Kubernetes deployment guide

This guide covers deploying Arbitex to Kubernetes using the official Helm charts. It covers the Platform (API + frontend + DLP microservices) and the Outpost (hybrid data plane).

---

Architecture overview

A full Arbitex deployment consists of two independent Helm releases:

┌─────────────────────────────────────────────────────────────┐
│  Kubernetes cluster                                          │
│                                                              │
│  namespace: arbitex                                          │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐   │
│  │ api      │  │ frontend │  │ ner-gpu  │  │ deberta  │   │
│  │ :8000    │  │ :8080    │  │ :8200    │  │ :8201    │   │
│  └──────────┘  └──────────┘  └──────────┘  └──────────┘   │
│       │             │         GPU pool        GPU pool       │
│  NGINX Ingress (port 443)                                    │
│       ↓                                                      │
│  api.arbitex.ai                                              │
└─────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────┐
│  namespace: arbitex-outpost (customer cluster)              │
│  ┌──────────────────────────────────────────────────┐       │
│  │ arbitex-outpost (replicas: 2)                    │       │
│  │  proxy :8300  admin :8301 (pod-local only)       │       │
│  │  policy cache PVC + audit buffer PVC             │       │
│  └──────────────────────────────────────────────────┘       │
│       │ mTLS                                                 │
│       ↓                                                      │
│  api.arbitex.ai  (management plane sync)                     │
└─────────────────────────────────────────────────────────────┘

The Platform chart deploys 4 workloads: API, frontend, NER GPU microservice, and DeBERTa validator. The Outpost chart deploys a single proxy workload with HA (2 replicas by default).

---

Part 1: Platform deployment

Prerequisites

• Kubernetes 1.25+
• Helm 3.12+
• kubectl configured for your target cluster
• NGINX ingress controller installed (helm install ingress-nginx ingress-nginx/ingress-nginx)
• PostgreSQL 14+ database accessible from the cluster
• Redis 7+ accessible from the cluster
• GPU nodes with the NVIDIA device plugin for DLP microservices (see GPU node pools)
• Container registry access to abtxacr.azurecr.io (imagePullSecret or registry credential)
• TLS certificate for api.arbitex.ai stored as a Kubernetes secret

Helm chart location

arbitex-platform/deploy/helm/arbitex-platform/
├── Chart.yaml
├── values.yaml          # base values — all environments
├── values-dev.yaml      # development overrides
├── values-staging.yaml  # staging overrides
├── values-prod.yaml     # production overrides
└── templates/
    ├── deployment-api.yaml
    ├── deployment-frontend.yaml
    ├── deployment-ner-gpu.yaml
    ├── deployment-deberta.yaml
    ├── job-alembic-migrate.yaml
    ├── ingress.yaml
    └── ...

Install — development

Development disables GPU services and uses relaxed Pod Security Standards:

helm upgrade --install arbitex-dev ./deploy/helm/arbitex-platform \
  -f values.yaml \
  -f values-dev.yaml \
  --set api.env.DATABASE_URL="postgresql+asyncpg://arbitex:arbitex@postgres-dev:5432/arbitex_dev" \
  --set api.env.REDIS_URL="redis://redis-dev:6379/0" \
  --set api.env.SECRET_KEY="<change-me-for-dev>" \
  --namespace arbitex-dev \
  --create-namespace

Dev overrides:

• namespace.podSecurityStandard: baseline (allows debug sidecars)
• nerGpu.enabled: false — no GPU node required
• deberta.enabled: false
• ingress.hosts[0].host: api.arbitex.local
• ingress.tls: [] (no TLS in dev)

Install — staging

helm upgrade --install arbitex-staging ./deploy/helm/arbitex-platform \
  -f values.yaml \
  -f values-staging.yaml \
  --set api.env.DATABASE_URL="postgresql+asyncpg://..." \
  --set api.env.REDIS_URL="redis://..." \
  --set api.env.SECRET_KEY="..." \
  --namespace arbitex-staging \
  --create-namespace

Staging overrides:

• GPU services enabled (nerGpu.enabled: true, deberta.enabled: true)
• ingress.hosts[0].host: api-staging.arbitex.ai
• TLS from secret arbitex-staging-tls

Install — production

Production uses sealed secrets or External Secrets Operator for credentials — do not pass them via --set in production pipelines:

helm upgrade --install arbitex ./deploy/helm/arbitex-platform \
  -f values.yaml \
  -f values-prod.yaml \
  --namespace arbitex \
  --create-namespace

