πͺ’ Open source agent evals & observability: Trace, evaluate, and improve LLM applications with one open platform.
TypeScript analyticsautogenevaluationlangchainlarge-language-modelsllama-indexllmllm-evaluationllm-observabilityllmopsmonitoringobservabilityopen-sourceopenaiplaygroundprompt-engineeringprompt-managementself-hostedycombinator
ci(preview): build and push the AI gateway image (#18697)
Tobias Wochinger4ceff5fa21ff
diff --git a/.github/workflows/preview-build.yml b/.github/workflows/preview-build.ymlindex 7919cf87c6..b7f6693cf0 100644--- a/.github/workflows/preview-build.yml+++ b/.github/workflows/preview-build.yml@@ -1,22 +1,21 @@name: AWS preview build-# Build the PR's web + worker images and push them to ECR as pr-<n>-<sha>.-# The Argo CD ApplicationSet in the infrastructure repo deploys whichever image-# tag exists for a labeled PR; this workflow only builds and pushes, then posts-# the preview URL on the PR β as the marker-tagged status comment, as a native-# GitHub deployment (the "View deployment" button), and as a machine-managed-# block pinned atop the PR description. All PR deployments-# share one GitHub environment; Argo CD still owns the real preview lifecycle.+# Build the PR's web, worker, and AI gateway images and push them to ECR as+# pr-<n>-<sha>. The Argo CD ApplicationSet in the infrastructure repo deploys+# all three tags for a labeled PR. This workflow does not wait for the gateway+# build before posting the preview URL: that URL is the web app, and a gateway+# failure must not replace it. This workflow posts that URL on the PR β as the+# marker-tagged status comment, as a native GitHub deployment (the "View+# deployment" button), and as a machine-managed block pinned atop the PR+# description. All PR deployments share one GitHub environment; Argo CD still+# owns the real preview lifecycle.#-# The two images are independent artifacts pushed to two independent ECR repos-# under the same tag, with no ordering dependency between them. They therefore-# build in PARALLEL: a `meta` job resolves the tags and posts the "building"-# comment once, a matrix `build` job builds web and worker concurrently on-# separate runners (so each gets a full runner's CPU/memory rather than sharing-# one β their heavy stages, turbo prune/pnpm install/turbo build, are-# scope-specific and share no layer cache anyway, so serializing them bought-# nothing), and a `notify` job posts the single success/failure comment once-# both builds finish. Wall-clock is now ~max(web, worker) instead of the sum.+# Web and worker are independent artifacts pushed to independent ECR repos+# under the same tag. They build in PARALLEL: a `meta` job resolves the tags+# and posts the "building" comment once, a matrix `build` job builds web and+# worker concurrently, and a `notify` job posts the single success/failure+# comment once those two finish. The gateway builds beside them in its own+# job. Wall-clock for the preview URL is ~max(web, worker).## Builds every SAME-REPO PR on open/update (write access is the gate β opening a# same-repo PR requires push access). It deliberately does NOT trigger on the@@ -74,6 +73,7 @@ jobs:commit_sha: ${{ steps.meta.outputs.commit_sha }}web_image: ${{ steps.meta.outputs.web_image }}worker_image: ${{ steps.meta.outputs.worker_image }}+ ai_gateway_image: ${{ steps.meta.outputs.ai_gateway_image }}preview_host: ${{ steps.meta.outputs.preview_host }}namespace: ${{ steps.meta.outputs.namespace }}login_email: ${{ steps.meta.outputs.login_email }}@@ -102,6 +102,7 @@ jobs:HEAD_SHA: ${{ github.event.pull_request.head.sha }}WEB_REPO: ${{ vars.AWS_PREVIEW_WEB_ECR_REPOSITORY_URL }}WORKER_REPO: ${{ vars.AWS_PREVIEW_WORKER_ECR_REPOSITORY_URL }}+ AI_GATEWAY_REPO: ${{ vars.AWS_PREVIEW_AI_GATEWAY_ECR_REPOSITORY_URL }}DOMAIN_NAME: ${{ vars.AWS_PREVIEW_DOMAIN_NAME }}run: |set -euo pipefail@@ -114,6 +115,7 @@ jobs:echo "commit_sha=${commit_sha}"echo "web_image=${WEB_REPO}:${image_tag}"echo "worker_image=${WORKER_REPO}:${image_tag}"+ echo "ai_gateway_image=${AI_GATEWAY_REPO}:${image_tag}"echo "preview_host=pr-${PR_NUMBER}.