| 2 Sep 2022 |
David Arnold (blaggacao) | (which may be the main current justification) | 18:34:24 |
Linux Hackerman | In reply to @zhaofeng:zhaofeng.li
To further expand on that, consuming the existing nixosConfigurations is problematic, since Colmena needs to insert two things into the configuration: Modules (deployment.* options and configs) as well as module arguments (name and nodes which allows you to refer to configurations of other nodes in the cluster). This is in part following the design decisions of NixOps and morph. Inserting them cannot be cleanly done in case of nixosConfigurations which contains complete, evaluated configurations rather than individual NixOS modules.
We could remove this need and make the user specify the deployment.* configs externally, outside the NixOS configuration. This is the approach that deploy-rs takes (the deploy output). The downside is two-fold:
- The level of configuration composition possible will be limited. As an example, I set
deployment.tags = [ "routers" ]; in my router module so enabling it will automatically make colmena apply --on @routers deploy to this node. Additionally, I use the added nodes argument extensively in my configuration to set up a WireGuard mesh and iBGP peerings between my nodes.
- The effect Colmena can have on the target node configuration will be limited. For instance, Colmena creates the
${name}-key.service for each secret file defined in deployment.keys, which will no longer be possible. deploy-rs uses a deploy binary copied alongside the profile to implement activation and other features.
The end result is a level of functionality noticeably different from "normal" Colmena that would necessitate a lot of documentation changes and still be confusing to users.
- The
nodes argument doesn't require colmena though. nixosConfigurations can reference self.outputs.nixosConfigurations to get at other hosts.
- (covers
deployment.tags as well) I'd prefer for there to be e.g. inputs.colmena.nixosModules.default that doesn't just supply the options (as inputs.colmena.nixosModules.deploymentOptions does), but also their implementation. I don't see (feel free to point it out) what needs to be injected there that isn't available at "any" evaluation time -- with nixops's model of resources that are also managed by the tool, that makes sense, but given that colmena doesn't "create" machines or otherwise manage resources (sensible choice IMO!) it's not necessary.
| 18:35:45 |
Linux Hackerman | In reply to @zhaofeng:zhaofeng.li
To further expand on that, consuming the existing nixosConfigurations is problematic, since Colmena needs to insert two things into the configuration: Modules (deployment.* options and configs) as well as module arguments (name and nodes which allows you to refer to configurations of other nodes in the cluster). This is in part following the design decisions of NixOps and morph. Inserting them cannot be cleanly done in case of nixosConfigurations which contains complete, evaluated configurations rather than individual NixOS modules.
We could remove this need and make the user specify the deployment.* configs externally, outside the NixOS configuration. This is the approach that deploy-rs takes (the deploy output). The downside is two-fold:
- The level of configuration composition possible will be limited. As an example, I set
deployment.tags = [ "routers" ]; in my router module so enabling it will automatically make colmena apply --on @routers deploy to this node. Additionally, I use the added nodes argument extensively in my configuration to set up a WireGuard mesh and iBGP peerings between my nodes.
- The effect Colmena can have on the target node configuration will be limited. For instance, Colmena creates the
${name}-key.service for each secret file defined in deployment.keys, which will no longer be possible. deploy-rs uses a deploy binary copied alongside the profile to implement activation and other features.
The end result is a level of functionality noticeably different from "normal" Colmena that would necessitate a lot of documentation changes and still be confusing to users.
*
- The
nodes argument doesn't require colmena though. nixosConfigurations can reference self.outputs.nixosConfigurations to get at other hosts.
