Skip to content

v0.166.0 — Prod on Google Cloud: the server, not yet reachable

The second step of the Google Cloud program's M3b. Nothing changes for players: production stays on AWS at the same addresses. This release builds production's server on Google Cloud, so the move can be rehearsed with a copy of production's data before the window. The program's documents are in docs/programs/gcp/; the move's runbook is infra/gcp/README.md, "Prod's move (M3)".

On Google Cloud

  • [x] Production's server and load balancer. Production's stack applies its second pass: the Cloud Run service (one always-on instance, 2 GiB), the load balancer and its address, and certificates for map, api, assets and help, proven by DNS records before any traffic moves. The service starts on a placeholder image, which production's own replaces on its first deploy.
  • [x] Deploying production on demand. GCP Deploy Prod gains Run workflow, for either owner on main: it deploys main's head to production on Google Cloud without waiting for a merge. It still does nothing while PROD_HOST is unset.
  • [x] 2 GiB per server. Server-side Remove Background loads a 213 MB model, which ran a 1 GiB instance out of memory on next. Every environment on Google Cloud now has 2 GiB. The remaining failure is issue #176.

Still to come (M3b)

  • The certificates' DNS records, production's first real deploy, and PROD_HOST=both.
  • The rehearsal, with a copy of production's data.
  • The window: AWS's production stopped, the final copy, DNS switched, PROD_HOST=gcp.