nandi/jolt-native · 33a5195

⑂ 33a5195

Give CI a realistic timeout and real local parallelism

The cold pipeline hit GitLab's default 1h wall while still vendoring, so
raise it to 3h. max-jobs goes from 2 to nproc: crane emits a per-crate
derivation for each of ~800 crates, and trickling those through two
slots (or dispatching them, at a serial ~1.5s ssh round trip each) is
most of the wall clock.

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

nandi committed 2026-09-07T21:01:46-07:00 Browse files

33a5195 parent: 8c5607c

modified .gitlab-ci.yml +11 -7 Viewed

@@ -14,18 +14,22 @@ stages: [check]
14	14	check:
15	15	stage: check
16	16	image: nixos/nix:latest
	17	+ # A cold run has to vendor ~800 crates and compile the workspace; the default
	18	+ # hour is not enough for that even with the builder doing the compiling.
	19	+ timeout: 3h
17	20	variables:
18	21	# The flake is fetched from the checkout, so the runner needs the git tree
19	22	# rather than a shallow single commit.
20	23	GIT_DEPTH: "0"
21	24	before_script:
22	25	# 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.
	26	+ # and uploads. max-jobs must not be zero: at zero the runner refuses to
	27	+ # build anything at all, which sends the fixed-output fetches (the rust
	28	+ # toolchain tarballs, every crate source) to nixbuild, where the build
	29	+ # sandbox has no network and they can only fail. Giving it nproc keeps
	30	+ # those fetches and crane's hundreds of trivial cargo-src/cargo-package
	31	+ # derivations local and parallel -- each one costs a serial ~1.5s ssh
	32	+ # round trip if dispatched -- while the expensive builds still go remote.
29	33	- \|
30	34	set -eu
31	35	mkdir -p ~/.ssh
@@ -38,7 +42,7 @@ check:
38	42	'experimental-features = nix-command flakes' \
39	43	'builders = ssh-ng://eu.nixbuild.net x86_64-linux - 16 1 big-parallel,benchmark' \
40	44	'builders-use-substitutes = true' \
41		- 'max-jobs = 2' >> /etc/nix/nix.conf
	45	+ "max-jobs = $(nproc)" >> /etc/nix/nix.conf
42	46	script:
43	47	- nix flake check -L
44	48	# The store paths are read-only symlink farms; the runner uploads plain

  check:
    stage: check
    image: nixos/nix:latest
++  # A cold run has to vendor ~800 crates and compile the workspace; the default
++  # hour is not enough for that even with the builder doing the compiling.
++  timeout: 3h
    variables:
      # The flake is fetched from the checkout, so the runner needs the git tree
      # rather than a shallow single commit.
      GIT_DEPTH: "0"
    before_script:
      # Compilation is offloaded to nixbuild.net; the runner evaluates, fetches
--    # and uploads. max-jobs stays small but NOT zero: at zero the runner refuses
++    # and uploads. max-jobs must not be zero: at zero the runner refuses to
--    # to build anything at all, which sends the fixed-output fetches (the rust
++    # build anything at all, which sends the fixed-output fetches (the rust
--    # toolchain tarballs) to nixbuild, where the sandbox has no network and they
++    # toolchain tarballs, every crate source) to nixbuild, where the build
--    # fail. Two local slots keep those fetches -- and the hundreds of trivial
++    # sandbox has no network and they can only fail. Giving it nproc keeps
--    # cargo-src/cargo-package derivations, each otherwise a 1.5s ssh round trip
++    # those fetches and crane's hundreds of trivial cargo-src/cargo-package
--    # -- on the runner, and everything expensive still lands on the builder.
++    # derivations local and parallel -- each one costs a serial ~1.5s ssh
++    # round trip if dispatched -- while the expensive builds still go remote.
      - |
        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 = 2' >> /etc/nix/nix.conf
++        "max-jobs = $(nproc)" >> /etc/nix/nix.conf
    script:
      - nix flake check -L
      # The store paths are read-only symlink farms; the runner uploads plain