Compute Overview

How Trovella's production compute infrastructure works -- a single Compute Engine VM running Docker Compose with Caddy, Next.js, and service containers.

Trovella runs on a single Google Compute Engine VM (e2-custom-2-6144) in us-central1-a. All services run as Docker containers orchestrated by Docker Compose. Caddy handles TLS termination and reverse proxying. There is no auto-scaling, no load balancer, and no multi-zone redundancy -- this is a deliberate MVP trade-off for simplicity and predictable cost.

Architecture at a Glance

Internet
  |
  | HTTPS (443) / HTTP (80, redirects to 443)
  v
Caddy (caddy:2-alpine)
  |
  | reverse_proxy web:3000
  v
Next.js (custom image, standalone server.js)
  |
  |-- cloud-sql-proxy:5432 --> Cloud SQL PostgreSQL 18
  |-- typesense:8108       --> Typesense 27.1 (search)
  |-- inngest:8288         --> Inngest Dev Server (background jobs)

All inter-service communication happens over the Docker internal network with sub-millisecond latency. No traffic between containers traverses the public internet.

Production Stack

Five containers run on the VM, defined in infra/docker-compose.prod.yml:

Container	Image	Port	Purpose
`caddy`	`caddy:2-alpine`	80, 443 (public)	TLS termination, HTTP/3, reverse proxy, www redirect
`web`	`us-central1-docker.pkg.dev/trovella-shared/trovella/web:latest`	3000 (internal)	Next.js application (standalone `server.js`)
`cloud-sql-proxy`	`gcr.io/cloud-sql-connectors/cloud-sql-proxy:2.21.2`	5432 (internal)	IAM-authenticated tunnel to Cloud SQL
`typesense`	`typesense/typesense:27.1`	8108 (internal)	Keyword search engine with persistent data volume
`inngest`	`inngest/inngest:latest`	8288 (internal)	Durable background job execution

Only Caddy exposes ports to the public internet. All other services communicate on the Docker internal network via service names.

VM Specifications

Property	Value
Machine type	`e2-custom-2-6144` (2 vCPU, 6 GB RAM)
OS	Ubuntu 24.04 LTS
Disk	50 GB `pd-standard`
Zone	`us-central1-a`
Static IP	Reserved via Terraform
Monthly cost	~$37
Upgrade path	`e2-standard-2` (8 GB, ~$49/month) -- one `gcloud` command, ~2 min downtime

Memory Budget

Service	Estimated Usage
Next.js	768 MB -- 1 GB
Typesense	1 -- 2 GB
Inngest	256 -- 512 MB
Cloud SQL Proxy	Minimal (~50 MB)
OS + Docker overhead	~512 MB
Total	~3.5 -- 5.5 GB (fits in 6 GB with headroom)

Local Development Stack

Local development uses docker-compose.yml (root of repo) with different services:

Container	Image	Port	Notes
`postgres`	`pgvector/pgvector:pg18`	5433	Local PostgreSQL (no Cloud SQL Proxy needed)
`redis`	`redis:8-alpine`	6379	Local Redis (production uses Upstash)
`mailpit`	`axllent/mailpit:latest`	1025, 8025	SMTP testing (UI at `localhost:8025`)
`inngest`	`inngest/inngest:latest`	8288	Dev mode (`inngest dev`)
`typesense`	`typesense/typesense:27.1`	8108	Same version as production

Start with pnpm docker:up, stop with pnpm docker:down.

Terraform Module

The VM and its associated resources are defined in infra/modules/compute-vm/:

File	Contents
`main.tf`	Service account, firewall rules, static IP, VM instance
`variables.tf`	Machine type, disk size, zone, notification email
`startup.sh`	First-boot script: installs Docker, configures Artifact Registry auth, installs Cloud Ops Agent
`monitoring.tf`	Alert policies (CPU, memory, disk) and HTTPS uptime check
`outputs.tf`	External IP, instance name, zone, service account email

The module is instantiated in infra/environments/prod/main.tf alongside the Cloud SQL and Secret Manager modules.