When using the "classic" (non-BuildKit) builder, ensure that
services are iterated in dependency order for a build so that
it's possible to guarantee the presence of a base image that's
been added as a dependency with `depends_on`. This is a very
common pattern when using base images with Compose.
A fix for BuildKit is blocked currently until we can rely on a
newer version of the engine (see docker/compose#9324)[^1].
[^1]: https://github.com/docker/compose/issues/9232#issuecomment-1060389808
Signed-off-by: Milas Bowman <milas.bowman@docker.com>
Within the Docker API/engine, networks have unique IDs, and the
name is a friendly label/alias, which notably does NOT have any
guarantees around uniqueness (see moby/moby#18864 [^1]).
During day-to-day interactive/CLI Compose usage, this is rarely
an issue, as Compose itself isn't creating networks concurrently
across goroutines. However, if multiple Compose instances are
executed simultaneously (e.g. as part of a test suite that runs
in parallel), this can easily occur.
When it does happen, it's very confusing for users and resolving
it via the `docker` CLI is not straightforward either [^2].
There's two primary changes here:
* Pass `CheckDuplicates: true` to the Docker API when creating
networks to reduce the likelihood of Compose creating duplicates
in the first place
* On `down`, list networks using a name filter and then remove
them all by ID, as the Docker API will return an error if the
name alias is used and maps to more than one network
Hopefully, this provides a better UX, since the issue should be
less likely to occur, and if it does, it can now be resolved via
standard Compose workflow commands.
[^1]: https://github.com/moby/moby/issues/18864
[^2]: https://stackoverflow.com/questions/63776518/error-2-matches-found-based-on-name-network-nameofservice-default-is-ambiguo
Signed-off-by: Milas Bowman <milas.bowman@docker.com>
This is mostly marking a bunch of the run methods as helpers so
that the internal assertions they do will show the line number of
the calling test instead.
There's also some small tweaks around the plugin initialization to
help with the output in the event that it fails to make it easier
to debug what went wrong.
Signed-off-by: Milas Bowman <milas.bowman@docker.com>
This was using `docker exec` on Compose containers instead of
`docker compose exec` (and `docker-compose exec` for standalone).
Thanks to @glours for catching!
Signed-off-by: Milas Bowman <milas.bowman@docker.com>
The E2E tests can be run in plugin (`docker compose`) or standalone
(`docker-compose`) mode. Existing logic was in place to ensure that
the helper method is always used, which will invoke the right one
based on how tests are being executed.
However, this logic was too easy to (unintentionally) bypass given
the myriad of ways that commands can be run. The check has been
made stricter and pushed to a lower-level to more aggressively
catch instances.
As a result, a bunch of calls to `RunDockerCmd` are now updated
to be `RunDockerComposeCmd`, which will ensure that the invocation
is correct based on test mode.
Signed-off-by: Milas Bowman <milas.bowman@docker.com>
Use the command `stdout` instead of combined `stdout` + `stderr`
for assertions to avoid failures from any CLI logging such as
warnings, which will be on `stderr`.
Signed-off-by: Milas Bowman <milas.bowman@docker.com>
The most important change here is to ensure that the correct Compose
standalone binary is used by `ddev`. Since it invokes Compose itself,
we need to ensure that `PATH` is set appropriately such that it finds
the binary we want to test rather than something from the system.
As part of this, the rest of the environment has been isolated, which
should make the test more reliable, and avoids polluting `~/.ddev`
with test artifacts by using a tmpdir as `HOME` for the test instead
of the user's real home folder.
Signed-off-by: Milas Bowman <milas.bowman@docker.com>
When running Docker / Compose commands, do NOT inherit the system
environment to ensure that the tests are reproducible regardless
of host settings.
Additionally, per-command environment overrides are provided to
the command instead of using `os.SetEnv`, as this is not safe when
running tests in parallel (`testing.T::SetEnv` will actually error
if used in this way!)
Signed-off-by: Milas Bowman <milas.bowman@docker.com>
The big mechanical change here is to NOT store `t` as a field on
the `CLI` object (which has been renamed as well to fix the odd
capitalization). The way the tests are structured meant that the
"subtests" were using the _parent_ `*testing.T` instance, which
causes various oddities including confusing messages on failure
about a child test causing a panic in the parent.
Additionally, a few tests have been blocked from running in
parallel because they are sharing `compose.yaml` fixtures and
can fail as a result (e.g. due to a port conflict on the host).
I'll fix these in follow-up PRs but want to focus on correctness
for the tests before optimizing them.
Signed-off-by: Milas Bowman <milas.bowman@docker.com>
Guillaume Lours
Milas Bowman
dependabot[bot]
Nick Sieger
Nicolas De Loof
Bastian Venthur
Laura Brehm