Adds the real-router BMP extension code and a redesign doc capturing
what changed in the test's intent.
scripts/router_bmp.py push/dry-run/remove `bmp server 2`
on all 18 lab routers via paramiko
SSH shell. Inventory + cred pattern
lifted from obmp-rib-poller/poller.py.
Activation scope: bmp-activate server 2
on the existing BMP-MONITORED
neighbor-group -- mirrors production.
terraform/router_bmp.tf terraform_data resource with two-step
safety (dry-run first, apply on
confirm), triggers_replace on action
change, destroy provisioner calls
`remove` (idempotent, on_failure=
continue so a stuck router doesn't
block VM teardown).
terraform/terraform.tfvars enable_router_bmp / confirm_router_push
both default false (dormant).
docs/redesign-real-routers.md captures the pivot from v1 "prove
deploy.sh works" to v2 "canary
collector receiving real router BMP".
Documents that v1 store VM OOM'd at
4 GB under the internal --profile
test feed alone, proposes 32 GB /
8 vCPU for v2 and dropping --profile
test so the collector receives ONLY
real router BMP.
State on prox940 as of this commit:
- v1 VMs (9001 store, 9002 bmpgen) destroyed via PVE API
(terraform destroy hung on the OOM'd guest agent)
- Terraform state cleaned
- Cached Ubuntu image kept for a fast next apply
- Lab routers untouched (extension code dormant)
Not yet done, awaiting user input:
- v2 sizing landed in terraform.tfvars
- Compose scope decision (add --no-feeders to deploy.sh vs
workaround in cloud-init)
- Re-apply
- Router extension dry-run + confirm
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The initial two-VM run against a real Proxmox host found three real
bugs in the obmp-docker stack, plus one design issue in this test:
Fixed in obmp-docker (deploy-refinement branch):
- 745f066 deploy.sh: probe sudo with `sudo -n true` instead of
`sudo -v` / `sudo -n -v`; the latter demands a credential
timestamp that NOPASSWD never produces, so both fail on
cloud-init / CI even with NOPASSWD:ALL configured.
- 6df67a7 .env.example: quote GNMI_ADDRESSES with single quotes so
docker compose's .env parser doesn't choke on the mid-line
comma in the double-quoted list form.
Fixed here:
- bmpgen cloud-init: add `port = 1179` under [global.config] to
avoid gobgpd's 179 listener bind failing under Ubuntu 24.04 +
Docker even with --network host and default NET_BIND_SERVICE cap.
We don't accept BGP peers on bmpgen so the port value doesn't
matter as long as it's unprivileged.
- run-test.sh: fix column name for bgp_peers.state -- the opstate
enum uses lowercase 'up', not uppercase 'UP'.
Verified against the live run: 6/6 assertions pass, ip_rib landed
251,760 routes from the store VM's own gobgp connecting to Bromirski's
public full-table feed via --profile test.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Terraform (bpg/proxmox) provisions two Ubuntu 24.04 cloud-init VMs on a
Proxmox host, one running obmp-docker's `deploy.sh --scope full-stack`
and the other running GoBGP as a BMP-source stand-in. `scripts/run-test.sh`
runs the six assertions from docs/design.md (containers healthy, BMP
peer up, ip_rib/bgp_peers non-zero, Grafana healthcheck).
Layout:
terraform/ provider + VM + cloud-init snippet resources
terraform/cloud-init/ user-data templates (store + bmpgen)
scripts/ run-test.sh, teardown.sh
docs/design.md topology, assertion set, and rationale
Cred handling:
terraform.tfvars is gitignored; endpoint + api_token can also come
from PROXMOX_VE_ENDPOINT / PROXMOX_VE_API_TOKEN env vars.
Not validated locally: `terraform init && terraform validate` -- the
terraform binary is not on the authoring host. YAML + shell + HCL brace
balance verified.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
