Terraform wandb-qa and wandb-production infrastructure.

[annirudh]

Background

We currently have the infra module which codifies resources in the wandb-qa and wandb-production GCP projects. Most of the infrastructure provisioned relates to the following services:

• gorilla
• weave-python

The terraform provisions the zoo-qa cluster in the wandb-qa project. We want to achieve a similar result for the wandb-production project, where applying the terraform in the listed modules provisions a GKE cluster in wandb-production called zoo-production, mirroring the same nodepool structure (but with larger node instance sizes as per the module variables).

Goals

We want to provision a new, parallel production environment that has fully codified versions of the following stateless resources:

• Static IP
• Kubernetes cluster
• Redis instances (1, 2, 3, 4)

Additionally, it would be good to import the following stateful resources:

• CloudSQL databases (production and its replicas, production-runs-us-central1-shard1 and its replica, production-usage)
• GCS buckets (wandb-production.appspot.com, wandb-artifacts-prod)
• Pub/Sub topics + subscriptions (filestream, wandb-production-gs, run-updates-v2, task-queue, task-queue-dead-letter)
• Bigtable (production-ssd, production-ssd-v3)

Our desired end-state is:

• All resources referenced in the QA and production Helm charts are managed by Terraform
• We generate a Helm values file which references the Terraform managed resources (will be new for stateless resources and same as existing for stateful ones)
• We can install the Helm chart onto the new, Terraform managed k8s cluster

Once this is done, we should have a fully functional mirror of our production environment that is completely managed by Terraform.