${DOMAIN_NAME}"# k8s namespace/release for this PR (web+worker deploys are# <namespace>-web / <namespace>-worker) β used for the log commands@@ -141,12 +143,13 @@ jobs:### π‘ AWS preview buildingBuilding images for `${{ steps.meta.outputs.commit_sha }}` (~5 min).- Argo CD deploys the preview once both images finish pushing; it may- briefly show `ImagePullBackOff` until they propagate.+ Argo CD deploys web and worker once those images finish pushing; it+ may briefly show `ImagePullBackOff` until they propagate.# web and worker build concurrently, each on its own runner. fail-fast: false- # so one failing build does not cancel the other β the notify job below reports- # the aggregate result.+ # so one failing build does not cancel the other β the notify job below+ # reports the aggregate result. The gateway is a separate job and does not+ # gate this result.build:name: Build and push ${{ matrix.target }}needs: meta@@ -166,6 +169,7 @@ jobs:- target: webimage: ${{ needs.meta.outputs.web_image }}dockerfile: web/Dockerfile+ context: .# Signup is on (Mailpit catches mail and does not relay). The cloud# region must be baked at build time (NEXT_PUBLIC_* is inlined into# the client bundle) so previews render the cloud-only AI features;@@ -177,6 +181,7 @@ jobs:- target: workerimage: ${{ needs.meta.outputs.worker_image }}dockerfile: worker/Dockerfile+ context: .extra_build_args: ""steps:# The PR head is the Docker build context β the exact commit the image tag@@ -202,6 +207,7 @@ jobs:COMMIT_SHA: ${{ needs.meta.outputs.commit_sha }}IMAGE: ${{ matrix.image }}DOCKERFILE: ${{ matrix.dockerfile }}+ CONTEXT: ${{ matrix.context }}# Static, workflow-defined literal (per-target flags) β no user input.EXTRA_BUILD_ARGS: ${{ matrix.extra_build_args }}TARGET: ${{ matrix.target }}@@ -264,10 +270,142 @@ jobs:"${preview_meta_args[@]}" \-f "${DOCKERFILE}" \-t "${IMAGE}" \- .+ "${CONTEXT:-.}"docker push "${IMAGE}"- # Single authoritative status comment once both builds settle, updating the+ # Argo deploys this image. This job does not wait for it: a gateway failure+ # must not replace the preview URL, deployment record, or description pin,+ # and a PR head from before the crate existed must still publish web and+ # worker. continue-on-error keeps that failure off the workflow conclusion;+ # the job still reports it in its own comment. -f is a path from the repo+ # root: Docker resolves a relative Dockerfile against the working directory,+ # not the context.+ build-gateway:+ name: Build and push ai-gateway+ needs: meta+ if: >-+ vars.AWS_PREVIEW_ECR_PUSH_ROLE_ARN != '' &&+ github.event.pull_request.head.repo.full_name == github.repository &&+ github.event.pull_request.state == 'open' &&+ (github.event.action != 'labeled' || github.event.label.name == 'preview')+ runs-on: blacksmith-4vcpu-ubuntu-2404+ continue-on-error: true+ permissions:+ contents: read+ id-token: write+ pull-requests: write+ steps:+ - name: Checkout PR head (build context)+ uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1+ with:+ persist-credentials: false+ ref: ${{ github.event.pull_request.head.sha }}++ - name: Skip when the repository or crate is absent+ id: gate+ env:+ IMAGE: ${{ needs.meta.outputs.ai_gateway_image }}+ run: |+ set -euo pipefail+ # An unset repository variable renders the image as ":<tag>".