| 19 Jan 2026 |
 Mic92 | Also because they have to be lexical sorted, no? | 14:56:23 |
 roberth | It's the simple thing to do, and simple is fast, usually. Not being simple is what stalled https://github.com/NixOS/nix/pull/4154 | 14:57:32 |
| roberth | Could be revisited, but it does change the language semantics so that's tricky, and hard to backtrack away from in the future | 14:58:10 |
| Mic92 | Yeah, feels hard to make this bug-compatible with the current behaviour. | 14:58:51 |
| Mic92 | One could have an infinite attrset that would work with new nix, but not with old nix. | 14:59:24 |
| Mic92 | Also that would be less likely the biggest issue | 14:59:51 |
| roberth | It'd be subtler current errors that would be the problem, mostly infinite recursions that would be lifted in the new Nix language | 15:01:10 |
| roberth | (or at least those would be the ones people would complain to me about :) ) | 15:01:50 |
 Taeer Bar-Yam | Generally turning errors into correct behaviour is the least bad way to break compatibility, right? Though it causes more problems now that we have other forks of nix | 15:07:26 |
 raboof | yeah I wasn't saying we should change this, just calibrating my expectations ;) - but nice to see more people have been pondering (and even hacking on) this ;) | 15:07:34 |
| roberth | no, raboof, you are in the rabbit hole with us now! | 15:08:04 |
| roberth | I agree it's a good kind of change, but the second order effects are real. E.g. you upgrade, you use it, I don't, it fails for me, I complain, we lose | 15:09:02 |
| roberth | not to say we can't get over that. It's still an improvement, but is it the right thing to do now? | 15:10:05 |
| roberth | I'd want to know the long term impact on evaluator performance including opportunities missed. tricky. | 15:10:47 |
| roberth | maybe it's a cognitohazard and we should just ignore it for another year | 15:12:57 |
| raboof | yeah, performance-wise it's hard to predict the trade-off between the benefit of not needing to evaluate the first parameter vs the cost of delaying when it's needed after all - AFAICT infinisil's work seems to suggest it's not a huge benefit (if at all). | 15:13:54 |
| Taeer Bar-Yam | Also would this change solve anything for anyone or improve anything? I don’t think I’ve ever run into a situation that needed it, personally | 15:13:58 |
| roberth | It could solve a portion of infinite recursions that necessitate lazyAttrsOf instead of attrsOf in module system use. Actually it might make that use case worse because definition merge order is deterministic but arbitrary in the module system. | 15:16:08 |
| roberth | It would speed up nix-env -q because // { type = "derivation"; } could just be the last thing in every mkDerivation (or similar) call. | 15:16:58 |
| Mic92 | A different rabbit hole is improving narinfo lookup in our current binary cache protocol by adding an efficient index that can be downloaded (incrementally): https://github.com/NixOS/rfcs/pull/195 For now kalbasit is implementing this in his binary cache proxy. And I am planning on matching this with my binary cache implementations, but at some point it would be also useful for Nix to have something like this. Currently it's in concept stage and we want to test with some implementations how it performs. It would be the dream if we also had this in cache.nixos.org, but it would be also useful otherwise i.e. if you have many small scoped caches and don't want to reach out to each of them all the time. It unlocks features like efficient p2p binary caches. | 15:17:35 |
 magic_rb | exciting times | 15:18:08 |
| roberth | If I were to shepherd that RFC I'd basically require cache.nixos.org integration before completing the RFC | 15:18:51 |
| roberth | If it doesn't work for that (almost) what's the point | 15:19:01 |
| roberth | * If it doesn't work for that what's the point (almost) | 15:19:55 |
| Mic92 | Not really, required. It doesn't even need to be accepted by the nix community as long as there are compatible implementations. | 15:19:57 |
| Mic92 | But it's nicer if everyone use the same standard of course | 15:20:29 |
| Mic92 | Right now cache.nixos.org is a rather big dataset that is hard to experiment with. | 15:22:14 |
| roberth | It is the most worthwhile though | 15:22:32 |
| roberth | Most worthwhile per byte even, I would argue ;) | 15:22:42 |
| roberth | But do ignore me if I'm just a whiny bitch | 15:23:19 |
