Repository navigation
brew install ghPatch 2 - #1158
Open
kimtrabach-spec wants to merge 13 commits into
Open
kimtrabach-spec wants to merge 13 commits into
kimtrabach-spec wants to merge 13 commits into
Conversation
…ht-size resource defaults (TheSpaghettiDetective#1144) * fix(chart): raise tasks and redis memory defaults The post-print compile_timelapse task shells out to ffmpeg/libx264 to encode the timelapse. At the old 512Mi limit the kernel OOM-kills the ffmpeg child on every compile, and because the Celery parent (PID 1) survives the container never restarts and nothing surfaces in kubectl — prints just silently never get their timelapse. Observed steady state of the tasks container is around 700Mi with pod-lifetime peaks near 745Mi, so raise the memory limit to 2Gi (headroom above the peak), the memory request to 512Mi, and the CPU limit to a full core so the encode is not throttled for minutes. The bundled redis sits flat around 250Mi in steady state, right at the old 256Mi limit. Raise its memory request to 256Mi to schedule for the real footprint and its limit to 512Mi for headroom before the OOM killer. Bump the chart version to 0.2.0 and regenerate the README defaults. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> * ci(workflows): publish the Helm chart as a signed OCI artifact The images already publish to GHCR, but the chart shipped in this repo still tracks the moving "release" tag, so a chart install picks up whatever images the latest release build produced rather than the ones it was authored against. Publish the chart itself as an OCI artifact with the matching images baked in, so a chart release pins the exact images it was built alongside. Hoist the namespace and version computation into a dedicated version job whose outputs feed both the image builds and the chart publish. Computing the version once removes the race where a commit crossing a UTC-midnight boundary between two independent computations would yield two different values and a chart pinning images that were never pushed. The new publish-chart job needs the build job, then in a throwaway CI workspace points each image repository at the namespace the build job pushed to and pins each image to its immutable per-commit sha-<short> tag. Pinning the repository (not just the tag) keeps the chart correct whether the workflow runs on upstream or a fork; pinning the immutable sha (not the day-granular CalVer) keeps a second release commit on the same UTC day from silently moving the images a published chart references. The chart itself is versioned by the build's CalVer, with appVersion set to the same CalVer. It then packages the chart, pushes it to oci://<owner>/charts and cosign-signs the pushed digest keyless — mirroring the image signing. helm push and cosign read separate credential stores, so the job logs in both ways. Helm maps Chart.yaml sources[0] to the org.opencontainers.image.source annotation on push (since 3.12), and GHCR links a package to a repository by matching that annotation. Bake sources[0] from the running repository too, so the package links to whoever publishes (upstream or a fork) and in the canonical trailing-slash-free form GHCR matches on; the version job pins a Helm new enough for the mapping. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> * docs(chart): document image variants and Kubernetes deployment notes Add README sections for installing from the signed OCI artifact (images pinned to the build commit) versus a git checkout (moving release tag), the data and media persistence volumes with existingClaim support, and why the tasks container memory limit matters: ffmpeg timelapse encoding is OOM-killed silently below it while the pod stays healthy. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> --------- Signed-off-by: Aleksei Sviridkin <f@lex.la> (cherry picked from commit 49c0bc7)
Every image fetched its model weights from a moving branch and trusted whatever came back. An error page served with a 200, or an LFS pointer instead of the file it points to, both land on disk and ship as a model; the container then dies at import with no hint of why. Pin the RKNN sources to commits, reject a non-success status, and refuse to build if a downloaded file is too small to be what it claims. The RK3588 runtime and the toolkit that talks to it are one matched pair, so pin both to 2.3.2 rather than letting either float independently. Order the layers so editing source does not reinstall dependencies, and keep the local 390 MB darknet weights out of the build context. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la>
…iant The web and CPU inference images were published for amd64 only, so an arm64 host had to build them itself. Build both on native runners per platform, push by digest, and merge the digests into one manifest list per image so a single tag serves either architecture. Nothing is tagged until it has been proven to work: the platform legs push no tag at all, and the single-platform images now do the same, so a build that finishes but cannot load a model never owns the per-commit tag that the chart pins and the docs recommend. Add an L4T image for NVIDIA Jetson. The same Dockerfile produces it, because on arm64 its base manifest resolves to the L4T half; publishing it under its own name keeps the hardware target visible instead of hiding it in a manifest list's arm64 slot. It builds best-effort: its base is old enough that a broken upstream archive is a realistic failure, and a release run that everyone learns to ignore is worse than a missing niche image. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la>
The pages describing how to build and troubleshoot the inference image had drifted from the code: they named a script and a directory that no longer exist, gave library and weight paths the loader never prints, and pointed at a base image tag that was never published without saying it has to be built by hand first. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la>
Nothing verified a change until it was already published. The chart's unit tests and lint ran only on a contributor's laptop, the images were first built on a push to the release branch, and the two hand-built ML bases were compiled by nobody at all — which is how one of them came to fetch its runtime from a link that answers 301 with an empty body without anyone noticing. Run all of it on a pull request instead, pushing nothing: images are built and started on the runner, so this works from a fork with no credentials. The bases also run weekly, since they pull third-party artifacts from URLs this project does not control and rot without anyone touching the repository. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la>
The single-platform image job built from a Dockerfile and signed in the same place, so the token that signs releases was held for the length of a build running pip, apt and a base image from somebody else's registry. Splitting naming and signing into their own job leaves that job with layers pushed by digest and nothing else, which is the split the multi-arch path already had. Two windows that stay open are written down where they are rather than left to be found: an index and a chart can only be pushed under a name, so each exists unsigned between its push and its signature. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la>
…ive/feat/compose-prebuilt-images ci(workflows): publish the server images for arm64 and verify what ships
Media URLs are stored with an absolute scheme and host. When a self-hosted server's domain changes, previously stored timelapse, prediction, poster and event snapshot URLs keep pointing at the old host, and resign_media_urls could not repair them because new_signed_url() only replaces the query string. With --rewrite-host, absolute URLs are rewritten to the given host while being re-signed, with the scheme taken from SITE_USES_HTTPS (matching how the server builds all other full URLs). Relative URLs are left untouched. The digest is calculated from the path only, so rewriting the host never invalidates the signature. Without the flag, behavior is unchanged.
apt-get install janus fails on Trixie. Link to the blog post covering the from-source build.
…rewrite-host Add rewrite_media_url_host command for repairing media URLs after a server address change
Contributor
|
Thanks for the contribution! Could you please change this PR’s base branch from |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.