跳到主要内容
Documentation

Upgrade and Rollback

This runbook covers how to upgrade, pin, and roll back each release surface of

Docs indexEdit page
版本：v0.3

Upgrade and Rollback

This runbook covers how to upgrade, pin, and roll back each release surface of the vLLM Semantic Router in a production environment.

Release Channels​

ChannelTag patternUpdated onUse case
Versionedv0.3.0 / 0.3.0Tagged releases onlyProduction — immutable, recommended
Nightlynightly-20260115Daily at 02:00 UTCPre-release testing
LatestlatestEvery push to main + releasesDevelopment only
Recommendation

Always use a versioned tag in production. It is immutable — the digest never changes for a given version tag. Find the latest release on the GitHub Releases page.

Prerequisites​

  • helm ≥ 3.14 (for Helm OCI operations)
  • kubectl configured for your target cluster
  • pip ≥ 22 (for Python CLI)
  • docker or podman (for direct image operations)

1. Checking Your Current Version​

Helm release​

helm list -n vllm-semantic-router-system
helm history semantic-router -n vllm-semantic-router-system

The CHART column shows the chart version (e.g. semantic-router-0.2.0) and APP VERSION shows the image tag that chart deployed.

Running container image​

# Get the image tag currently used by the extproc deployment
kubectl get deployment -n vllm-semantic-router-system \
  -o jsonpath='{range .items[*]}{.metadata.name}{"\t"}{.spec.template.spec.containers[0].image}{"\n"}{end}'

Python CLI​

vllm-sr --version
pip show vllm-sr

2. Upgrading​

2a. Helm chart upgrade​

Always upgrade to a specific version. Never rely on latest in production.

# Pull the chart metadata first (optional but useful to verify it exists)
helm show chart oci://ghcr.io/vllm-project/charts/semantic-router --version 0.3.0

# Upgrade to a specific version
# --reset-then-reuse-values (Helm ≥ 3.14) resets to the new chart's defaults
# first, then re-applies your previous overrides on top. This is safer than
# --reuse-values alone, which breaks if the new chart adds new required values.
helm upgrade semantic-router \
  oci://ghcr.io/vllm-project/charts/semantic-router \
  --version 0.3.0 \
  --namespace vllm-semantic-router-system \
  --reset-then-reuse-values \
  --wait \
  --timeout 10m
Use --reset-then-reuse-values, not --reuse-values, for cross-version upgrades

--reuse-values only merges your old stored values and skips new chart defaults, which causes template errors when a new chart version introduces new required values. --reset-then-reuse-values (Helm ≥ 3.14) resets to the new defaults first, then re-applies your overrides — it is always safe to use. If you are on Helm < 3.14, supply your configuration explicitly with -f your-values.yaml instead.

Verify after upgrade:

helm status semantic-router -n vllm-semantic-router-system
kubectl rollout status deployment/semantic-router -n vllm-semantic-router-system

2b. Docker image upgrade (non-Helm deployments)​

Find the latest version on the GitHub Releases page, then:

# Pull by version tag (substitute podman for docker if using podman)
docker pull ghcr.io/vllm-project/semantic-router/extproc:v0.3.0
docker pull ghcr.io/vllm-project/semantic-router/vllm-sr:v0.3.0
docker pull ghcr.io/vllm-project/semantic-router/anthropic-shim:v0.3.0

# Get the immutable digest for maximum pinning stability
DIGEST=$(docker inspect --format='{{index .RepoDigests 0}}' \
  ghcr.io/vllm-project/semantic-router/extproc:v0.3.0)
echo "Use digest: ${DIGEST}"

For Kubernetes manifests, pin to the digest, not the tag:

image: ghcr.io/vllm-project/semantic-router/extproc@sha256:<digest>

Published versioned images for a full release:

ImageTypical owner
ghcr.io/vllm-project/semantic-router/extproc:v0.3.0Router ExtProc runtime
ghcr.io/vllm-project/semantic-router/extproc-rocm:v0.3.0ROCm router ExtProc runtime
ghcr.io/vllm-project/semantic-router/vllm-sr:v0.3.0Local/runtime CLI image
ghcr.io/vllm-project/semantic-router/vllm-sr-rocm:v0.3.0ROCm local/runtime CLI image
ghcr.io/vllm-project/semantic-router/anthropic-shim:v0.3.0Anthropic-compatible API shim image
ghcr.io/vllm-project/semantic-router/dashboard:v0.3.0Dashboard backend/frontend image
ghcr.io/vllm-project/semantic-router/llm-katan:v0.3.0Fleet simulation service image
ghcr.io/vllm-project/semantic-router/operator:v0.3.0Kubernetes operator image
ghcr.io/vllm-project/semantic-router/operator-bundle:v0.3.0Operator bundle image

2c. Python CLI upgrade​

pip install --upgrade vllm-sr==0.3.0
vllm-sr --version    # verify

To upgrade to the latest stable release:

pip install --upgrade vllm-sr

2d. Fleet simulator Python package upgrade​

vllm-sr-sim is a separate PyPI package with its own release cadence. Pin it explicitly when you depend on the simulator CLI or dashboard sidecar package data:

pip install --upgrade vllm-sr-sim==0.1.0

Fleet simulator package releases use the independent vllm-sr-sim-v<version> tag stream and the pypi-publish-vllm-sr-sim.yml workflow; they are not published by the main v<version> router release tag.