After a CI build, pin the image digests to the promoted artifacts:

helm upgrade arbitex ./deploy/helm/arbitex-platform \
  -f values.yaml \
  -f values-prod.yaml \
  --set api.image.digest="sha256:<api-digest>" \
  --set frontend.image.digest="sha256:<frontend-digest>" \
  --set nerGpu.image.digest="sha256:<ner-digest>" \
  --set deberta.image.digest="sha256:<deberta-digest>" \
  --namespace arbitex

Configuration reference

The base values.yaml documents all supported keys. Key sections:

API service (api.*)

Key	Default	Description
api.replicas	1	Pod replica count
api.image.repository	abtxacr.azurecr.io/platform-api	Container image
api.image.tag	latest	Image tag
api.image.digest	""	SHA256 digest — overrides tag when set
api.resources.requests.cpu	250m	CPU request
api.resources.requests.memory	512Mi	Memory request
api.resources.limits.cpu	1000m	CPU limit
api.resources.limits.memory	1Gi	Memory limit
api.env.DATABASE_URL	""	Async PostgreSQL connection string (required)
api.env.REDIS_URL	""	Redis connection string (required)
api.env.SECRET_KEY	""	JWT signing secret (required)
api.env.ENVIRONMENT	production	Environment label
api.env.NER_GPU_URL	auto	NER GPU service URL (auto-resolved from chart name)
api.env.DEBERTA_URL	auto	DeBERTa service URL (auto-resolved from chart name)

NER GPU microservice (nerGpu.*)

Key	Default	Description
nerGpu.enabled	true	Enable NER GPU DLP microservice
nerGpu.replicas	1	Replica count
nerGpu.resources.requests.nvidia.com/gpu	1	GPU request
nerGpu.nodeSelector.accelerator	nvidia	Node selector for GPU pool
nerGpu.tolerations	nvidia.com/gpu:NoSchedule	Toleration for GPU taint

DeBERTa validator (deberta.*)

Same structure as nerGpu.*. Runs the DeBERTa v3 Tier 3 DLP model on GPU.

Ingress (ingress.*)

Key	Default	Description
ingress.className	nginx	Ingress class name
ingress.hosts	api.arbitex.ai	Hostname and path routing
ingress.tls[0].secretName	arbitex-tls	Kubernetes TLS secret name
ingress.annotations	ssl-redirect, proxy-body-size 50m	NGINX annotations

Alembic migration (alembic.*)

Key	Default	Description
alembic.runAsInitContainer	true	Run DB migration as init container before API starts
alembic.ttlSecondsAfterFinished	300	Job TTL after completion

GPU node pools

The NER GPU and DeBERTa services require nodes with NVIDIA GPUs. Configure your cluster:

AKS (Azure Kubernetes Service)

# Create a GPU node pool
az aks nodepool add \
  --resource-group myRG \
  --cluster-name myAKS \
  --name gpupool \
  --vm-size Standard_NC6s_v3 \
  --node-count 1 \
  --node-taints nvidia.com/gpu=present:NoSchedule \
  --labels accelerator=nvidia

The chart’s nodeSelector: { accelerator: nvidia } and tolerations target this pool automatically.

Self-managed clusters

1. Label GPU nodes:

   kubectl label node <gpu-node> accelerator=nvidia

2. Taint GPU nodes (optional but recommended to prevent non-GPU workloads on expensive nodes):

   kubectl taint node <gpu-node> nvidia.com/gpu=present:NoSchedule

3. Install the NVIDIA device plugin:

   kubectl apply -f https://raw.githubusercontent.com/NVIDIA/k8s-device-plugin/v0.14.5/nvidia-device-plugin.yml

Disabling GPU services

If you don’t have GPU nodes, disable both services:

values-nogpu.yaml

nerGpu:
  enabled: false
deberta:
  enabled: false

The API falls back to CPU-based DLP processing. DeBERTa Tier 3 is unavailable; DLP runs regex + NER (CPU) only.

Required secrets

Never put credentials directly in values files committed to source control. Use one of:

Kubernetes Secrets (minimum viable):

kubectl create secret generic arbitex-secrets \
  --from-literal=DATABASE_URL="postgresql+asyncpg://..." \
  --from-literal=REDIS_URL="redis://..." \
  --from-literal=SECRET_KEY="$(openssl rand -hex 32)" \
  --namespace arbitex

Reference in values:

api:
  envFrom:
    - secretRef:
        name: arbitex-secrets

Azure Key Vault (production — CSI driver):

api:
  env:
    SECRETS_BACKEND: vault
    VAULT_URL: "https://my-keyvault.vault.azure.net/"

See Security hardening guide for the full Key Vault CSI driver setup.

