Files

T

agent-radar 04bac1f6e2 源码快照 @ 2026-09-02 (source-only, 剔除非源码素材)

2026-09-02 01:05:41 +08:00

19 KiB

Raw Permalink Blame History

Deploy OpenWork EE on AWS with EKS and Helm

Status: self-host operator guide Related: packaging/helm/openwork-ee, packaging/helm/openwork-ee/examples/values.aws-load-balancer.yaml, packaging/helm/openwork-ee/examples/values.aws-load-balancer-http-smoke.yaml

This is the recommended AWS path for a first production-like OpenWork EE self-host install. Use Helm on Amazon EKS with Amazon RDS for MySQL. For the simplest customer deployment, use EKS Auto Mode and Kubernetes LoadBalancer Services so AWS provisions Network Load Balancers directly from the chart. Use an Ingress/ALB path later when you need one shared ALB, advanced HTTP routing, WAF rules, or an existing ingress platform.

What this deploys

Den API on port 8788
Den Web on port 3005
optional inference service, disabled by default
one RDS MySQL database
one single-org OpenWork deployment
two public AWS Network Load Balancers by default: one for web and one for API

AWS owns the EKS cluster, node lifecycle, VPC networking, load balancers, RDS, DNS, TLS certificates, IAM, and security groups. The OpenWork Helm chart owns OpenWork Deployments, Services, ConfigMaps, Secrets, health probes, and the database migration Job.

Use Helm or something else?

Use Helm on EKS for AWS unless the customer explicitly cannot run Kubernetes. The OpenWork EE release artifact is already a Helm chart, the service split maps cleanly to Kubernetes, and EKS Auto Mode removes most node and load balancer setup from the customer path. A VM or ECS guide can be useful later, but it would be a separate packaging surface to maintain. The practical gap found by AWS testing was not Helm itself; it was missing AWS-specific infra guidance and missing chart knobs for AWS service annotations.

Prerequisites

AWS CLI authenticated to the target account.
kubectl, helm, and eksctl.
eksctl version 0.195.0 or newer for EKS Auto Mode.
Permission to create EKS, EC2/VPC, IAM, Elastic Load Balancing, RDS, Secrets Manager, Route 53, and ACM resources.
A real admin email address for the first owner account.
A domain you control, such as openwork.example.com and api.openwork.example.com.

Do not run production installs as the AWS root account. Use AWS SSO or an IAM role with only the permissions needed for the cluster, VPC/load balancer, RDS, DNS, and certificate resources.

AWS docs used for this guide:

EKS Auto Mode: https://docs.aws.amazon.com/eks/latest/userguide/automode.html
EKS Auto Mode with eksctl: https://docs.aws.amazon.com/eks/latest/userguide/automode-get-started-eksctl.html
EKS Auto Mode NLB service annotations: https://docs.aws.amazon.com/eks/latest/userguide/auto-configure-nlb.html
AWS Load Balancer Controller: https://docs.aws.amazon.com/eks/latest/userguide/aws-load-balancer-controller.html
RDS TLS: https://docs.aws.amazon.com/AmazonRDS/latest/UserGuide/UsingWithRDS.SSL.html

CloudShell quickstart

If you start from the AWS Console, CloudShell is a reasonable test environment, but it may not have every required tool or a default region configured. Set the region explicitly even if the console region selector already shows the right region:

export AWS_REGION=us-east-1
aws sts get-caller-identity
aws configure get region

Install pinned Helm and eksctl releases if they are missing. Verify each archive against the checksum published for that exact release before installing the binary:

HELM_VERSION=v3.18.4
HELM_ARCHIVE="helm-${HELM_VERSION}-linux-amd64.tar.gz"
curl -fsSL "https://get.helm.sh/${HELM_ARCHIVE}" -o "/tmp/${HELM_ARCHIVE}"
printf '%s  %s\n' \
  'f8180838c23d7c7d797b208861fecb591d9ce1690d8704ed1e4cb8e2add966c1' \
  "/tmp/${HELM_ARCHIVE}" | sha256sum --check --strict
tar -xzf "/tmp/${HELM_ARCHIVE}" -C /tmp
sudo install -m 0755 /tmp/linux-amd64/helm /usr/local/bin/helm

