Deployment#
Four supported deployment modes, ordered by typical maturity:
- Docker Compose — single-node dev / laptop / tiny PoC. The quickstart mode.
- Kubernetes + Helm — recommended production default. Multi-replica, HPA, cluster-native.
- Terraform modules — AWS + GCP stubs for customer-VPC deployments.
- Air-gapped — offline / banking-core / defence. Helm chart with pinned images.
Cross-cutting concerns#
- Data residency — India / EU / US regionality, BYOK
- Upgrade & versioning — LTS, semver, migration paths
Picking a mode#
| If you're … | Use |
|---|---|
| Running on a laptop / demo | Docker Compose |
| Standing up a dev / staging env | Docker Compose → Kubernetes |
| Evaluating for an enterprise PoC | Kubernetes + Helm in your VPC |
| Deploying in BFSI / healthcare with compliance sensitivity | Kubernetes + Terraform module (customer-VPC) |
| Running in a zero-egress zone (defense, banking core) | Air-gapped Helm with pinned image bundle |
Minimum resources#
| Mode | CPU | RAM | Storage |
|---|---|---|---|
| Docker Compose (dev) | 4 cores | 8 GB | 20 GB |
| Kubernetes (single-node) | 6 cores | 16 GB | 50 GB |
| Kubernetes (multi-replica) | 12 cores | 32 GB | 200 GB (Postgres) + 500 GB object store |
Scale up for LLM workloads — the swarm API itself is lightweight, but training workloads + concurrent batch runs can push RAM hard.