Verify the deployment

kubectl get pods -n arbitex
# Expected: api, frontend, ner-gpu, deberta all Running

kubectl get ingress -n arbitex
# Expected: ADDRESS shows the external IP

curl https://api.arbitex.ai/healthz
# Expected: 200 OK

Check migration ran:

kubectl logs -n arbitex -l job-name=arbitex-alembic-migrate --tail=20
# Expected: "INFO  [alembic.runtime.migration] Running upgrade ..."

---

Part 2: Outpost deployment

Prerequisites

• Kubernetes 1.25+ (separate cluster from Platform, or separate namespace)
• Helm 3.12+
• Outpost registered in the Arbitex admin console — you need the outpost UUID
• mTLS certificates issued by Arbitex (outpost.pem, outpost.key, ca.pem)
• GPU node (optional — for DeBERTa Tier 3 DLP)

Helm chart location

arbitex-outpost/charts/arbitex-outpost/
├── Chart.yaml
├── values.yaml          # all configuration keys with comments
├── values.schema.json   # JSON schema validation
└── templates/
    ├── deployment.yaml
    ├── service.yaml
    ├── configmap.yaml
    ├── secret.yaml
    ├── pvc.yaml
    ├── hpa.yaml
    └── poddisruptionbudget.yaml

Step 1: Create the mTLS certificate secret

The outpost requires three certificate files issued by Arbitex:

kubectl create secret generic outpost-certs \
  --from-file=outpost.pem=./certs/outpost.pem \
  --from-file=outpost.key=./certs/outpost.key \
  --from-file=ca.pem=./certs/ca.pem \
  --namespace arbitex-outpost

Step 2: Create the secrets secret

kubectl create secret generic outpost-secrets \
  --from-literal=OUTPOST_AUDIT_HMAC_KEY="$(openssl rand -base64 32)" \
  --from-literal=OUTPOST_ADMIN_API_KEY="$(openssl rand -hex 32)" \
  --namespace arbitex-outpost

Step 3: Install the chart

Minimal install (no GPU):

helm upgrade --install arbitex-outpost ./charts/arbitex-outpost \
  --set outpost.id="your-outpost-uuid" \
  --set outpost.platformManagementUrl="https://api.arbitex.ai" \
  --set outpost.orgId="your-org-uuid" \
  --set certs.existingSecret="outpost-certs" \
  --set outpost.auditHmacKey="$(kubectl get secret outpost-secrets -o jsonpath='{.data.OUTPOST_AUDIT_HMAC_KEY}' | base64 -d)" \
  --set outpost.adminApiKey="$(kubectl get secret outpost-secrets -o jsonpath='{.data.OUTPOST_ADMIN_API_KEY}' | base64 -d)" \
  --namespace arbitex-outpost \
  --create-namespace

GPU-enabled install (DeBERTa Tier 3):

helm upgrade --install arbitex-outpost ./charts/arbitex-outpost \
  --set outpost.id="your-outpost-uuid" \
  --set outpost.platformManagementUrl="https://api.arbitex.ai" \
  --set outpost.orgId="your-org-uuid" \
  --set certs.existingSecret="outpost-certs" \
  --set outpost.auditHmacKey="..." \
  --set outpost.adminApiKey="..." \
  --set outpost.gpuEnabled=true \
  --set outpost.dlpDebertaEnabled=true \
  --set outpost.debertaModelPath="/app/models/deberta/model.onnx" \
  --namespace arbitex-outpost \
  --create-namespace

Outpost values reference

Identity and platform connection

Key	Required	Default	Description
outpost.id	Yes	""	Outpost UUID issued when the outpost is registered in the Arbitex admin console
outpost.platformManagementUrl	Yes	""	Platform management plane URL
outpost.orgId	Yes	""	Organization UUID
outpost.auditHmacKey	Yes	""	Base64-encoded HMAC key (min 32 bytes)
outpost.adminApiKey	Yes	""	Emergency admin API key
certs.existingSecret	Yes	""	K8s secret with outpost.pem, outpost.key, ca.pem

Storage

Key	Default	Description
outpost.policyCacheSize	1Gi	PVC size for policy cache snapshots
outpost.auditBufferSize	5Gi	PVC size for offline audit buffer (~1KB/event, 5Gi ≈ 5M events)
outpost.maxAuditBufferEntries	100000	Ring buffer limit before oldest events are pruned

Sync intervals

Key	Default	Description
outpost.auditSyncIntervalSeconds	30	How often audit events push to Platform
outpost.policySyncIntervalSeconds	60	Policy cache refresh interval

