This is based on nixpkgs commit 862277ac9d34273cd953f42061e23d488d6b7e8b.
The resulting image is less than 25 MiB compressed.
The actual crosslink closure got shrunk from ~180 MiB to ~60 MiB just by using beam_minimal rather than beam as the base set in beams.nix!
| pkgs: let | |
| beamPkgs = with pkgs.beam_minimal; packagesWith interpreters.erlangR25; | |
| erlang = beamPkgs.erlang; | |
| elixir = beamPkgs.elixir_1_14; | |
| fetchMixDeps = beamPkgs.fetchMixDeps.override {inherit elixir;}; | |
| buildMix' = beamPkgs.buildMix'.override {inherit fetchMixDeps;}; | |
| mixRelease = beamPkgs.mixRelease.override {inherit elixir erlang fetchMixDeps;}; | |
| in {inherit erlang elixir buildMix' mixRelease fetchMixDeps;} |
| { | |
| mixRelease, | |
| fetchMixDeps, | |
| lib, | |
| }: | |
| mixRelease rec { | |
| pname = "cross-link"; | |
| version = "0.0.0"; | |
| src = ../.; | |
| mixFodDeps = fetchMixDeps { | |
| pname = "${pname}-deps"; | |
| inherit src version; | |
| sha256 = "sha256-NazEvRNJydlMuLXYRXFxnQSyhCVVy4J0G6q3nDV9XAQ="; | |
| }; | |
| } |
| { | |
| inputs.nixpkgs.url = "github:nixos/nixpkgs?ref=nixpkgs-unstable"; | |
| inputs.nobbz.url = "github:nobbz/nixos-config?ref=main"; | |
| outputs = { | |
| self, | |
| nixpkgs, | |
| nobbz, | |
| }: let | |
| pkgs = nixpkgs.legacyPackages.x86_64-linux; | |
| npkgs = nobbz.packages.x86_64-linux; | |
| spkgs = self.packages.x86_64-linux; | |
| bpkgs = import ./.nix/beams.nix pkgs; | |
| callPackage = pkgs.lib.callPackageWith (pkgs | |
| // npkgs | |
| // spkgs | |
| // { | |
| inherit callPackage; | |
| inherit (bpkgs) mixRelease fetchMixDeps; | |
| }); | |
| in { | |
| inherit (nobbz) formatter; | |
| devShells.x86_64-linux.default = callPackage ./.nix/shell.nix {}; | |
| packages.x86_64-linux = { | |
| inherit (bpkgs) erlang elixir; | |
| yamlfmt = callPackage ./.nix/yamlfmt.nix {}; | |
| crosslink = callPackage ./.nix/cross_link.nix {}; | |
| image = callPackage ./.nix/image.nix {}; | |
| }; | |
| }; | |
| } |
| { | |
| dockerTools, | |
| crosslink, | |
| }: | |
| dockerTools.buildLayeredImage { | |
| name = "${crosslink.pname}"; | |
| tag = "${crosslink.version}"; | |
| contents = crosslink; | |
| } |
$ nix build nixpkgs\#beam_minimal.interpreters.erlang_25
$ nix path-info ./result -Sh
/nix/store/0b6ywhdaa7q5w7n8xnprl2sy0amgkzpz-erlang-25.3.2.4 337.4M
This is indeed much beyond what I would expect and I will need to bisect that in more detail later today.
Oh, wait!
This is as expected!
The "minimal" Erlang is indeed ~300 to 350 MiB in its closure size ("full" is about twice), only the final applications closure will actually benefit from the partially copied OTP. As that then will only copy relevant OTP-apps.
In hindsight, using beam_minimal and beam_nox is not to remove wx, as the application will be built anyway, but no hard linking to the wx libraries happens.
The actual reason for using beam_minimal is to exclude systemd from the closure, which would be kept referenced in the final closure otherwise due to it being "wrapped" into the EPMD for which I have so far not found a way to exclude.
So for checking the size of your deployment, please build the final closure and check its size, not the intermediate erlang artifact, for which we take steps to get rid of it anyway.
If the final closure though still exceeds your expectations, further debugging has to be done on that.
And another one: I briefly checked your mix.exs, which loads the :runtime_tools, which again require the :wx application.
Remove that and your final app will not include wx anymore.
Thanks for the advice! Makes sense about the elimination of erlang. I see that the elimination does happen in the closure, but when dockerising I see the erlang stuff return with a vengeance. I'll dig a bit deeper and report!
As I have trouble following discussions on a gist, I'd suggest to move any further questions to the official discourse, which I can more actively watch, and where your problem is also more visible to other users.
I usually do not care for the size of individual folders in the store, but closure sizes only.
Still the mentioned differences make me wonder where and why that is happening and whether this might be a regression.
I am currently trying to check the closure size of the
beam_minimal.interpreters.erlang_25, but have to build that one first, which can take up to an hour on my weak machine.If I can confirm this variancy I will definitely take a closer look into what is going on. The commit linked in the
README.mddefinitely gave me the mentioned closure size of ~60MiB which not only included erlang, but also elixir and my deployment subject (at that point a freshly generated phoenix application).