You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
CI/build failure: official Nix 2.34.4 installer expects a different SHA-256 than the currently served x86_64-linux tarballΒ #15564
Thank you for posting this issue! I was thinking if I did something wrong.
Also, you'll need to thank :the ai: ;)
I wouldn't be able to dig this deep into the build system without it.
(well, I could but not within work hours)
I digged in hydra logs and here is the breakdown of what I found:
The x86_64-linux tarball derivation (9piv8l3vvj5iabsjsnjyj754xcy0mfni-nix-binary-tarball-2.34.4.drv) was built on two different builders within minutes of each other. They produced different contents for the same output path x7svz9b68aj74cslr96hgkf784shxla8:
22:16 on sleepy-brown β NAR size 25,695,176 bytes (wrong tarball, hashes to f5096772...)
22:21 on elated-minsky β NAR size 25,745,752 bytes (correct tarball, hashes to 595669eb...)
The second upload to the binary cache overwrote the first, so the cache ended up with the correct tarball. But sleepy-brown still had the wrong version sitting in its local Nix store.
When the installer-script build (2ri7ilj7..., build 325163674) got scheduled on sleepy-brown at 22:31, it picked up the wrong local copy of the tarball as its input dependency. nix hash-file dutifully hashed that wrong tarball and embedded f5096772... into the install script. A second installer-script (gjgln519, on sleepy-brown at 22:28) has the same wrong hash, confirming it's builder-specific.
Installer-scripts built on other machines only hash their native platform's tarball and got correct results. You can see this in the build logs where maximum-snail (macOS) only emits one hash-file warning vs eight on sleepy-brown.
Reacted by Alexander Shestakov, Kevin Veen-Birkenbach, G and Travis A. Everett
Activity
SubaruArai commentedon Mar 26, 2026
Just FYI, it doesn't match one version earlier (2.34.3) either.
7888ca3b586d026e149e866cda9ecb228811bd98d9f56b8c58a5e9106151ffd0SubaruArai commentedon Mar 26, 2026
What's the proper way to circumvent? I.e. how can I know the proper hash or an older version's hash?
SubaruArai commentedon Mar 26, 2026
Found it.
https://hydra.nixos.org/job/nix/maintenance-2.34/installerScript/all
Build 325163674 is for 2.34.4, and it has the hash mismatch.
Build 324420758 is for 2.34.3, and the hash is correct.
This is the link to the install script for 2.34.3, for anyone that need it now:
https://hydra.nixos.org/build/324420758/download/1/install
kevinveenbirkenbach commentedon Mar 26, 2026
Thanks a lot :) Still keeping this open until a sustainable solution is implemented :)
SubaruArai commentedon Mar 26, 2026
Thank you for posting this issue! I was thinking if I did something wrong.
Also, you'll need to thank :the ai: ;)
I wouldn't be able to dig this deep into the build system without it.
(well, I could but not within work hours)
kevinveenbirkenbach commentedon Mar 26, 2026
FYI: Also raised awreness in the Matrix channel :)
xokdvium commentedon Mar 26, 2026
Somehow hydra built the wrong install script:
Tarball is from this https://hydra.nixos.org/build/325162956
This is what we uploaded too:
The install script for 2.34.4 (https://releases.nixos.org/nix/nix-2.34.4/install):
Notice that the input addressed path is the same
x7svz9b68aj74cslr96hgkf784shxla8/nix-2.34.4-x86_64-linux.tar.xz- but the hash is unexpected. https://hydra.nixos.org/build/325163674/download/1/installSo hydra calculated the wrong hash for reasons unclear. All other tarball hashes in the installer seem to be correct.
kevinveenbirkenbach commentedon Mar 26, 2026
Could you please manual trigger hydra again? May it was just an Single-event upset, if it worked until now :)
packaging: Make binaryTarball more reproducible
xokdvium commentedon Mar 26, 2026
Ah seems like the build got retried https://hydra.nixos.org/build/325162956#tabs-buildsteps on different machines. And with the tarball being non-reproducible due to an ordering issue #15565 we wound up with a franken-build.
Mic92 commentedon Mar 26, 2026
I digged in hydra logs and here is the breakdown of what I found:
The x86_64-linux tarball derivation (9piv8l3vvj5iabsjsnjyj754xcy0mfni-nix-binary-tarball-2.34.4.drv) was built on two different builders within minutes of each other. They produced different contents for the same output path x7svz9b68aj74cslr96hgkf784shxla8:
The second upload to the binary cache overwrote the first, so the cache ended up with the correct tarball. But sleepy-brown still had the wrong version sitting in its local Nix store.
When the installer-script build (2ri7ilj7..., build 325163674) got scheduled on sleepy-brown at 22:31, it picked up the wrong local copy of the tarball as its input dependency. nix hash-file dutifully hashed that wrong tarball and embedded f5096772... into the install script. A second installer-script (gjgln519, on sleepy-brown at 22:28) has the same wrong hash, confirming it's builder-specific.
Installer-scripts built on other machines only hash their native platform's tarball and got correct results. You can see this in the build logs where maximum-snail (macOS) only emits one hash-file warning vs eight on sleepy-brown.
Mic92 commentedon Mar 26, 2026
Fixed in the releases bucket directly by updating the script to contain the right hash.
packaging: Make binaryTarball more reproducible
packaging: Make binaryTarball more reproducible
packaging: Make binaryTarball more reproducible
packaging: Make binaryTarball more reproducible
packaging: Make binaryTarball more reproducible
packaging: Make binaryTarball more reproducible