nandi/jolt-native · 8c5607c

⑂ 8c5607c

Let the runner build locally again, just barely

max-jobs = 0 sent the fixed-output rust toolchain fetches to nixbuild,
whose sandbox has no network, so they failed; it also pushed every
trivial cargo-src derivation across an ssh round trip. Two local slots
fix both without giving the slow runner any real compiling to do.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

nandi committed 2026-09-07T19:59:09-07:00 Browse files

8c5607c parent: fccae88

modified .gitlab-ci.yml +8 -7 Viewed

@@ -19,12 +19,13 @@ check:
19	19	# rather than a shallow single commit.
20	20	GIT_DEPTH: "0"
21	21	before_script:
22		- # Everything heavy is offloaded to nixbuild.net; the runner only evaluates,
23		- # copies results back and uploads. max-jobs = 0 means it refuses to build
24		- # anything locally, so a misconfigured builder fails loudly instead of
25		- # silently falling back to the slow runner. 16 concurrent remote jobs, not
26		- # more: every in-flight goal costs the runner's single nix daemon an ssh-ng
27		- # connection plus the path copy, and at 100 the daemon OOMs mid-pipeline.
	22	+ # Compilation is offloaded to nixbuild.net; the runner evaluates, fetches
	23	+ # and uploads. max-jobs stays small but NOT zero: at zero the runner refuses
	24	+ # to build anything at all, which sends the fixed-output fetches (the rust
	25	+ # toolchain tarballs) to nixbuild, where the sandbox has no network and they
	26	+ # fail. Two local slots keep those fetches -- and the hundreds of trivial
	27	+ # cargo-src/cargo-package derivations, each otherwise a 1.5s ssh round trip
	28	+ # -- on the runner, and everything expensive still lands on the builder.
28	29	- \|
29	30	set -eu
30	31	mkdir -p ~/.ssh
@@ -37,7 +38,7 @@ check:
37	38	'experimental-features = nix-command flakes' \
38	39	'builders = ssh-ng://eu.nixbuild.net x86_64-linux - 16 1 big-parallel,benchmark' \
39	40	'builders-use-substitutes = true' \
40		- 'max-jobs = 0' >> /etc/nix/nix.conf
	41	+ 'max-jobs = 2' >> /etc/nix/nix.conf
41	42	script:
42	43	- nix flake check -L
43	44	# The store paths are read-only symlink farms; the runner uploads plain

      # rather than a shallow single commit.
      GIT_DEPTH: "0"
    before_script:
--    # Everything heavy is offloaded to nixbuild.net; the runner only evaluates,
++    # Compilation is offloaded to nixbuild.net; the runner evaluates, fetches
--    # copies results back and uploads. max-jobs = 0 means it refuses to build
++    # and uploads. max-jobs stays small but NOT zero: at zero the runner refuses
--    # anything locally, so a misconfigured builder fails loudly instead of
++    # to build anything at all, which sends the fixed-output fetches (the rust
--    # silently falling back to the slow runner. 16 concurrent remote jobs, not
++    # toolchain tarballs) to nixbuild, where the sandbox has no network and they
--    # more: every in-flight goal costs the runner's single nix daemon an ssh-ng
++    # fail. Two local slots keep those fetches -- and the hundreds of trivial
--    # connection plus the path copy, and at 100 the daemon OOMs mid-pipeline.
++    # cargo-src/cargo-package derivations, each otherwise a 1.5s ssh round trip
++    # -- on the runner, and everything expensive still lands on the builder.
      - |
        set -eu
        mkdir -p ~/.ssh
          'experimental-features = nix-command flakes' \
          'builders = ssh-ng://eu.nixbuild.net x86_64-linux - 16 1 big-parallel,benchmark' \
          'builders-use-substitutes = true' \
--        'max-jobs = 0' >> /etc/nix/nix.conf
++        'max-jobs = 2' >> /etc/nix/nix.conf
    script:
      - nix flake check -L
      # The store paths are read-only symlink farms; the runner uploads plain