Repository navigation
feat(platform): lean system pool, fit the platform on one e2-medium when idle #208
Description
🤖 AI agent: requested by the owner after the rc.4 bootstrap validation on a private GKE instance.
Problem
With every agent scaled to zero, the system pool still runs 2 × e2-medium (Spot), because the platform and the GKE add-ons together request about 1.7 CPU and one e2-medium offers about 940m. Measured on the instance (requests per node 793m and 900m, 85% and 96% of allocatable). Most of that is GKE add-ons agentdev does not use or use only in gateway ingress mode:
| Add-on | Request (measured) | Used by agentdev? |
|---|---|---|
Managed Prometheus (gmp-operator, collectors) |
collector on each node | no (no reference in docs, Terraform or Ansible) |
Filestore CSI (filestore-node on each node, lock-release controller) |
25m + 160 MB per node | no (PD only) |
Monitoring components POD/DAEMONSET/DEPLOYMENT/STATEFULSET (kube-state-metrics) |
105m + 130 MB | no |
httpLoadBalancing + Gateway API (l7-default-backend) |
small | only with ingress.mode: gateway |
Must stay: Cloud Logging SYSTEM_COMPONENTS and WORKLOADS (the persistence step 1 of security events, docs/security-model.md), Dataplane V2, PD CSI, Workload Identity, konnectivity.
Plan (two stacked PRs, a third in the deployment repo)
- Customisation points (non-breaking, defaults unchanged): Terraform inputs
enable_managed_prometheus,enable_filestore_csi,monitoring_components,gateway_enabled. A deployment can then trim the cluster without patching the module. - Lean defaults (on top of 1): Managed Prometheus and Filestore CSI off, monitoring components reduced to
SYSTEM_COMPONENTS, Gateway and load balancer on only whenpublic_hostsis set. - A deployment repository sets the inputs for its instance (here
valentinalexeev/agentdev-valabs).
Acceptance
- On a live cluster with no public hosts, the platform fits on one
e2-medium: the system pool autoscaler settles at 1 node with all agents at 0. - Cloud Logging still carries
AGENTDEV_SECURITY_EVENTlines. terraform validatepasses; a plan on an existing instance shows in-place cluster updates only.
Questions
- Milestone for the lean defaults. Flipping defaults alters behaviour of existing instances, so under
CLAUDE.md§7 it is a MAJOR change once 1.0.0 is tagged. Options: (a) land it in 1.0.0 (still rc), (b) wait for 2.0.0, (c) keep the defaults as they are and ship only the customisation points in 1.0.x/1.1.0. Proposal: (a) if the owner is willing to hold 1.0.0 for it, otherwise (c). Blocks the lean-defaults PR only.
Activity
docs: add "add-on" to the STE technical names
fix(ansible): name gateway_enabled when ingress.mode gateway finds no…
[-]Lean system pool: fit the platform on one e2-medium when idle[/-][+]feat(platform): lean system pool, fit the platform on one e2-medium when idle[/+]