Azure Container Deployment

Azure Container Deployment

The deploy is one script: deploy-chuuk.sh. It deploys two Container Apps to a single Container Apps environment, sourcing secrets from Key Vault and Cosmos credentials at deploy time. No docker build is run locally — images are built remotely in ACR.

Resources (defaults; override via env)

The script sets the subscription explicitly (deploy-chuuk.sh), so accidentally being on the wrong subscription is not a failure mode.

Prerequisites (failure-fast)

ensure_keyvault_and_secrets (deploy-chuuk.sh) requires:

• The Key Vault exists and you have read access.
• Both flask-secret-key and google-cloud-api-key secrets exist in it.

If either is missing the script aborts before doing any work. Provision them once via:

az keyvault secret set --vault-name "$KEY_VAULT_NAME" --name flask-secret-key --value "$(openssl rand -hex 32)"
az keyvault secret set --vault-name "$KEY_VAULT_NAME" --name google-cloud-api-key --value "<your-key>"

Cosmos DB credentials are fetched at deploy time via az cosmosdb keys list and passed in directly — no Key Vault entry needed for them.

Build flow (no local docker)

build_images (deploy-chuuk.sh) runs:

az acr build --registry chuukdictregistry --image chuuk-dictionary-app:latest \
  --file Dockerfile --platform linux/amd64 .
az acr build --registry chuukdictregistry --image chuuk-ollama:latest \
  --file Dockerfile.ollama --platform linux/amd64 .

This pushes the build context up to ACR and builds inside Azure — works on M-series Macs where local linux/amd64 builds would otherwise be cross-arch.

PREPULL_LLM is not set by default — the Ollama image starts lean and pulls llama3.2:3b on first boot (~30s warm-up). Add --build-arg PREPULL_LLM=true (and a much larger image) if cold starts are a problem.

Container Apps configuration (current)

Main app (deploy-chuuk.sh):

• --cpu 2.0 --memory 4.0Gi
• --min-replicas 0 --max-replicas 2
• --target-port 8000 --ingress external (public HTTPS)

Ollama sidecar (deploy-chuuk.sh):

• --cpu 2.0 --memory 4Gi
• --min-replicas 1 --max-replicas 1 (always on; keeps the model loaded)
• --target-port 11434 --ingress internal (private FQDN, reachable only inside the env)

The internal Ollama FQDN is captured into OLLAMA_BASE_URL and injected into the main app's env (deploy-chuuk.sh).

Env vars passed to the main app

From deploy-chuuk.sh:

DB_TYPE=cosmos
COSMOS_DB_URI=<from az cosmosdb>
COSMOS_DB_KEY=<from az cosmosdb keys list>
FLASK_ENV=production
FLASK_DEBUG=0
FLASK_SECRET_KEY=<from KV>
GOOGLE_CLOUD_API_KEY=<from KV>             # if present
OLLAMA_BASE_URL=https://<internal-fqdn>    # if Ollama deployed

Important: secrets are passed as plain env values in the current script, not via secretRef. If you want to switch to secretRef, update deploy_main to first az containerapp secret set and then reference secretref:<name> in --env-vars — and remember to refresh secrets on every key rotation.

Architecture facts (don't forget)

• The frontend is bundled into the Flask image (COPY --from=frontend-builder /app/frontend/dist ./frontend/dist in Dockerfile). There is no Static Web App, no separate CDN, no Cloudflare. Same-origin everything.
• One Container Apps environment hosts both apps. Splitting them across environments breaks the internal FQDN-based connectivity to Ollama.
• Models are baked into the image (COPY models/ ./models/ in Dockerfile) — image size is ~2 GiB. Retraining requires a new image build + deploy. See production-retraining-orchestration.

Operational commands

# Tail logs
az containerapp logs show -n chuuk-dictionary -g rg-chuuk-beta-eastus2 --follow

# Open a shell in the running container
az containerapp exec -n chuuk-dictionary -g rg-chuuk-beta-eastus2 --command /bin/bash

# Force a restart by bumping the revision
az containerapp update -n chuuk-dictionary -g rg-chuuk-beta-eastus2 --image <same-tag>

Pitfalls

• The script auto-generates FLASK_SECRET_KEY and writes it to .env if missing — that's a local-only fallback. In CI/CD, ensure the Key Vault secret exists; otherwise users get logged out on every redeploy because the cookie signature changes.
• min-replicas=0 on the main app means cold starts. The first request after idle takes 5–15s. Don't set this to 1 without budget consideration.
• min-replicas=1 on Ollama is required — at 0, the model unloads and every translate request pays the model-load cost.
• If you change MAIN_APP_NAME/OLLAMA_APP_NAME, the deploy will create new resources rather than update existing ones. Always check az containerapp list -g $RESOURCE_GROUP -o table before re-running.
• Key Vault access: the deploy runs as the user invoking it, not as a managed identity — that user needs Key Vault Secrets User (or higher) on the vault.