HydraIssues

Issues / #716
GCP bootstrap: upgrade billing to paid, L4 quota in europe-west1, connect in iamnim
open unclassified Priority: high Project: hydrahyperscaler Parent: #714 Reporter: 14 Sep 2026 11:22

Description

Child of #714. GCP is the first-class provider.

STATE 2026-09-14 evening: GCP IS LIVE AND MEASURING

  • Project hydrahyperscaler, number 185030409246, ACTIVE. Billing UPGRADED to paid (billingAccounts/01698B-F2BD9A-B12123, billingEnabled true). Compute Engine API enabled. gcloud 584.0.0 installed on neoremote.
  • Quota picture, verified by API, not by reading console rows:
    • NVIDIA_L4_GPUS europe-west1 = 1 OK
    • CPUS europe-west1 = 200 OK
    • GPUS_ALL_REGIONS = 0 ← this was the real blocker. The owner saw a 1 on an NVIDIA row and reasonably read it as done, but a GPU instance needs BOTH the regional L4 quota AND the global all-regions cap. The global cap was zero.
    • g2-standard-8 confirmed available in europe-west1-b.
  • Quota increase FILED BY API (Cloud Quotas API, QuotaPreference gpus-all-regions-1, GPUS-ALL-REGIONS-per-project, preferred 4). Requested 4 rather than 1 so a pilot fleet does not need a second two-day wait. Status: pending, grantedValue still 0. Note: the Cloud Quotas API needs an x-goog-user-project header when called with a Cloud Shell token, otherwise it 403s against Cloud Shell's own project.
  • Reflector LIVE: reflector-gcp-europe-west1, e2-micro in europe-west1-b, 34.79.137.44, running hydraprobe server from the repo cloud-init, firewall rule hydraprobe-udp2112 on tag hydraprobe-reflector. Needs no GPU, so the campaign is NOT blocked by the quota.
  • FIRST REAL MEASUREMENT PASSES. 60 s B1 from neoremote (Hetzner Falkenstein) to the reflector: RTT p50 10.698 ms, p95 10.976, p99 11.35, max 16.472; IPDV p99 0.902 ms against the 5 ms gate; zero loss, zero reordering; verdict PASS. Committed to hydrahyperscaler docs/testbooks/region-qualification/results/. This is from GERMANY, so it is the pessimistic bound. Belgian venues should measure better.

Credential note

Bootstrap is running on a SHORT-LIVED Cloud Shell access token pasted by the owner, expiring about an hour after 21:10 UTC. No long-lived cloud credential is stored on neoremote. The durable path is unchanged: the owner pastes a service account key into iamnim, mycelium vends it, nimsforesthyperscaler consumes it. That paste has NOT happened yet.

Next

When the quota is granted: g2-standard-8 plus nvidia-l4-vws Windows smoke test (#707, gate G2b). Independently of the quota: extend the probe campaign to the real venues (#706), which needs the hydraprobe binary on venue probe nodes and the shared HMAC key distributed out of band.