'COPY *' has a different behavior than 'COPY .' - when regexes are
involved, COPY behave differently, by unpacking directory content to the
root.
Enhance unit test to cover the scenario as well
We used to create dockerfiles blindly assuming there is content, but
that's not the case for virtual packages.
Due to https://github.com/moby/moby/issues/38039 we are forced for a
"unpleasant" workaround, as we can't create empty FROM scratch images
and export them.
The realtime output could be configured through
LUET_GENERAL__SHOW_BUILD_OUTPUT environment
variable or related config option or through
`--live-output` option.
This commit removes the Domain Name, if any, from the cached image
reference before computing the image fingerprint. This way the same
image, if stored in some oter mirror, is still seen as the same one.
Fixes#158
CopyFile relies on copy.Copy from https://github.com/otiai10/copy which
doesn't handle named pipes copy. Handle it here until
https://github.com/otiai10/copy/issues/47 is fixed.
This causes luet to hang while copying packages that have named pipes in
it.
Also invert compression argument for gzip, it causes slowliness.
The backend will figure out if we have the image or not, otherwise will
atempt to pull if not there.
Skip retrieve integration test with img as its not supported.
This changeset also drops --keep-exported-images, which is quite unused
and can be replaced with a plugin, or either by manually exporting the
resulting images.
Several changes are included:
- Expose ensureDir in helpers, and call it in the Docker client. In
other implementations that was handled by CopyFile behind the scenes,
but that's not the case here
- Create accessor in Artifact to create Artifact objects from files.
This is handy when we have to carry over downloaded package content
into caches when artifacts are already verified
- Fix various issues around the imagePush flag, so now trees are pushed
forcefully each time
- Take into consideration the real artifact name when pushing single
files in the docker image. This behavior should be changed eventually,
because single files which aren't repository packages now are in its
own docker image, but we should have just one that brings the required
metadata alltogether.
Drop image-repository on create-repo. In case of a docker repository, --output is the image reference to use.
Also restore default output build dir.
See also: #169
This is currently not a valid use case. Virtuals are empty packages and
if the `build.yaml` is completely empty, nothing could depend on them.
Let's try to not be too smart and build the package image if a source
image is supplied, and fail hardly when we depend on a virtual in build
time.
We don't need to look after the builder image as its optional. In this
way we can also reduce the compiler options, as we don't require a
--clean flag anymore. --only-target-package is sufficient to determine
what we can skip and how.
As now we build only when necessary, we need to make sure the images are
built similarly. The discrepancies between the two are less now, and
they can share the same logic.
This fixes a regresion where when no prelude is defined, the build
context isn't copied over
Collections, similarly to packages, have a `build.yaml` and
a `finalize.yaml` that are templated for each package.
They have a `collection.yaml` containing a list of
packages that are part of the tree.
Is it more faster in this way as we already have all the needed folders
to the comparison extracted. In this way we don't repeat I/O operation
twice by calling container-diff.
Do not depend on container-diff anymore
Annotate the package build time when compiling, and use that from the
client to force upgrade of packages that changed the artifact, but
didn't changed any version.
The client can trigger this behavior with `luet upgrade --sync`
This allows to have an unique identifier for the builder image id against
different depgraphs combinations. The package fingerprint is not enough,
as an atom could have a difference deptree depending on the requires
constraints.
TODO: Don't use the full image name, but only the hash as a salt
(currently the salt contains ALSO a reference of the image-repository,
as such it doesn't allow to port a tree in a different docker registry)
Currently, it's used the archive.ReplaceFileTarWrapper
that requite a []byte of the files replaced. This is not
a good idea if files are big and instead could be better
in the near future reimplement ReplaceFileTarWrapper with
a callback that return io.Reader instead of []byte.
If a protected file is already present on target rootfs
it is created a file with the same prefix used in Gentoo:
._cfgXXXX_<filename>