+ if [ -z "${IMAGE%%:*}" ]; then+ echo "AWS_PREVIEW_AI_GATEWAY_ECR_REPOSITORY_URL is unset β skipping gateway image."+ echo "skip=true" >> "$GITHUB_OUTPUT"+ exit 0+ fi+ if [ ! -f ai-gateway/Dockerfile ]; then+ echo "PR head has no ai-gateway/Dockerfile β skipping gateway image."+ echo "skip=true" >> "$GITHUB_OUTPUT"+ exit 0+ fi+ echo "skip=false" >> "$GITHUB_OUTPUT"++ - name: Configure AWS credentials+ if: steps.gate.outputs.skip != 'true'+ uses: aws-actions/configure-aws-credentials@e6de054238d6b7531b4efff3b6587d9aade6a06c # v6.2.3+ with:+ aws-region: ${{ vars.AWS_PREVIEW_REGION }}+ role-to-assume: ${{ vars.AWS_PREVIEW_ECR_PUSH_ROLE_ARN }}++ - name: Login to AWS ECR+ if: steps.gate.outputs.skip != 'true'+ uses: aws-actions/amazon-ecr-login@d539f0932e70871a027e9d5a9d8fc38589180a64 # v2.1.6++ - name: Build and push ai-gateway image+ if: steps.gate.outputs.skip != 'true'+ env:+ COMMIT_SHA: ${{ needs.meta.outputs.commit_sha }}+ IMAGE: ${{ needs.meta.outputs.ai_gateway_image }}+ run: |+ set -euo pipefail++ # Same immutable-tag skip as web and worker. Only ImageNotFoundException+ # means "absent -> build"; any other describe failure is inconclusive.+ repo="${IMAGE%:*}"; repo="${repo#*/}"+ tag="${IMAGE##*:}"++ if describe_out="$(aws ecr describe-images --repository-name "${repo}" \+ --image-ids imageTag="${tag}" 2>&1)"; then+ echo "${IMAGE} already present β skipping."+ exit 0+ fi+ if ! grep -q 'ImageNotFoundException' <<<"${describe_out}"; then+ echo "::error::Cannot confirm ${IMAGE} in ECR β describe-images failed" \+ "for a reason other than ImageNotFound; refusing to build+push" \+ "against the tag-immutable repo. ${describe_out}"+ exit 1+ fi++ echo "${IMAGE} not found β building and pushing."+ docker build \+ --target runtime \+ --build-arg "BUILD_ID=${COMMIT_SHA}" \+ -f ai-gateway/Dockerfile \+ -t "${IMAGE}" \+ ai-gateway+ docker push "${IMAGE}"++ # Own marker, not <!-- aws-preview -->, so this cannot replace the+ # preview URL comment the notify job owns.+ - name: Report gateway build failure+ if: failure()+ uses: actions/github-script@3a2844b7e9c422d3c10d287c895573f7108da1b3 # v9.0.0+ with:+ script: |+ const marker = "<!-- aws-preview-gateway -->";+ const { owner, repo } = context.repo;+ const issue_number = context.payload.pull_request.number;+ const runUrl = `${process.env.GITHUB_SERVER_URL}/${owner}/${repo}/actions/runs/${process.env.GITHUB_RUN_ID}`;+ const body = [+ marker,+ "### β οΈ AI gateway image failed to build",+ "",+ "The preview URL is unaffected β web and worker still deploy.",+ `Gateway build: ${runUrl}`,+ ].join("\n");+ const comments = await github.paginate(github.rest.issues.listComments, {+ owner,+ repo,+ issue_number,+ per_page: 100,+ });+ const previous = comments.find(+ (comment) =>+ comment.user?.login === "github-actions[bot]" &&+ comment.body?.includes(marker),+ );+ if (previous) {+ await github.rest.issues.updateComment({+ owner,+ repo,+ comment_id: previous.id,+ body,+ });+ } else {+ await github.rest.issues.createComment({ owner, repo, issue_number, body });+ }++ # Single authoritative status comment once web and worker settle, updating the# same marker-tagged comment the meta job posted. Runs on always() so it still# reports after a failure, but each step keys on a definite success/failure# result (see the per-step guards) so a cancelled run β e.g. superseded by a