Costs for GitHub Actions

[zimbatm]

Somethign that might be interesting to @wolfgangwalther and the rest of the CI team. I wasn't sure exactly what the best channel was to share those numbers.

Note: the costs are discounted at 100% from GitHub.

[infinisil]

I also just highlighted this to some others. We should definitely offer GitHub (and others like AWS) who are sponsoring us in non-financial ways the corresponding perks of the new sponsorship tiers: https://nixos.org/sponsorship/

@Ra33it0 should keep track of this kind of thing

[wolfgangwalther]

Thanks for sharing these numbers. The Actions Usage Metrics are visible for all of us (maintainers? committers? not sure exactly), but the Metered Usage is not, I think.

Here's a summary of how the actions usage developed over the last year:

Month	Total minutes	Total job runs	New PRs
2024-07	169,700	142,933	6,923
2024-08	164,270	136,388	6,570
2024-09	152,030	124,772	6,192
2024-10	223,768	168,185	6,293
2024-11	353,685	223,418	6,959
2024-12	799,372	427,054	8,307
2025-01	780,516	416,463	7,676
2025-02	725,851	390,231	6,809
2025-03	895,154	495,004	8,370
2025-04	948,555	472,216	7,556
2025-05	916,811	413,981	8,957
2025-06	836,074	311,404	7,914
2025-07	910,827	329,171	7,932
2025-08	810,647	335,599	8,370

In November 2024, Eval was moved into GHA, which caused the huge increase. At the same time, we can see that the number of new PRs is increasing over time as well.

In May and June of this year some of the work I did on GHA reduced the number of jobs we ran, which is visible. It seems like it did have a brief effect on the total CI minutes as well, but by now those numbers are back to where they were before.. probably because we run more stuff now: parse checks for multiple Nix versions, more formatters via treefmt, more of the build jobs for the manual etc.

I expect quite a reduction in total minutes for August, though, when we switch Eval to Nix 2.30 - see the numbers here. A rough calculation gives me up to 100k minutes that this could save us.

cc @NixOS/nixpkgs-ci

[wolfgangwalther]

A rough calculation gives me up to 100k minutes that this could save us.

I don't think I ever estimated something more on point than that. Pretty much exactly 100k minutes less in August compared to July - despite having more PRs and running more jobs. Updated the table above.

[Mic92]

With recent memory optimizations that will be part of nix 2.32, we should be further able to increase the chunk size. Nix-eval-jobs might also be able to speed up things a bit instead of nix-env. I was also playing around with a bump allocator in nix-eval-jobs instead of boehmgc to speed up things further ( 10% on nixpkgs on top of the 10% from using the nogc option in boehm).

[wolfgangwalther]

With recent memory optimizations that will be part of nix 2.32, we should be further able to increase the chunk size.

My experiments have shown that increasing the chunk size doesn't help performance. Thus, we are running at a much lower-than-possible chunk size anyway.

[wolfgangwalther]

Closing since the number have been seen by the relevant parties.