Kea may clean the runtime directory when starting (or maybe systemd does
it). I ran into this issue when restarting Kea after changing its
configuration, so I think the fact it normally doesn't clean it is a
race condition (it's cleaned on service start, and normally all Kea
services start at roughly the same time).
The previous implementation works fine when the plugins do not already
contain store paths, which is the case for stuff from munin-contrib.
However, for plugins generated via nix (e.g. with writeShellScriptBin),
it tries to fix the paths in it which already point to the nix store,
ruining everything.
If extraAutoPlugins contains values that carry context (e.g. it comes
from a flake input), the keys generated from them using baseNameOf
inherit that context and the config doesn't compile.
This doesn't actually need to be an attrset anyways, so a bit of
internal refactoring lets us fix this without changing the visible API.
nixosTests.forgejo: test backup/dump service; nixos/forgejo: pass {env}`GIT_PROTOCOL` via ssh to forgejo; nixosTests.forgejo: test git wire protocol version
Otherwise the tests will fail with `networking.useNetworkd = true;`
because `systemd-resolved` ignores invalid hostnames in `/etc/hosts`
(which is where all hosts from the `nodes`-attribute set end up) and
subsequently e.g. `ssh server_lazy` will fail because the name cannot be
resolved.
In d6e84a4574 the test-framework was
changed to replace all dashes with underscores of hostnames in the
python code to have readable hostnames that are valid. I.e.
nodes.foo-bar = {}
represents a host with a valid hostname and it can be referenced in the
`testScript` with `foo_bar`.
Applying this here fixes the test for both scripted networking and
networkd.
when using the host's openssh service (not the builtin golang one).
This enables the use of the much faster and more efficient wire protocol
version 2.
See https://git-scm.com/docs/protocol-v2
This should allow us to catch issues regarding that in the future.
nixos/gitea had an issue with the dump service recently, which didn't
affect us, fortunately.
But to be fair, it only affected non-default-y setups.
Not something we are able to catch in the current, rather simple, config
of our test.
Still, I see a lot of value adding this new subtest to our test suite.
Anyhow, this patch also exposes the resulting tarball as test (build)
output, which is a nice addition IMHO, as it allows some sort of
external sanity-check, if needed, without running the test interactive.
The current state is certainly very wrong - testing ZFS only on i686.
I suspect it was a typo (?) in commit 2de3caf011.
The current practical problem is that the test fails,
though in a part that looks cross-platform (which adds confusion):
https://hydra.nixos.org/build/239290208#tabs-buildsteps
