Containers Without the Cluster
You want the portability of containers without managing Kubernetes. Or you have Kubernetes but want scale-to-zero for bursty workloads. Serverless containers solve both.
The Contenders
Google Cloud Run
# Deploy a container in one command
gcloud run deploy my-api \
--image gcr.io/myproject/api:v1 \
--region europe-west4 \
--allow-unauthenticated \
--min-instances 0 \
--max-instances 100 \
--memory 512Mi \
--cpu 1 \
--concurrency 80Strengths: Best developer experience, fastest cold starts (~200ms), generous free tier, scale-to-zero Weaknesses: GCP only (unless using Cloud Run on GKE), 60-min max request timeout
AWS Fargate
# ECS Fargate task definition
{
"family": "my-api",
"networkMode": "awsvpc",
"requiresCompatibilities": ["FARGATE"],
"cpu": "512",
"memory": "1024",
"containerDefinitions": [{
"name": "api",
"image": "123456.dkr.ecr.eu-west-1.amazonaws.com/api:v1",
"portMappings": [{"containerPort": 8080}],
"logConfiguration": {
"logDriver": "awslogs",
"options": {
"awslogs-group": "/ecs/my-api",
"awslogs-region": "eu-west-1"
}
}
}]
}Strengths: Deep AWS integration, long-running tasks, ECS + EKS support, Fargate Spot for cost savings Weaknesses: Slower cold starts (~1-3s), more configuration needed, no scale-to-zero without extra tooling
Knative (Self-Hosted)
apiVersion: serving.knative.dev/v1
kind: Service
metadata:
name: my-api
spec:
template:
metadata:
annotations:
autoscaling.knative.dev/minScale: "0"
autoscaling.knative.dev/maxScale: "100"
spec:
containers:
- image: registry.company.com/api:v1
resources:
limits:
cpu: "1"
memory: 512Mi
ports:
- containerPort: 8080Strengths: Cloud-agnostic, runs on any Kubernetes, full control, no vendor lock-in Weaknesses: You manage Kubernetes, operational overhead, slower iteration than managed services
Comparison Matrix
Cloud Run Fargate Knative
Scale to zero Yes No (native) Yes
Cold start ~200ms ~1-3s ~500ms-2s
Max timeout 60min Unlimited Configurable
GPU support Yes No Yes (with K8s)
Pricing model Per-request Per-second Your K8s cost
Vendor lock-in GCP AWS None
Min config 1 command ~50 lines K8s + Knative
Observability Built-in CloudWatch BYO (Prometheus)When to Use What
"I want the simplest possible deployment"
→ Cloud Run
"I'm already on AWS and need long-running containers"
→ Fargate
"I need cloud portability and already run Kubernetes"
→ Knative
"I need scale-to-zero AND I'm on AWS"
→ Cloud Run (yes, consider multi-cloud) or Karpenter + KEDA
"I have bursty traffic with long idle periods"
→ Cloud Run (scale-to-zero saves money)Cost Comparison (1M Requests/Month)
Assumptions: 200ms avg response, 512MB memory, 1 vCPU
Google Cloud Run:
Compute: 1M × 0.2s × $0.00002400/vCPU-s = $4.80
Memory: 1M × 0.2s × $0.00000250/GiB-s = $0.25
Requests: 1M × $0.40/million = $0.40
Total: ~$5.45/month
AWS Fargate (always-on, 2 tasks):
vCPU: 2 × $0.04048/hr × 730h = $59.10
Memory: 2 × 1GB × $0.004445/hr × 730h = $6.49
Total: ~$65.59/month
Knative (on existing K8s):
Incremental cost on existing cluster = ~$0 (idle)
(But you're paying for the cluster anyway)Cloud Run wins dramatically for bursty workloads. Fargate wins for steady-state. Knative is “free” if you already have Kubernetes capacity.
My Recommendation
For new projects without cloud commitment: start with Cloud Run. The developer experience is unmatched, and you can always migrate to Kubernetes later (the container is portable).
For existing Kubernetes users wanting serverless-like scaling: add Knative to your cluster. I deploy it with Ansible (patterns at Ansible Pilot) and detail the Kubernetes integration at Kubernetes Recipes.
For Terraform-managed infrastructure across clouds, including serverless container services, see Terraform Pilot.
The serverless container space is mature. The question isn’t whether to use it — it’s which flavor fits your architecture.
