Running Containers Without Kubernetes
Updated
IntermediateWhat you'll learn
~20 min- Deploy a container to a managed container platform with no cluster to own
- Wire secrets, scaling rules, and health probes at the platform layer
- Enumerate exactly which operational concerns the platform absorbed -- and verify the list
Rung 3 pays off
Lesson 2.2 produced a production-grade container image for AssetTrack’s API. The ladder placed it on rung 3: a managed container service — concretely, Azure Container Apps. Today you deploy it and take inventory of everything you did not have to build.
Prompt first
Switch your workspace agent to Plan mode on your containerized API project and ask it to plan the managed container deployment (if you do not have live exports or files handy, ask it to generate a small Meridian practice sample from the lesson’s scenario first):
Plan the az containerapp CLI deployment for assettrack-api on Azure Container Apps. Use managed identity for registry and Key Vault secrets, keep at least one warm replica for storm-day operations, and list what the platform absorbs versus Kubernetes.Check that the plan or draft covers:
- Pulling
assettrack-apifrom Azure Container Registry using a system-assigned managed identity (--registry-identity system) and a pinned version tag - External HTTPS ingress on port 8080 with platform-managed TLS and readiness/liveness probes against
/healthz - Resolving the database connection string from Azure Key Vault via managed identity rather than inline secrets
- Scaling from
1to5replicas on HTTP concurrency of 50 (never scaling to0so storm-day dispatches never hit cold starts), plus an inventory of cluster ops concerns absorbed by the platform
If the draft uses --min-replicas 0 or :latest, follow up with: Pin the image tag to a release version and set --min-replicas 1 so field operations never hit a cold start.
The shape of the answer:
REFERENCE SCRIPT Azure Container Apps provisioning script for assettrack-api open reference hide reference Reference example of the az containerapp provisioning script your agent generates. Open to inspect the managed identity, Key Vault secret reference, and scaling flags.
az containerapp env create \ --name meridian-prod --resource-group rg-meridian \ --location eastus2
az containerapp create \ --name assettrack-api \ --resource-group rg-meridian \ --environment meridian-prod \ --image acrmeridian.azurecr.io/assettrack-api:1.4.2 \ --registry-identity system \ --ingress external --target-port 8080 \ --min-replicas 1 --max-replicas 5 \ --scale-rule-name http-concurrency \ --scale-rule-type http \ --scale-rule-http-concurrency 50 \ --secrets "db-conn=keyvaultref:https://kv-meridian.vault.azure.net/secrets/assettrack-db,identityref:system" \ --env-vars "ConnectionStrings__Oracle=secretref:db-conn"Three details in those flags carry most of this lesson’s weight:
--registry-identity systemandkeyvaultref:...,identityref:system— the app pulls its image and reads its secret using a managed identity. There is no registry password and no connection string in any config file, pipeline variable, or shell history. Credential-less by construction.--min-replicas 1— scale-to-zero is a great default for internal tools, and the wrong one here: a cold start in the middle of a storm-day dispatch surge is a field tech staring at a spinner. Scaling policy is a business decision expressed in one flag.- Version-pinned image tag —
:1.4.2, not:latest. Deploys should change because you chose to change them. (Lesson 2.5 has more to say about:latest.)
The inventory: what you never built
This table is the argument for rung 3. For this one API, the platform absorbed:
| Concern | On Container Apps | On a cluster you own |
|---|---|---|
| TLS certificates | Issued and rotated by the platform | Your cert-manager config, your renewals |
| OS and node patching | Invisible to you | Your maintenance windows |
| Ingress routing | One flag | Ingress controller to choose, deploy, patch |
| Autoscaling | One scale rule | Metrics pipeline + HPA tuning |
| Zero-downtime deploys | Revisions built in | Rollout strategy + PodDisruptionBudgets |
| Registry auth & secret storage | Managed identity references | Pull secrets, CSI drivers, rotation story |
Every row is work that exists somewhere — the question the ladder keeps asking is whether it exists on your payroll.
In multiple-revision mode, Container Apps keeps prior revisions addressable. Ship 1.4.3, watch it misbehave, and rollback is one command: route traffic back to the 1.4.2 revision. You get blue-green semantics without building blue-green infrastructure — remember this in Lesson 4.3 when the promotion runbook needs its backout step.
When rung 3 is not enough
Honesty requires the failure modes. Managed container platforms will fight you when the workload needs: privileged containers or kernel modules, custom node types or GPUs with exotic drivers, long-lived non-HTTP protocols the ingress cannot route, or network topologies the platform cannot express. Those are real forcing tests — name one, and Lesson 2.4’s descent to Kubernetes is justified. AssetTrack’s API needs none of them.
The deploy references the database credential as a Key Vault reference resolved through a managed identity, rather than passing the connection string directly. What is the operational win?
Key takeaway
A managed container platform gives you the container’s benefits — dependency control, immutable versioned artifacts, instant rollback via revisions — while absorbing TLS, patching, ingress, scaling, and secret plumbing. That inventory table is your template: run it for any workload, and if every row lands in the platform column, you have no business on a cluster. The one Meridian workload that fails that test is next.