EKSCTL_VERSION=v0.195.0
EKSCTL_ARCHIVE=eksctl_Linux_amd64.tar.gz
curl -fsSL \
  "https://github.com/eksctl-io/eksctl/releases/download/${EKSCTL_VERSION}/${EKSCTL_ARCHIVE}" \
  -o "/tmp/${EKSCTL_ARCHIVE}"
printf '%s  %s\n' \
  '1cc86fd94da2378687f4db7c537da3c7316f2dca1347a4223a2ed86ad94dd818' \
  "/tmp/${EKSCTL_ARCHIVE}" | sha256sum --check --strict
tar -xzf "/tmp/${EKSCTL_ARCHIVE}" -C /tmp
sudo install -m 0755 /tmp/eksctl /usr/local/bin/eksctl

helm version
eksctl version
kubectl version --client

The checksums above are the official Linux AMD64 values for Helm v3.18.4 and eksctl v0.195.0. When updating either pin, replace its checksum from the corresponding official release rather than changing the version alone.

1. Create the EKS cluster

For a first deployment, create an EKS Auto Mode cluster:

export AWS_REGION=us-east-1
export CLUSTER_NAME=openwork-ee

eksctl create cluster \
  --name "$CLUSTER_NAME" \
  --region "$AWS_REGION" \
  --enable-auto-mode

aws eks update-kubeconfig \
  --region "$AWS_REGION" \
  --name "$CLUSTER_NAME"

kubectl get nodes

EKS Auto Mode handles default compute, pod networking, DNS, block storage, and load balancer integration. It also makes Network Load Balancers available for Kubernetes Services of type LoadBalancer.

Fresh EKS Auto Mode clusters can show no nodes at first. Auto Mode provisions compute when there is pending workload demand, so kubectl get nodes may return No resources found until the first pods are scheduled.

2. Create RDS MySQL

Create a MySQL database reachable from the EKS worker security group. The exact VPC and subnet commands vary by account, so the important requirements are:

RDS MySQL 8-compatible engine.
Database name: openwork_den.
Private subnets in the same VPC as the EKS cluster.
RDS security group inbound TCP 3306 from the EKS node/pod security group.
Storage encryption enabled.
Backups enabled.
Public accessibility disabled for production.
TLS required or at least supported.

Example database URL:

mysql://openwork:<password>@<rds-endpoint>:3306/openwork_den?sslaccept=accept

Use ?sslaccept=accept for the simple private-RDS smoke path. This keeps TLS on but does not require the RDS CA bundle to be mounted into the OpenWork image. Use strict certificate verification later, after you provide the RDS CA bundle, with a hardened value such as sslmode=verify-ca or sslmode=verify-full.

Wait until the EKS cluster is ACTIVE before wiring RDS security groups. Then allow MySQL traffic from both the EKS cluster security group and the eksctl shared node security group:

export CLUSTER_SG_ID=$(aws eks describe-cluster \
  --region "$AWS_REGION" \
  --name "$CLUSTER_NAME" \
  --query 'cluster.resourcesVpcConfig.clusterSecurityGroupId' \
  --output text)

export NODE_SG_ID=$(aws cloudformation describe-stacks \
  --region "$AWS_REGION" \
  --stack-name "eksctl-${CLUSTER_NAME}-cluster" \
  --query 'Stacks[0].Outputs[?OutputKey==`SharedNodeSecurityGroup`].OutputValue' \
  --output text)

aws ec2 authorize-security-group-ingress \
  --region "$AWS_REGION" \
  --group-id "$RDS_SG_ID" \
  --protocol tcp \
  --port 3306 \
  --source-group "$CLUSTER_SG_ID"

aws ec2 authorize-security-group-ingress \
  --region "$AWS_REGION" \
  --group-id "$RDS_SG_ID" \
  --protocol tcp \
  --port 3306 \
  --source-group "$NODE_SG_ID"

