Deployment Examples

Reference deployments of the FermiHDI stack for Ansible, Chef, Puppet, Kubernetes (raw manifests and Helm) and Nomad. Each subdirectory has its own README.

All of these assume DMS-managed nodes

Every example configures the Storage Cluster Controller and Storage Nodes with identity only — DMS URL, node name, bootstrap token, certificate material and the per-plane subnets. None sets RECORD_SIZE, SCC_IP, MASTER_CPUSET or a storage device address, and that is correct for this deployment mode: the DMS delivers record size, the SCC address, storage devices and the hash election while each node's data plane waits at its boot gate.

A node reporting DP_HOLDING under these examples is waiting for the DMS, not misconfigured. If it never releases, investigate the DMS side.

For a pre-provisioned or standalone deployment — no DMS config pushes — those settings become required and the node exits without them. See the per-component configuration guides:

• hd_scc/docs/configuration_guide.md
• hd_sn/docs/configuration_guide.md
• hd_proxy/docs/configuration_guide.md
• hd_dms/docs/configuration_guide.md

Two things to check before adapting these

1. Telemetry export is not configured. The examples set FERMIHDI_OTEL_URL, which is a routing probe used to pick the node's observability IP — it configures no exporter. The OpenTelemetry SDK reads OTEL_EXPORTER_OTLP_ENDPOINT and OTEL_EXPORTER_OTLP_PROTOCOL, and neither appears in any example. Add them if you expect traces and metrics to leave the node.

2. The gRPC ports are not declared. Beyond the documented 6986 / 6987 / 8080, the cluster uses three derived gRPC listeners:

No example declares them in a containerPort, Service or Nomad port block. Within a flat pod or host network this is cosmetic, since the traffic is direct peer-to-peer. It stops being cosmetic the moment you add a default-deny NetworkPolicy, a service mesh or a host firewall — the existing NetworkPolicies in the Kubernetes example cover only node→DMS traffic on 6985/6986/6989, so adding cluster-wide isolation without these three ports silently breaks batch verification and query fan-out.

That failure is quiet by design elsewhere in the stack: gRPC's BuildAndStart() does not fail when a listener cannot bind, and a blocked verification stream presents as batches that never get confirmed rather than as a connection error.

Also note: FERMIHDI_FORCE_POSIX is set on the SCC and SN in several examples. Neither binary reads it — it is a client-side switch in hd_sdk. Leave it on ingester and proxy containers; it is inert on the storage components.

Verified against hd_scc@c915937, hd_sn@fea7c61, hd_proxy@2888e6c and hd_dms@6311f08 on 2026-08-18.