CI targets for //tvix are very expensive to evaluate
#426
Opened by sterni at
I've been wondering how 🦙 has started to take ~6min. Seems like //tvix is at least partially responsible. I've taken the following measurement on whitby (!):
time nix-instantiate -E 'with import ./. {}; nix.readTree.gather third_party.nixpkgs.lib.isDerivation tvix' … real 2m58.487s user 2m32.097s sys 0m9.247s
It appears that the NixOS VM tests are mostly unproblematic, main contributor seem to be the feature variants of //tvix/store and //tvix/castore. Evaluating the only the subtargets of //tvix/build, //tvix/castore and //tvix/store takes ~1m50s on whitby.
For comparison,
- //ops (without //ops/pipelines) takes ~30s
- //users takes 1m50s
so the classic expensive sub trees that have a ton of NixOS machines easily outperform //tvix.
sterni at 2024-11-21T17·51+00