This regression was introduced in 616fb3e55c,
with the node resolution refactor.
The core issue here is just that we would unconditionally set the
solve opt's platform to the default current platform, which was
incorrect. We can prevent this easily by having a special case for the
default case, like we had before, and then not setting the platforms
field on this (which keeping the resolution behavior which was
introduced).
Signed-off-by: Justin Chadwell <me@jedevc.com>
This patch reworks and updates the node resolution logic for selecting a
node from a builder.
The new implementation reworks the logic to make use of containerd's
platforms.Matcher interface instead of manually associated strings, and
additionally provides a few behavioural changes over the original
implementation, namely platforms can be matched with non-strict
semantics. e.g. i386 builds can be scheduled on an amd64 node, arm/v6
builds can be scheduled on an arm/v7 node.
We also add a new collection of tests for tracking regressions and
making the intended behaviour clearer.
Signed-off-by: Justin Chadwell <me@jedevc.com>
Previously, we would fail while trying to use the global "target" field
when using a matrix. The contents of the matrix really don't matter for
this.
What was happening was that we would copy the "target" property into the
child evaluation context, so that when it was updated on the parent, it
wouldn't propagate to the child. The correct behavior here is to avoid
copying variables from the target evaluation context if it is the
root.
Signed-off-by: Justin Chadwell <me@jedevc.com>
Without this, we attempt to overwrite the previous display in the lines
above which have been filled with logs from the monitor (see when used
with `--invoke`).
Signed-off-by: Justin Chadwell <me@jedevc.com>
This extracts the same logic for parsing annotations from the imagetools
create command, and allows the same flags to be attached to the build
command.
These annotations are then merged into all provided exporters.
Signed-off-by: Justin Chadwell <me@jedevc.com>
Also, tidy up the error printing, so that now we always print out the
"current commit information was not captured by the build" message,
instead of just for not locating the git binary.
Before:
WARNING: buildx: git was not found in the system. Current commit information was not captured by the build
After:
WARNING: current commit information was not captured by the build: git was not found in the system: <error message>
Signed-off-by: Justin Chadwell <me@jedevc.com>
Fixes the race condition where two boots are executed simultaneously
across multiple processes.
We initially check to see if the container exists, but if during
container creation we get a name conflict, we don't treat this error as
a hard failure, and instead move immediately into waiting for the node
to boot.
Signed-off-by: Justin Chadwell <me@jedevc.com>
On drivers that do not support multi-platform builds (the default
`docker` driver), we do not support building attestations (unless using
the containerd store).
We need to check this feature before attempting to build using
attestations.
Also adds a test to ensure that attestations can be pushed to registries
at all, and that it adequately fails on the docker driver.
Signed-off-by: Justin Chadwell <me@jedevc.com>
No other parts of the progress rendering modify the inputs, so we should
avoid this as well.
This actually fixes an edge case in pushWithMoby which writes the same
VertexStatus multiple times, modifying the timestamps and similar.
However, if the operation takes long enough the small time difference
can accumulate, and move the Start time far into the past.
Signed-off-by: Justin Chadwell <me@jedevc.com>
This native implementation uses GRPC level waiting, instead of starting
a busy loop. We also a manual max backoff of one second to improve
responsiveness.
Signed-off-by: Justin Chadwell <me@jedevc.com>
This means that we can run our docker-container tests in parallel again,
which can help speed up our test runs by a *significant* amount.
Signed-off-by: Justin Chadwell <me@jedevc.com>
We need to resolve the strip the cwd:// prefix before attempting to
resolve the dockerfile. Otherwise, we'll get the cwd:// prefix in the
dockerfile name, which isn't stripped out later.
Signed-off-by: Justin Chadwell <me@jedevc.com>
ReadLocalFiles should allow passing the stdin file as an argument, which
allows us to read from dockerCli.Stdin() to be consistent with other
commands in the same package.
Signed-off-by: Justin Chadwell <me@jedevc.com>
Fixed in 12b6a3ad9a, but now we have
regression tests! So we can add a check that we don't break this
behavior again.
Signed-off-by: Justin Chadwell <me@jedevc.com>
- Adds a new GitServeHTTP function to start an http server to serve a
target git repository.
- Adds a new GitDir helper method to get the path to the .git
directory
- Updates the GitAdd method to take a variable number of files
Signed-off-by: Justin Chadwell <me@jedevc.com>
Now, we always perform the full solve request in the main gateway call.
This ensures that progress works properly, and makes the lifetime
semantics much clearer.
NewResultContext abstracts the details of a successful/failed build, to
always return a single ResultContext, even though the details of how a
gateway is created is different:
- For a failed build, we can just keep the gateway open.
- For a successful build, we immediately open another gateway and
re-evaluate the build definition in that gateway. This should give an
instant cache hit (since the build was just successful).
Signed-off-by: Justin Chadwell <me@jedevc.com>
Subrequests have been included in docker/dockerfile:1.5 labs, so we can
update the fallback to point to this release.
Signed-off-by: Justin Chadwell <me@jedevc.com>
This was missing, since the driver property can only be fully populated
after loading nodes from disk. So we add logic to load the nodes, and
check for an error, which ensures that the "docker" driver is always
correctly present in the progress description.
Signed-off-by: Justin Chadwell <me@jedevc.com>
Justin Chadwell