Deployment

This is the detailed deployment reference for the Dory Orchestrator. The orchestrator's own README covers the quick build/deploy; this page documents the full build pipeline, manifests, RBAC, database schema, secrets, and Karpenter integration.

See Configuration for the flags/env wired in here, HTTP API Reference and Metrics for the exposed surfaces, and Edge Failover for the failover behavior these manifests enable.

Prerequisites

Requirement	Version / value
Go	1.25+
k8s.io/client-go	v0.35
EKS	1.33+ (tested 1.35)
Karpenter	v1.7+
System nodegroup	role=system label, taint node-role=system:NoSchedule (hosts the orchestrator)
ECR registry	<ACCOUNT_ID>.dkr.ecr.us-east-2.amazonaws.com (region us-east-2)

Build

# Binary
go build -o orchestrator ./cmd/orchestrator

# Container image (amd64)
docker build --platform linux/amd64 -t dory-orchestrator:latest ./cmd/orchestrator

Run locally against a cluster:

export DORY_DATABASE_URL=postgres://...
export KUBECONFIG=~/.kube/config
go run ./cmd/orchestrator --namespace default --poll-interval 30s --log-level info

The pushed image is referenced as <ACCOUNT_ID>.dkr.ecr.us-east-2.amazonaws.com/dory-orchestrator:latest.

Applying manifests

The manifests live in deploy/. Apply RBAC, ConfigMap, Deployment, and Services.

Deployment

deploy/deployment.yaml — Deployment dory-orchestrator in namespace dory-system:

Setting	Value
replicas	1
strategy	Recreate
ServiceAccount	dory-orchestrator
Pod securityContext	runAsNonRoot, runAsUser 65534, fsGroup 65534
nodeSelector	role: system
toleration	node-role=system:NoSchedule
image	<ACCOUNT_ID>.dkr.ecr.us-east-2.amazonaws.com/dory-orchestrator:latest
imagePullPolicy	Always
Container securityContext	allowPrivilegeEscalation: false, readOnlyRootFilesystem: true, runAsUser 65534, drop ALL
Port	metrics 8080
Resources (requests)	cpu 100m, mem 128Mi
Resources (limits)	cpu 500m, mem 512Mi
Liveness	GET /livez (initialDelay 30 / period 10 / timeout 5 / failure 3)
Readiness	GET /readyz (10 / 5 / 3 / 3)
terminationGracePeriodSeconds	90

Args:

--config-db=$(DATABASE_URL) --namespace=$(NAMESPACE) --poll-interval=$(POLL_INTERVAL) --log-level=$(LOG_LEVEL) --enable-monitor

Env:

Env var	Source
DATABASE_URL	Secret dory-db-secret key database-url
NAMESPACE	ConfigMap dory-orchestrator-config
POLL_INTERVAL	ConfigMap dory-orchestrator-config
LOG_LEVEL	ConfigMap dory-orchestrator-config
DORY_STATE_TOKEN	Secret dory-state-secret key state-token (optional)

Services

deploy/service.yaml — two ClusterIP Services:

Service	Port mapping	Notes
dory-orchestrator-metrics	8080 → metrics	Prometheus scrape annotations; edge pods reach the orchestrator here.
dory-orchestrator	80 → metrics

Edge pods address the orchestrator at dory-orchestrator-metrics.dory-system.svc.cluster.local:8080.

ConfigMap

deploy/configmap.yaml — dory-orchestrator-config:

Key	Value
NAMESPACE	default
POLL_INTERVAL	10s
LOG_LEVEL	info

RBAC

deploy/rbac.yaml defines ServiceAccount dory-orchestrator (ns dory-system), a ClusterRoleBinding to ClusterRole dory-orchestrator:

Resource	Verbs
namespaces	get, list, watch
pods	get, list, watch, create, delete, patch, update
pods/status, pods/log	get
pods/eviction	create
nodes	get, list, watch, patch, update, delete
nodes/status	get, patch, update
events	create, patch
configmaps	get, list, watch, create, update (drain sentinel)
secrets	get, list, watch
karpenter.sh nodepools, nodeclaims	get, list, watch
karpenter.k8s.aws ec2nodeclasses	get, list, watch

A separate processor ServiceAccount dory-processor (ns default) plus a Role grants full configmaps verbs for SDK state persistence to dory-state-{processor_id}.

Secrets

Secret	Key	Use
dory-db-secret	database-url	PostgreSQL connection string.
dory-state-secret	state-token	Optional state transfer token (DORY_STATE_TOKEN).
ecr-registry-secret	docker-registry	Image pull secret (refreshed by CronJob — see below).

Database schema

PostgreSQL backs the orchestrator. Key tables:

Table	Notes
processors	One row → one pod. Columns: id UUID, processor_template_id FK→processor_templates.id, node_type, node_id FK→edge_nodes.id, k8s_namespace, k8s_pod_name, node_ip VARCHAR(45), sensor_id FK→sensors.id, status (pending/starting/running/terminated/failed/failover), health_status jsonb, last_health_check_at, consecutive_failures, terminated_at, created_at, updated_at.
processor_templates	The application / "slug" table (formerly processing_applications; PK still named processing_applications_pkey). Columns: id, slug, name.
processor_template_versions	processor_template_id FK, image_uri, digest, version, runtime_config_template JSON, build_config JSON, is_active bool.
sensors	id, sensor_type, connection_config jsonb, metadata jsonb, location_point PostGIS.
edge_nodes	id, organization_id, name, status (online/offline/decommissioned), failover_enabled, failover_target_node_id, last_heartbeat_at.
edge_node_apps	id, edge_node_id, processor_config_id FK→processor_templates.id, failover_enabled, state_storage_path, status (active/failover/stopped), current_processor_id.
edge_node_events	id, edge_node_id, processor_config_id (nullable), event_type, details jsonb, created_at.

Warning

The two edge_node_* tables key on column processor_config_id, while processors and processor_template_versions use processor_template_id. Both FK to processor_templates.id. Mixing them up causes column ... does not exist (SQLSTATE 42703) errors.

Karpenter

Manifests in deployments/karpenter/.

NodePool dory-app-pool (karpenter.sh/v1)

Field	Value
Template labels	workload-type=application, managed-by=dory-orchestrator
Taints	none
Requirements	arch amd64, os linux, capacity-type on-demand, instance-types [t3.small, t3.medium, t3a.small, t3a.medium]
nodeClassRef	dory-app-nodeclass
Limits	cpu 100, memory 400Gi
Disruption	consolidationPolicy WhenEmpty, consolidateAfter 3m, budget 20%
Weight	10

EC2NodeClass dory-app-nodeclass (karpenter.k8s.aws/v1)

Field	Value
amiFamily	AL2023, alias al2023@latest
IAM role	KarpenterNodeRole-dory-demo
Subnet & SG	tag karpenter.sh/discovery: dory-demo
kubelet	reservations configured
blockDeviceMappings	/dev/xvda gp3 30Gi
IMDS	IMDSv2 required
detailedMonitoring	true

ECR token refresh

CronJob ecr-token-refresh (ns kube-system):

Field	Value
Schedule	0 */6 * * * (ECR tokens expire after 12h)
Action	aws ecr get-login-password (region us-east-2, registry <ACCOUNT_ID>.dkr.ecr.us-east-2.amazonaws.com)
Result	Recreates docker-registry secret ecr-registry-secret in namespaces default, kube-system, dory-system.

Warning

The orchestrator must run on a system node. System pods require nodeSelector: {role: system} and a toleration for the node-role=system:NoSchedule taint — the Deployment already sets both.