To make the logic a bit more clear, when snapshotting files, the
parent dirs are now snapshotted in a different loop from the files we
are actually trying to snapshot. Unfortunately this loop is nearly
duplicated but I did managed to group some fo the related logic
together:
- A function to check if the file should be snapshotted (e.g. isn't
whitelisted, etc.)
- Created a `Tar` type to handle some of the logic around tar-ing, e.g.
tracking hardlinks and stat-ing files before adding them
One side effect of this is that now when snapshoting the file system,
files will be stat-ed twice.
Kaniko uses mtime (as well as file contents and other attributes) to
determine if files have changed. COPY and ADD commands should _always_
update the mtime, because they actually overwrite the files. However it
turns out that the mtime can lag, so kaniko would sometimes add a new
layer when using COPY or ADD on a file, and sometimes would not. This
leads to a non-deterministic number of layers.
To fix this, we have updated the kaniko commands to be more
authoritative in declaring when they have changed a file (e.g. WORKDIR
will now only create the directory when it doesn't exist) and we will
trust those files and _always_ add them, instead of only adding them if
they haven't changed.
It is possible for RUN commands to also change the filesystem, in which
case kaniko has no choice but to look at the filesystem to determine
what has changed. For this case we have added a call to `sync` however
we still cannot guarantee that sometimes the mtime will not lag, causing the
number of layers to be non-deterministic. However when I tried to cause
this behaviour with the RUN command, I couldn't.
This changes the snapshotting logic a bit; before this change, the last
command of the last stage in a Dockerfile would always scan the whole
file system and ignore the files returned by the kaniko command. Instead
we will now trust those files and assume that the snapshotting
performed by previous commands will be adequate.
Docker itself seems to rely on the storage driver to determine when
files have changed and so doesn't have to deal with these problems
directly.
An alternative implementation would use `inotify` to track which files
have changed. However that would mean watching every file in the
filesystem, and adding new watches as files are added. Not only is there
a limit on the number of files that can be watched, but according to the
man pages a) this can take a significant amount of time b) there is
complication around when events arrive (e.g. by the time they arrive,
the files may have changed) and lastly c) events can be lost, which
would mean we'd run into this non-deterministic behaviour again anyway.
Fixes#251
In #251 we are investigating test flakes due to layer offsets not
matching, this change will give us a bit more context so we can be sure
which image has which number of layers, and it will also include the
digest of the image, since kaniko always pushes images to a remote repo,
so if the test fails we can pull the digest and see what is up.
Also updated reproducible Dockerfile to be built with reproducible flag,
which I think was the original intent (without this change, there is no
difference between how `kaniko-dockerfile_test_copy_reproducible` and
`kaniko-dockerfile_test_copy` are built.
To allow contributors to run the integration tests with their own GCS
buckets and image repos (since not all contributors will have accesss to
the projects used by the kaniko maintainers) this updates the
integration tests so that these can be provided on the command line.
This allows tests to be run individually, without using `make
integration-test`. Previously, part of the test setup was done
in the shell script (creating the context tarball that is required
for the tests that build images with context). Instead it will be
done in the test iself, so we can use `go test` to run tests
individually if we want to.
If we are running only one individual test, we don't want to build
all of the images, so this commit creates a builder which tracks which
images it has built and can be used by a tests to check if it should
build an image before running, or it will use the images that have
already been built by a previous test.
The name of the context tarball has also been made unique (it includes
the unix timestamp) to avoid potential test flakes if two tests using
the same GCS bucket run simultaneously.
* added switch to extract different sources as build context
* first rough implementation of aws s3
* added buildcontext package and interface
* added GetBuildContext func to buildcontext.go
added fallback to gcs
renamed GC struct to GCS
* improved the default behavior of build context retrieval
* renamed gc:// to gs:// in order to follow common standards
* renamed struct File to Dir and some cleanup work
* moved context.tar suffix to the buildcontext processors where it is needed
* added buildcontext retrieval as struct variable
added fallback if prefix in bucket specifier is present
* cleanup if structures
* added prefix to s3
* WIP
* Fixed build context bugs
* refactored build context
* adding reproducible test
* newer version of go-containerregistry
* new ImageOptions
* switch reproducible flag to default to false
* small fixes
* update dep
* adding metadata tests back to integration tests and fixing resulting bugs
* fix onbuild and default env
* removing old test files
* adding the ArgsEscaped boolean on CMD commands
* fix onbuild test
* ignore failing test until container-diff is fixed
* code comments
* adding todo to remove uncomment failing test
Christie Wilson
priyawadhwa
Christian Jantz
Sharif Elgamal