some queries | 00:35:33 |
| raitobezarius | (or you should teach me a bit your clickhouse workflow and i will mimic it so i become autonomous) | 00:36:27 |
| edef | if you want a lot of stuff, we can plausibly pull and dedupe it for you EC2-side | 00:36:27 |
| raitobezarius | i'm basically working on a project to enable a small 2x10Gbps appliance to serve offline "fresh" nixpkgs/nixos | 00:36:52 |
| raitobezarius | and in that sense, i need to do regular syncs with last N releases of nixos | 00:37:47 |
| edef | In reply to @whentze:matrix.org
so the thing to self host would be large, robust, highly available but not necessarily screaming fast (object?) storage? to elaborate on this more: we basically have three tiers. frontend cache, semi-hot fast-enough storage, cold storage / DR | 00:37:51 |
| raitobezarius | (actually i mostly need latest release) | 00:37:53 |
| edef | hmm, okay | 00:38:01 |