Before installing OpenWork, verify network access from the cluster. One simple way is to run a temporary MySQL client pod:

kubectl run mysql-client \
  --rm \
  -it \
  --restart=Never \
  --image=mysql:8 \
  -- mysql \
    --host="$RDS_ENDPOINT" \
    --user=openwork \
    --password \
    --ssl-mode=REQUIRED \
    --execute "select 1"

3. Prepare Helm values

Copy the starter file:

cp packaging/helm/openwork-ee/examples/values.aws-load-balancer.yaml values.aws.yaml

If you do not have DNS and ACM ready yet, use the HTTP smoke-test starter instead:

cp packaging/helm/openwork-ee/examples/values.aws-load-balancer-http-smoke.yaml values.aws.yaml

Replace every REPLACE_* placeholder.

Generate secrets:

openssl rand -base64 48
openssl rand -base64 48

Use the first value for secret.values.betterAuthSecret and the second for secret.values.denDbEncryptionKey. Do not reuse either value across environments.

To send transactional email, configure SMTP in the same values file:

secret:
  values:
    emailFrom: "OpenWork <no-reply@example.com>"
    smtpHost: "smtp.example.com"
    smtpPort: "587"
    smtpUser: "openwork@example.com"
    smtpPass: "REPLACE_SMTP_PASSWORD"
    smtpSecure: "false"

These values become EMAIL_FROM, SMTP_HOST, SMTP_PORT, SMTP_USER, SMTP_PASS, and SMTP_SECURE in Den API. If you use secret.create=false, add those keys to the existing Kubernetes Secret referenced by secret.existingSecret. SMTP delivery requires both EMAIL_FROM and SMTP_HOST; leave smtpHost blank only when SMTP-backed transactional email should be disabled.

Use a values file, not a long list of --set flags. Several OpenWork values are comma-separated strings, such as config.public.corsOrigins, and plain --set parsing commonly breaks them.

Make sure your file uses the current chart keys. Public URL values belong under config.public.*; database and app secrets belong under secret.values.*. Values such as config.urls, config.databaseUrl, or secrets.* are ignored by the chart.

Before installing, render the chart and verify the migration Job will use your RDS URL:

helm template openwork-ee oci://ghcr.io/different-ai/charts/openwork-ee \
  --version REPLACE_OPENWORK_VERSION \
  --namespace openwork-ee \
  -f values.aws.yaml > /tmp/openwork-rendered.yaml

grep -E 'DATABASE_URL|DEN_BASE_URL|DEN_WEB_PUBLIC_ORIGIN|EMAIL_FROM|SMTP_HOST|SMTP_PORT|SMTP_SECURE' /tmp/openwork-rendered.yaml

Redact secrets before sharing rendered manifests or terminal output.

4. Install OpenWork

Published chart releases live in GHCR:

helm upgrade --install openwork-ee oci://ghcr.io/different-ai/charts/openwork-ee \
  --version REPLACE_OPENWORK_VERSION \
  --namespace openwork-ee \
  --create-namespace \
  -f values.aws.yaml

For a checkout-local test:

helm upgrade --install openwork-ee ./packaging/helm/openwork-ee \
  --namespace openwork-ee \
  --create-namespace \
  -f values.aws.yaml

If GHCR image pulls fail with ImagePullBackOff, authenticate to GHCR and add an imagePullSecrets entry. Public releases should not require this, but private packages or private forks do:

kubectl create secret docker-registry ghcr-pull-secret \
  --namespace openwork-ee \
  --docker-server=ghcr.io \
  --docker-username="$GITHUB_USER" \
  --docker-password="$GITHUB_TOKEN"

imagePullSecrets:
  - name: ghcr-pull-secret

5. Migration troubleshooting

The migration Job runs before the Deployments and Services are installed. If it fails, fix that before debugging web/API readiness.

Avoid kubectl describe job openwork-ee-migrate in shared reports because the hook Job currently includes DATABASE_URL and DEN_DB_ENCRYPTION_KEY in the rendered environment. Use logs and redacted rendered manifests instead.

