STREAMABLE transport holds session state in-memory per-pod. With HPA
scaling to 2+ replicas and round-robin routing, requests hit different
pods and fail with "Session not found". Pin each client to one pod via
a secure HttpOnly cookie.
BuildKit rootless failed on k3s (mount propagation), privileged mode
fixed that but buildkitd gRPC is incompatible with the Docker REST API
that forgejo-runner needs to manage job containers. Net security change
vs. the original socket approach was zero (privileged ≈ socket access).
Remove buildkitd.yaml entirely. Restore docker-sock hostPath mount on
both runners. Builds work again.
Add MARKETPLACE_DB_URL/USER to ConfigMap and MARKETPLACE_DB_PASSWORD
ExternalSecret (sourced from secret/legion-db in Vault). Remove the
SQLite subPath volume mount and fix-data-ownership initContainer from
the blue deployment — marketplace storage is now in Postgres.
subPath mounts require the source file to exist on the PVC before the
main container starts. Add 'touch /data/neuron-marketplace.db' to the
initContainer so the subPath mount can bind the file at /app/neuron-marketplace.db.
App hard-codes relative path 'neuron-marketplace.db' from working dir /app.
Mount the PVC file at /app/neuron-marketplace.db via subPath so the app can
write to it as UID 1000. PVC persists the file across pod restarts.
Also keep NEURON_MARKETPLACE_DB_PATH in ConfigMap for future app versions.
workingDir: /data broke the JVM (app.jar is a relative path in /app).
Add NEURON_MARKETPLACE_DB_PATH=/data/neuron-marketplace.db env var
following the same pattern as NEURON_DB_PATH for the core DB.
neuron-marketplace.db was opening relative to /app (image root,
not writable by UID 1000). Set workingDir to /data so all relative
file opens land on the writable PVC volume.
PVC was written as root before runAsUser: 1000 was added; SQLite
refused writes with SQLITE_READONLY. initContainer runs as root to
chown the volume, then the app runs as UID 1000 as required.
Pod was crash-looping: image has no USER directive so kubelet rejected it
with runAsNonRoot. Add runAsUser: 1000 at both pod and container level.
MCP paths moved from neurontechnologies.ai (now a GCP Cloud Run A record,
unreachable on Legion) to private subdomain neuron.neurontechnologies.ai.
Cloudflare tunnel and DNS CNAME updated out-of-band.
Adds complete k8s manifests, ArgoCD app, and Terraform namespace
resources for the Neuron swarm self-improvement loop system.
Each variant (alpha, beta, gamma) gets its own isolated namespace,
PVC, MCP/REST deployments, ExternalSecrets from Vault, RBAC for CI,
and a SQLite clone Job template for session startup.
Architecture: intelligence stays on Legion; only compiled artifacts cross
to GCP. Source code and Neuron's knowledge base never leave the system.
Artifact Registry:
- neuron-marketing, neuron-accounts, neuron-api repos in us-central1
- Keep-last-10 cleanup policy; ci-pusher SA with writer access
- Legion CI runners authenticate via GCP_SA_KEY Gitea secret
Cloud SQL (cloud-sql.tf):
- postgres-15 on db-g1-small, us-central1 (scale up to REGIONAL HA at 1k users)
- Point-in-time recovery, 14-day backup retention
- Accounts DB + user; password generated and stored in Secret Manager
- JWT signing key in Secret Manager (shared by accounts + api)
- Cloud Run connects via built-in Auth Proxy (Unix socket volume mount)
Accounts Cloud Run (cloud-run-accounts.tf):
- 3 regions (us-central1, europe-west1, asia-northeast1), min:1 max:50
- Cloud SQL proxy volume mount; secrets via Secret Manager
- Stripe + JWT env vars; health probe on /health
API Cloud Run (cloud-run-api.tf):
- 3 regions, min:1 max:100, cpu_idle=false (always-hot)
- Validates JWTs from accounts service; no direct DB connection
- License admin token from Secret Manager
Load balancer (host-based routing):
- Same global anycast IP for all three services
- URL map routes by Host: neurontechnologies.ai→marketing,
api.neurontechnologies.ai→api, accounts.neurontechnologies.ai→accounts
- New managed SSL certs for api.* and accounts.* added to HTTPS proxy
- Cloud Armor (WAF + rate limit) applied to all backends
Service accounts + IAM:
- neuron-accounts-sa: secretmanager.secretAccessor + cloudsql.client
- neuron-api-sa: secretmanager.secretAccessor
- allUsers invoker on all prod Cloud Run services (LB health checks)
bootstrap.sh:
- One-shot setup: pulls Stripe secrets from Vault → Secret Manager,
creates CI SA JSON key, prints DNS + next-step instructions
Replace the aspirational alpha/beta/gamma model with the actual
deployment topology: prod runs blue/green in neuron-prod namespace,
stage is the single experiment slot in neuron-stage namespace.
The old script referenced neuron-alpha/beta/gamma deployments that
never existed. The new script uses blue-green-deploy.sh for prod
promotion and kubectl set image for stage experiments.
Loop: snapshot → deploy stage → evaluate → promote via blue/green.
Bucket created, SA key stored in Vault at secret/gcs.
CronJob ExternalSecret updated to pull from secret/gcs.
Hourly restic backup now runs to both R2 and GCS.
Will Anderson
will.anderson