Infrastructure readiness for RunPod inference workloads:
- runpod-inference.yaml: ConfigMap with pod creation payloads for RTX 4090,
A40 (single+dual), and custom templates
- runpod-lb-configmap.yaml: nginx least-conn load balancer for inference
endpoint distribution (Deployment + ClusterIP Service)
- runpod-provision.sh: bash provisioner script — reads RUNPOD_API_KEY/HF_TOKEN,
creates pods via GraphQL, polls until RUNNING, outputs endpoint URLs.
Does NOT spin up any pods (dry-run flag available).
The `alloy` ClusterIP service only exposes port 12345 (UI/metrics).
OTLP HTTP ingestion on port 4318 is on the separate `alloy-otlp` service.
Pointing OTEL_EXPORTER_OTLP_ENDPOINT at `alloy:4318` caused continuous
SocketTimeoutException in neuron-mcp. Updated both blue and green deployments.
Green manifest was committed but missing from kustomization.yaml so Argo CD
never created the resource. Wires it in so the blue/green slot swap can work.
Remove neuron/neuron-alpha/neuron-beta/neuron-gamma Deployments and
Services, the slot LB service, Python ingress, slot PVCs, snapshot job,
neuron-config ConfigMap, and neuron-secrets ExternalSecret. The Python
MCP backend has been superseded by the Spring Boot backend in neuron-prod.
Cloudflared Deployment and cloudflared-secret ExternalSecret are
retained in the neuron namespace. The neuron-data PVC is preserved
for data safety.
Single-node k3s cluster cannot schedule a second pod during rolling update
(512Mi request × 2 exceeds available memory). Recreate terminates the old
pod before starting the new one, trading brief downtime for schedulability.
Add marketing/ manifests (Deployment, Service, ExternalSecret) for the
Next.js marketing site in neuron-prod namespace, an Argo CD app pointing
to the new path, and update the prod ingress to serve / from neuron-marketing
instead of neuron-rest.
OAuth discovery endpoints (/.well-known/oauth-protected-resource and
/.well-known/oauth-authorization-server) and the authorization server
paths (/oauth2/*, /login) were routing to neuron-rest, which has none
of those endpoints. MCP also uses /sse and /message directly.
All five path prefixes now route to neuron-mcp on port 8080.
Adds the neuron-tim namespace (via Terraform) and a full Argo CD
app + k8s manifests mirroring neuron-prod. All deployments set to
replicas=0 — nothing runs until Tim's imprint is ready and Vault
secrets are populated at secret/neuron-technologies/tim.
Deploy single-broker Redpanda in dedicated namespace with 20Gi storage,
1 CPU / 2Gi memory limits, and a one-shot topic init job for neuron.swarm,
neuron.ci, neuron.vcs, and neuron.webhooks with appropriate retention.
Will Anderson