For a retained-log debug attempt, temporarily disable hook behavior:

migrations:
  enabled: true
  hook: false
  backoffLimit: 0

Then run Helm and inspect the normal Job logs:

helm upgrade --install openwork-ee oci://ghcr.io/different-ai/charts/openwork-ee \
  --version REPLACE_OPENWORK_VERSION \
  --namespace openwork-ee \
  --create-namespace \
  -f values.aws.yaml \
  --wait=false

kubectl get jobs,pods -n openwork-ee
kubectl logs -n openwork-ee -l job-name=openwork-ee-migrate --all-containers=true

If you see self-signed certificate in certificate chain against RDS, use the documented private-RDS smoke URL with ?sslaccept=accept, or mount/configure the RDS CA bundle before using strict certificate verification.

Return to the default hook mode after debugging:

migrations:
  enabled: true
  hook: true
  backoffLimit: 2

6. Point DNS at AWS load balancers

Wait for AWS to allocate load balancer hostnames:

kubectl get svc -n openwork-ee

You should see external hostnames for:

openwork-ee-den-web
openwork-ee-den-api

Create DNS records:

openwork.example.com -> Den Web load balancer hostname.
api.openwork.example.com -> Den API load balancer hostname.

The starter values terminate TLS on port 443 at each Network Load Balancer and forward clear HTTP to the Kubernetes service target port. Use an ACM certificate that covers both the web and API hosts:

denWeb:
  service:
    port: 443
    annotations:
      service.beta.kubernetes.io/aws-load-balancer-ssl-cert: arn:aws:acm:...
      service.beta.kubernetes.io/aws-load-balancer-ssl-ports: "443"
denApi:
  service:
    port: 443
    annotations:
      service.beta.kubernetes.io/aws-load-balancer-ssl-cert: arn:aws:acm:...
      service.beta.kubernetes.io/aws-load-balancer-ssl-ports: "443"

For production domains, prefer HTTPS before testing SSO. Browser auth cookies and identity-provider callback policies are much easier to validate on stable HTTPS origins than on raw load balancer hostnames.

For a temporary HTTP-only smoke test, remove the SSL annotations, set denWeb.service.port back to 3005, set denApi.service.port back to 8788, and use explicit http://host:port origins in values.aws.yaml.

If you are still using raw AWS load balancer hostnames before DNS/TLS is ready, temporarily update the corresponding config.public.* origins in values.aws.yaml, then run helm upgrade again. Do not leave production deployments on raw load balancer hostnames.

The current chart rolls the Den API, Den Web, and inference pods automatically when ConfigMap or Secret content changes. On older chart versions, manually restart the deployments after changing public origin values:

kubectl rollout restart deployment/openwork-ee-den-api deployment/openwork-ee-den-web -n openwork-ee
kubectl rollout status deployment/openwork-ee-den-api -n openwork-ee --timeout=180s
kubectl rollout status deployment/openwork-ee-den-web -n openwork-ee --timeout=180s

7. Verify readiness

Check Kubernetes state:

helm status openwork-ee -n openwork-ee
kubectl get pods -n openwork-ee
kubectl get jobs -n openwork-ee
kubectl describe pods -n openwork-ee
kubectl logs -n openwork-ee deploy/openwork-ee-den-api
kubectl logs -n openwork-ee deploy/openwork-ee-den-web

Check service readiness from your machine:

curl -fsS https://api.openwork.example.com/ready
curl -fsS https://openwork.example.com/api/ready

For HTTP smoke tests, use the raw NLB hostnames and ports:

curl -fsS http://REPLACE_API_NLB_HOST:8788/ready
curl -fsS http://REPLACE_WEB_NLB_HOST:3005/api/ready

8. Bootstrap the first owner

The chart defaults to single_org. Set these before first sign-in:

config:
  tenancy:
    mode: single_org
    singleOrgName: "Acme"
    singleOrgSlug: "acme"
    ownerEmails: "admin@acme.com"
    requireEmailVerification: "false"
  public:
    bootstrapAdminEmails: "admin@acme.com"
secret:
  values:
    initialAdminBootstrapCode: "REPLACE_BOOTSTRAP_CODE"

For releases that include initial-administrator bootstrap, inject the release-documented one-time setup secret through the Kubernetes Secret referenced by secret.existingSecret. Do not store the code in the values file or a ConfigMap. Then open https://openwork.example.com/setup, enter the configured owner email and one-time operator code, and create the first account. OpenWork creates the singleton organization, grants owner and configured platform-admin access, and signs the administrator in. Public signup remains disabled. After the first user exists, the setup code cannot bootstrap another account.

ownerEmails and bootstrapAdminEmails authorize roles; neither setting creates an account or password. There is no default administrator password. Chart versions without /setup do not support private initial-administrator bootstrap and must be upgraded before following this step.