- (covers
deployment.tags as well) I'd prefer for there to be e.g. inputs.colmena.nixosModules.default that doesn't just supply the options (as inputs.colmena.nixosModules.deploymentOptions does), but also their implementation. I don't see (feel free to point it out if I'm missing something) what needs to be injected there that isn't available at "any" evaluation time -- with nixops's model of resources that are also managed by the tool, that makes sense, but given that colmena doesn't "create" machines or otherwise manage resources (sensible choice IMO!) it's not necessary.
| 18:35:53 |
David Arnold (blaggacao) | (I might be totally wrong, too) | 18:36:10 |
@yuka:yuka.dev | In reply to @blaggacao:matrix.org But, I don't know if that may break the paraleval I don't think that would be influenced by this kind of change | 18:38:13 |
Linux Hackerman | same | 18:38:38 |
@yuka:yuka.dev | I will try to prepare a PR that solves most of the problem without too many changes | 18:38:40 |
@yuka:yuka.dev | If that's fine with y'all | 18:38:51 |
@yuka:yuka.dev | first not even focusing on producing a nixosConfigurations, but giving a first stepping stone in that direction | 18:39:14 |
Zhaofeng Li | In reply to @yuka:yuka.dev You could just tell the users that the colmena binary has to be from the exact same version as the hive expression There still need to be some kind of versioning and/or a contract, because David Arnold (blaggacao) is making a framework (std) that will emit this structure with its own logic | 18:41:35 |
@yuka:yuka.dev | I mean, how is the current interface versioned? | 18:42:16 |
Zhaofeng Li | The current interface isn't versioned at all and the expressions are embedded into the Colmena executable | 18:43:44 |
@yuka:yuka.dev | I mean the interface between that expression and my flake | 18:44:03 |
@yuka:yuka.dev | There is still an interface, even if it's at a different level | 18:44:11 |
Linux Hackerman | (aside: an interface that I'd love to see changed, to avoid mixing the nodes with special attrs like defaults and meta!) | 18:45:07 |
Zhaofeng Li | In reply to @yuka:yuka.dev I mean the interface between that expression and my flake That's also poorly defined at the moment, documented mostly as examples. One possibility is to use the NixOS module system to define the schema, which is what NixOps may be heading towards: https://github.com/NixOS/nixops/pull/1508 | 18:47:31 |
Linux Hackerman | nixos modules for that would be great, I think that approach was pioneered by infinisil in nixus | 18:48:21 |
Linux Hackerman | defining modules that have the entire deployment as their scope instead of an individual host is 😘🤌 | 18:49:15 |
@yuka:yuka.dev | definitely, but I'd like to not mix up multiple issues/improvements too much | 18:49:49 |
Linux Hackerman | yeah. I'm also not sure that should be colmena's responsibility. | 18:50:11 |
@yuka:yuka.dev | https://github.com/yu-re-ka/colmena/commit/cd558fee7f4251719d1101f9677dce2581f6f58e | 19:19:20 |
@yuka:yuka.dev | This is enough changes to make it work for me | 19:19:34 |
@yuka:yuka.dev | In my flake, I now do:
outputs = { self, nixpkgs, colmena, ... }@inputs: {
colmenaEval = colmena.evalHive {
# whatever was in outputs.colmena before
};
nixosConfigurations = self.outputs.colmenaEval.nodes;
};
| 20:11:45 |
@yuka:yuka.dev | though somehow it does not like to do pure eval | 20:12:14 |
Zhaofeng Li | Yeah, if the git repo is dirty, nix flake metadata will not return a locked flake URI that's usable in pure evaluation mode. That's one gotcha I encountered when looking into pure eval. | 20:13:32 |
@yuka:yuka.dev | Does the assets wrapper flake solve this? | 20:14:10 |
Zhaofeng Li | Yes, the locked URI of the assets flake encompasses the inputs as well. | 20:16:01 |
Zhaofeng Li | Your solution should work if the git repo is clean | 20:16:35 |
@yuka:yuka.dev | In reply to @zhaofeng:zhaofeng.li Yes, the locked URI of the assets flake encompasses the inputs as well. also if it is a dirty git repo | 20:16:50 |
@yuka:yuka.dev | In reply to @zhaofeng:zhaofeng.li Yes, the locked URI of the assets flake encompasses the inputs as well. * also if one of the inputs is a dirty git repo? | 20:16:56 |