DLP configuration

Key	Default	Description
outpost.dlpEnabled	true	Enable DLP pipeline
outpost.dlpNerDevice	auto	NER inference device: auto, cpu, or cuda
outpost.dlpDebertaEnabled	false	Enable DeBERTa Tier 3 (requires GPU + model path)
outpost.debertaModelPath	""	Path to ONNX model inside the container
outpost.gpuEnabled	false	Switch resource profile to GPU (nvidia.com/gpu: 1)

High availability

Key	Default	Description
replicaCount	2	Pod replica count (min 2 for PDB)
autoscaling.enabled	false	Enable HPA
autoscaling.minReplicas	2	HPA min replicas
autoscaling.maxReplicas	10	HPA max replicas
autoscaling.targetCPUUtilizationPercentage	70	HPA CPU target
podDisruptionBudget.minAvailable	1	PDB minimum available during rollouts

Example values files

outpost-values-dev.yaml

replicaCount: 1

outpost:
  id: "dev-outpost-uuid"
  platformManagementUrl: "https://api-staging.arbitex.ai"
  orgId: "dev-org-uuid"
  auditHmacKey: ""           # set via --set-string
  adminApiKey: ""            # set via --set-string
  logLevel: "debug"
  policyCacheSize: "256Mi"
  auditBufferSize: "512Mi"
  dlpEnabled: true
  dlpNerDevice: "cpu"
  dlpDebertaEnabled: false
  gpuEnabled: false

certs:
  existingSecret: "outpost-certs-dev"

autoscaling:
  enabled: false

podDisruptionBudget:
  minAvailable: 0            # single-node dev — PDB would block rollouts

outpost-values-prod.yaml

replicaCount: 2

outpost:
  id: ""                     # set via CI --set
  platformManagementUrl: "https://api.arbitex.ai"
  orgId: ""                  # set via CI --set
  auditHmacKey: ""           # set via external-secrets
  adminApiKey: ""            # set via external-secrets
  logLevel: "info"
  policyCacheSize: "1Gi"
  auditBufferSize: "5Gi"
  maxAuditBufferEntries: 100000
  auditSyncIntervalSeconds: 30
  policySyncIntervalSeconds: 60
  dlpEnabled: true
  dlpNerDevice: "auto"
  dlpDebertaEnabled: true
  debertaModelPath: "/app/models/deberta/model.onnx"
  gpuEnabled: true

certs:
  existingSecret: "outpost-certs"

resources:
  requests:
    memory: "512Mi"
    cpu: "250m"
  limits:
    memory: "2Gi"
    cpu: "2"

gpuResources:
  requests:
    memory: "2Gi"
    cpu: "500m"
    nvidia.com/gpu: "1"
  limits:
    memory: "8Gi"
    cpu: "4"
    nvidia.com/gpu: "1"

autoscaling:
  enabled: true
  minReplicas: 2
  maxReplicas: 10
  targetCPUUtilizationPercentage: 70

podDisruptionBudget:
  minAvailable: 1

Accessing the outpost admin API

The admin interface runs on port 8301 and is intentionally not exposed via a Kubernetes Service. Access it via port-forward:

kubectl port-forward \
  -n arbitex-outpost \
  deployment/arbitex-outpost \
  8301:8301

Then access the admin API locally:

curl http://localhost:8301/admin/api/health

---

Production readiness checklist

Before going live:

Platform:

• Image digests pinned in values-prod.yaml (no latest tags in production)
• DATABASE_URL, REDIS_URL, SECRET_KEY stored in Key Vault or Sealed Secrets — not in ConfigMaps
• TLS certificate in arbitex-tls secret, ingress TLS configured
• alembic.runAsInitContainer: true (database migration runs before API starts)
• podSecurityStandard: restricted in production namespace
• GPU nodes available and labeled/tainted for NER GPU + DeBERTa pods
• OTEL_EXPORTER_OTLP_ENDPOINT set for distributed tracing (see Distributed tracing guide)

Outpost:

• mTLS certificate secret created before chart install
• outpost.id and outpost.orgId set to correct production values
• auditHmacKey and adminApiKey stored in external secret manager
• replicaCount: 2 with PDB minAvailable: 1
• Persistent volumes provisioned for policy cache and audit buffer
• Outpost reachable from client workloads on port 8300

---

See also

• Outpost deployment guide — Docker Compose deployment + air-gap setup
• Distributed tracing guide — OTLP configuration for Platform + Outpost
• Security hardening guide — Key Vault, mTLS, CSRF Redis config
• Outpost software updates — Helm upgrade procedure + rollback