3. Rollback​

3a. Helm rollback (fastest path)​

Helm keeps a history of every deployed revision. Rolling back requires no re-download and takes effect immediately.

# View history
helm history semantic-router -n vllm-semantic-router-system

# Roll back to the previous revision
helm rollback semantic-router -n vllm-semantic-router-system --wait

# Roll back to a specific revision number (e.g. revision 3)
helm rollback semantic-router 3 -n vllm-semantic-router-system --wait

# Verify
helm status semantic-router -n vllm-semantic-router-system
kubectl rollout status deployment/semantic-router -n vllm-semantic-router-system

Alternatively, roll back by re-installing an older chart version:

helm upgrade semantic-router \
  oci://ghcr.io/vllm-project/charts/semantic-router \
  --version 0.2.0 \
  --namespace vllm-semantic-router-system \
  --reset-then-reuse-values \
  --wait

3b. Docker / Kubernetes manifest rollback​

If you are managing Kubernetes manifests directly (without Helm), roll back the Deployment to the previous revision using the built-in rollout history:

# View rollout history
kubectl rollout history deployment/semantic-router -n vllm-semantic-router-system

# Undo the last rollout
kubectl rollout undo deployment/semantic-router -n vllm-semantic-router-system

# Undo to a specific revision
kubectl rollout undo deployment/semantic-router \
  --to-revision=3 -n vllm-semantic-router-system

# Verify
kubectl rollout status deployment/semantic-router -n vllm-semantic-router-system

If using pinned image digests, update your manifest to the previous image digest and kubectl apply.

3c. Python CLI rollback​

pip install vllm-sr==0.2.0
vllm-sr --version

4. Version Pinning Reference​

Makefile variables​

When building or deploying locally via make, override these variables to target a specific release instead of latest:

# Use a specific image tag for all docker-* targets
make docker-build-extproc DOCKER_TAG=v0.3.0

# Pull all production images at a specific version
make docker-pull-release DOCKER_TAG=v0.3.0

# Install/upgrade the Helm chart at a pinned chart version
make helm-upgrade-version CHART_VERSION=0.3.0

Create a values-production.yaml that explicitly pins image tags:

image:
  tag: "v0.3.0"   # pin to an immutable release tag
  pullPolicy: IfNotPresent

Then deploy with:

helm upgrade semantic-router \
  oci://ghcr.io/vllm-project/charts/semantic-router \
  --version 0.3.0 \
  -f values-production.yaml \
  --namespace vllm-semantic-router-system

5. Nightly Builds​

Nightly images and charts are built every day at 02:00 UTC and tagged nightly-YYYYMMDD. They are intended for pre-release testing only.

# Pull the nightly image built on a specific date
docker pull ghcr.io/vllm-project/semantic-router/vllm-sr:nightly-20260115

# Install the nightly Helm chart
helm install semantic-router \
  oci://ghcr.io/vllm-project/charts/semantic-router \
  --version 0.0.0-nightly.20260115 \
  --namespace vllm-semantic-router-system --create-namespace

Nightly builds are not automatically promoted to a versioned release. Promotion happens only via a tagged release.

6. Promotion Policy​

nightly-YYYYMMDD  ──→  (manual QA + CI green)  ──→  v0.3.0

A nightly build is promoted to a release by:

  1. Verifying all CI checks pass on the candidate commit.
  2. Bumping version fields in src/vllm-sr/pyproject.toml and candle-binding/Cargo.toml to the target version.
  3. Pushing a v<version> tag — this triggers docker-release.yml, helm-publish.yml, pypi-publish.yml, publish-crate.yml, and release.yml simultaneously.
  4. The release.yml workflow validates all surfaces are consistent before the GitHub Release is created.

For dashboard images, the Docker release workflow also passes the pushed v<version> tag into the dashboard backend build. The dashboard /api/status response therefore reports the same tag-form version as the released image. Non-release Docker publish builds use the src/vllm-sr/pyproject.toml version plus PR, commit, or nightly metadata, for example v0.3.0-dev.<sha> or v0.3.0-nightly.<date>.<sha>.

Fleet simulator package releases are promoted separately by bumping src/fleet-sim/pyproject.toml and pushing a vllm-sr-sim-v<version> tag, which triggers pypi-publish-vllm-sr-sim.yml.

There is no automated gating from nightly → release; that decision is made by the release owner.

7. Troubleshooting​

Helm: Error: chart not found​

# List available versions in the OCI registry (requires oras CLI)
oras repo tags ghcr.io/vllm-project/charts/semantic-router

# Verify a specific version exists before installing
helm show chart oci://ghcr.io/vllm-project/charts/semantic-router --version 0.3.0

Helm: release is in a broken state after failed upgrade​

helm rollback semantic-router -n vllm-semantic-router-system --wait
# If rollback also fails due to a bad state, force-reinstall:
helm uninstall semantic-router -n vllm-semantic-router-system
helm install semantic-router \
  oci://ghcr.io/vllm-project/charts/semantic-router \
  --version <last-known-good> \
  -f your-values.yaml \
  --namespace vllm-semantic-router-system --create-namespace

Kubernetes: ImagePullBackOff after upgrade​

The image tag may not exist yet (release still publishing) or the pull secret is missing. Check:

kubectl describe pod -n vllm-semantic-router-system <pod-name>
# Look for "ErrImagePull" and the exact tag that failed

If the tag genuinely does not exist, roll back while the release completes:

helm rollback semantic-router -n vllm-semantic-router-system