9. Configure SSO with a test IdP

Self-hosted installs keep plan gating off unless the operator explicitly sets DEN_PLAN_GATING_ENABLED=true, so SSO management should be available in the EE self-host default.

Use OIDC first for a smoke test because most IdPs provide discovery metadata. Auth0, Okta trial, and Microsoft Entra test tenants all work as realistic demo IdPs.

Configure the IdP application with these callback URLs:

https://openwork.example.com/api/auth/sso/callback/openwork-sso-<org-id>

In OpenWork, sign in as the owner, open the organization SSO settings, and enter the IdP issuer/client details. After saving, the organization sign-in path is:

https://openwork.example.com/sso/<singleOrgSlug>

For SAML, OpenWork shows the generated ACS URL and metadata URL after the SAML connection is registered. Use those values in the IdP rather than guessing. OpenWork rejects unsigned or weak SAML responses, so configure the IdP to sign assertions.

After SSO is configured, root sign-in shows the SSO-only experience for the single organization. Password sign-in for that organization is rejected.

10. Troubleshooting

Symptom	Likely cause	Fix
`helm` is missing in CloudShell	CloudShell does not always include Helm	Install Helm in CloudShell or run Helm locally with AWS credentials
`aws sts get-caller-identity` returns `NoCredentials`	AWS CLI is not authenticated	Configure AWS SSO/profile or use authenticated CloudShell
Initial `kubectl get nodes` is empty	EKS Auto Mode has not needed to provision nodes yet	Continue after the cluster is active; nodes appear when workloads are pending
Pods show untolerated Auto Mode taints	Auto Mode is still selecting/provisioning matching capacity	Wait for the NodeClaim, then check pod events again
Pod sandbox fails with `aws-cni failed ... failed to assign an IP address`	Subnet/IP or EKS CNI capacity problem	Check subnet free IPs, node events, and retry after Auto Mode provisions replacement capacity
Migration Job fails to connect to MySQL	RDS security group, DB credentials, or TLS mode are wrong	Allow TCP `3306` from EKS, verify with `mysql-client`, and use `?sslaccept=accept` for the private-RDS smoke path
Migration Job logs show `self-signed certificate in certificate chain`	Strict certificate verification is being used without the RDS CA bundle	Use `?sslaccept=accept` for the smoke path or mount/configure the RDS CA bundle before strict verification
Runtime is ready but migration failed	Helm hook did not complete	Check `kubectl get jobs` and the migration Job logs before testing web
`ImagePullBackOff` from GHCR	Private image or missing pull token	Add `imagePullSecrets`
Browser auth loops or CORS errors	Public origins do not match DNS/TLS	Set `webOrigin`, `apiOrigin`, `corsOrigins`, `betterAuthTrustedOrigins`, and `authCallbackUrl` to the final HTTPS domains
SSO callback rejected	IdP callback URL does not match OpenWork	Use the callback/ACS URL shown by OpenWork for that org/provider
SSO settings show Enterprise gating	`DEN_PLAN_GATING_ENABLED=true` or org is not entitled	Leave plan gating off for self-host smoke tests, or grant enterprise entitlement

11. Cleanup

For a disposable test:

helm uninstall openwork-ee -n openwork-ee
eksctl delete cluster --name "$CLUSTER_NAME" --region "$AWS_REGION"

Delete RDS snapshots, Secrets Manager secrets, Route 53 records, ACM certificates, and any manually created security groups if they were only for the test.

19 KiB Raw Permalink Blame History