From Docker Compose to Kubernetes: Deploying Agentic Workloads

Two Deployment Targets

Agentic Assistants is designed to run in two modes: local development with Docker Compose, and production deployment on Kubernetes. The architecture is the same in both cases -- only the orchestration layer changes.

Docker Compose for Development

The docker-compose.yml defines the full stack:

services:
  agentic-api:
    build:
      context: .
      dockerfile: docker/api.Dockerfile
    ports: ["8000:8000"]
    volumes:
      - ./src:/app/src
      - ./conf:/app/conf
    environment:
      OLLAMA_HOST: http://host.docker.internal:11434

  agentic-ide:
    build:
      context: ./frontend
      dockerfile: Dockerfile
    ports: ["3000:3000"]

  mlflow:
    image: ghcr.io/mlflow/mlflow:latest
    ports: ["5000:5000"]

  redis:
    image: redis:7-alpine
    ports: ["6379:6379"]

  postgres:
    image: postgres:16-alpine
    environment:
      POSTGRES_DB: agentic
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD}

  minio:
    image: minio/minio:latest
    ports: ["9000:9000", "9001:9001"]

This gives you the complete platform with hot-reloading for development. Source volumes are mounted so code changes are reflected immediately.

Kubernetes Manifests

The k8s/ directory contains Kustomize-based manifests organized by concern:

k8s/
├── namespace.yaml
├── configmap.yaml
├── kustomization.yaml
├── agentic/
│   ├── deployment.yaml
│   └── ingress.yaml
├── dagster/
│   └── dagster-deployment.yaml
├── mlflow/
│   └── mlflow-deployment.yaml
├── postgres/
│   └── postgres-deployment.yaml
├── redis/
│   ├── redis-deployment.yaml
│   └── redis-configmap.yaml
└── cronjobs/
    ├── data-sync.yaml
    └── global-repo-ingestion.yaml

Each service gets its own directory with deployment, service, and any required configmaps or secrets.

Key Kubernetes Patterns

Resource limits. Every deployment specifies CPU and memory requests/limits. This is critical for ML workloads where a training job can consume all available resources if unconstrained.

CronJobs for maintenance. Data synchronization and repository ingestion run as Kubernetes CronJobs. The data-sync job runs every 6 hours to update cached datasets, and global-repo-ingestion re-indexes monitored repositories nightly.

ConfigMaps for configuration. The framework's conf/base/ directory is mounted as a ConfigMap, so configuration changes don't require image rebuilds.

Kustomize Overlays

Different environments use Kustomize overlays:

# kustomization.yaml
apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization

namespace: agentic-platform

resources:
  - namespace.yaml
  - configmap.yaml
  - agentic/deployment.yaml
  - mlflow/mlflow-deployment.yaml
  - postgres/postgres-deployment.yaml
  - redis/redis-deployment.yaml
  - dagster/dagster-deployment.yaml
  - cronjobs/data-sync.yaml
  - cronjobs/global-repo-ingestion.yaml

Migration Path

The migration from Docker Compose to Kubernetes is incremental. Start with the stateless services (API, workers), then migrate stateful services (Postgres, Redis, MinIO) once you have persistent volume claims configured. The framework's configuration system uses environment variables for service discovery, so the same application code works in both environments.