The remote builder was hardcoded to use --platform linux/amd64
regardless of user input or runner architecture. This caused
performance issues on ARM runners and cache inefficiencies.
Now properly uses the platforms input or detects host architecture
to avoid unnecessary QEMU emulation and improve build performance.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Firstly this was a bug where we were trying to commit in the post
step even if we had already committed at the end of the main step in
a non-setup-only invocation.
Secondly, if the action is canceled before the exposeID is set in the main
process, we don't want to send a commit request with an empty exposeID.
Previously, we were firing off an async buildkit prune to clean
up layers unused in 14 days. This changes that to cleanup layers
unused in 7 days and fires it off inline on cleanup. It just seems
easier to reason about that way.
This setup-only mode will setup a docker builder with the stickydisk
mounted but will not run a Docker build. The use case here is to allow
customers to then run their custom Tilt files or Docker commands against
our builder. The other subtle change is that we only cleanup in the post
step of this builder action. It is still to be seen if you can start several
of these builders at the same time in a workflow but we can do that as a follow
on.
We have reason to believe that automatic GC is affecting
daemon startup times. In this patch we disable automatic GC
and instead rely on manual pruning of the buildkit cache.
Once the daemon is ready we spawn an async task to run prune
on any objects older than 14 days. We are already manaing the
ceph volume approaching its size limit ourselves in the VM
Agent.
Patch also adds some alerting when inode usage is high on a mountpoint.
Previosuly, we only killed the buildkitd process and unmounted
if builderInfo was non null. This was wrong cause we could have setup
builkdkitd, but failed after that step. This would then rely on the last
ditch effort by the post action to cleanup. We now change the proc kill
and unmount to happen on any build error.
Also increase retries when trying to unmount the buildkit directory.
Retry up to 3 seconds now, previously we were only retrying 3 times
with a 100ms backoff.
Additionally, write some tests to ensure the driver method
`startBlacksmithBuilder` handles all exceptions correctly in
both nofallback=true and nofallback=false configurations.
Previously, we were incorrectly formatting the block device
on the ceph server. This could cause corruption. As a mitigation
we now conditionally format the block device on the client. This
should only kick in if a previous parent clone has not already been
formatted.
