`sync` system call triggers a full page cache sync which may not always
work, especially in kubernetes environment where it is easy to be
interfered by others. I have seen several cases where a broken nfs mount
is blocking kaniko from doing its job.
With `syncfs`, it only writes cache back to disk for the current
filesystem that is used by kaniko which is supposed to be more reliable.
If a non-empty directory gets replaced with something other than a
directory (e.g. file or symlink), the files in that directory also get
deleted. However, there should not be any whiteout files for them in the
layer. If there were, the resulting tar file would not be extractable.
Fixes#2308
* fix(ci): Bump golangci-lint to 1.51.1
* chore(lint): fix gofmt and goimport issues
* chore(lint): fix linter issues
- Adapted error comparison according to linter recommendation
- Disabled noctx linting for http request where canceling makes no sense
- Disabled nilerror linting where nil error is returned on purpose
- Disabled makezero linter where slice is explicitly deepcopied
* chore(ci): Update go version in tests workflows
* fix(ci): Allow boilerplate years from 2000-2099
Previously the regex only allowed the copyright notice to contain the
years 2018,2019,2020,2021, or 2022. This commit widens to regex to
20\d\d allowing any year in the range [2000-2099]
* feat(ci): Replace minikube with k3s for intregration tests
The existing setup for minikube is very complicated, replicating most of
the setup steps for a full kubernetes cluster in an only partially
supported minikube configuration (driver=none). Furthermore the existing
setup has been broken for sometime, likely, at least in part due to the
changes to CNI and CRI in recent kubernetes versions.
Since what we actually need is only a running Kubernetes cluster on the
node and access to a registry on localhost:5000, we can switch the
extremely complicated minikube setup for a lightweight cluster using
k3s. Minikube came with a default addon for running a registry on every
node, but the same is not the case for k3s, instead we make use of the
package helm controller and its HelmChart CR to deploy twuni/docker-registry.helm
and expose it on localhost using the integrated LoadBalancer controller.
* fix(test-684): pin base container version
The dockerfile for the regression test connected to issue 684 used a
rolling tag as base image, making it flaky and fail since it was
introduced.
This commit pins the base image to the digest of bionic-20200219, which,
based on the date of the commit that introduced to the dockerfile would
be the most newest ubuntu build and likely what the "rolling" tag
resolved to back then. Since this also an image from the pre-oci days of
ubuntu, this circumvents a bug in container-diff as well
(https://github.com/GoogleContainerTools/container-diff/issues/389)
* Write parent directories to tar before whiteout files
Fixes#1149
The OCI image spec does not specify this order but it's a good idea and Docker
does the same.
When manually comparing layers created by Docker and Kaniko there are still
some differences (that container-diff does not show):
* Kaniko adds / to layers
* For `mkdir /test`, docker adds `/test` and an opaque whiteout file
`/test/.wh..wh..opq`. Kaniko only adds `/test/` (and /).
* snapshot_test: cleanup
Fix typos and use listFilesInTar() where possible
* fix: Correct flatten function in layers
- Added a test.
- Cache current image, track deletes in `whiteouts` as well as normal adds in `layers`.
- Fix ugly delete behavior of `layerHashCache`.
Delete it when crerating a new snapshot.
- Slight cleanup in `snapshot.go`.
- Format ugly `WalkFS` function.
* fix: Add symbolic link changes to Hasher and CacheHasher
* fix: Better log messages
* fix(ci): Integration tests
* fix(ci): Add `--no-cache` to docker builds
* fix(ci): Pass credentials for error integration test
* np: Missing .gitignore in `hack`
* np: Capitalize every log message
- Correct some linting.
* fix: Key function
- Merge only last layer onto `currentImage`.
* fix: Remove old obsolete `cacheHasher`
The directory created by `T.TempDir` is automatically removed when the
test and all its subtests complete.
Reference: https://pkg.go.dev/testing#T.TempDir
Signed-off-by: Eng Zer Jun <engzerjun@gmail.com>
This allows the value to be determined on the fly, which supports consumers that use Kaniko snaphot as a library and may need to change the value of config.KanikoDir
When deleting files of previous layers, the whiteout files
were not added to the tar file in a consistent order.
This change adds a stable sorting to the whiteout files and
adds unit tests to check for stable sorting.
During a snapshot, when a file changed and not its parent directories,
the parent directories weren't added to the layer. This is inconsistent
with Docker's behavior which always add parent directories to the layer.
In some edge-cases, it could lead to problems with docker considering
that parent directories where owned by root in forthcoming layers
although they shouldn't (see #1163).
Also, Docker seems to be POSIX compliant regarding the name of
directories in the archive, which always have a slash appended. This
commit also fixes this.
Fixes#1163
When using cache, the rootfs may not have been extracted. This prevents uname/gname from resolving
as there is no /etc/password or /etc/group. This makes this layer unnecessarily differ from
a cached layer which does contain this information. Omitting these should be consistent with Docker's
behavior.
filesToAdd is sorted in TakeSnapshotFS, but not here. This makes ordering unpredictable within the layer's tarball,
causing the SHA to differ even if layer contents haven't changed
Manish Giri
zhouhaibing089
Andreas Fleig
Joël Pepper
Gabriel Nützi
Eng Zer Jun
Tejal Desai
Rhianna
Patrick Barker
Josh Chorlton
Joe Kutner
Christopher Hlubek
Tejal Desai
Gábor Lipták
Gilbert Gilb's
cvgw
tinkerborg
Thomas Bonfort