Serverless vs Containers Decision

The Cost Crossover

Cost is usually the first question. Here is how the math works at different scales for a typical API workload (256MB Lambda function, 200ms average duration, compared against a t3.medium reserved instance):

The crossover sits around 20-30 million requests per month for simple workloads. But these numbers shift dramatically based on three factors: function memory allocation (512MB doubles the cost), execution duration (a 1-second function costs 5x more than a 200ms one), and whether you compare against on-demand, reserved, or spot instance pricing. Always run the math with your actual workload profile, not generic benchmarks.

When Serverless Wins

Serverless shines for event-driven workloads with unpredictable traffic patterns. Image processing triggered by S3 uploads. Webhook handlers that spike during business hours and go silent overnight. Cron jobs that run for 30 seconds once an hour. Anything where you would otherwise pay for idle compute.

It also wins for small teams. A startup with three engineers should not be managing Kubernetes clusters. Lambda with API Gateway gets you from zero to production API in an afternoon. The operational savings are worth more than the per-request cost premium.

When Containers Win

Containers win when you need predictable latency, long-running processes, or fine-grained control over the networking stack. WebSocket servers, gRPC services, ML model serving, and any workload that benefits from connection pooling or in-memory caching are better on containers.

ECS Fargate splits the difference: container packaging without managing EC2 instances. You pay a 20-30% premium over self-managed ECS but eliminate node management. For many teams this is the right middle ground.

The Hybrid Approach

Most production systems at scale use both. Shopify runs its core commerce engine on containers but uses serverless for event processing and background jobs. This is not indecision. It is matching the execution model to the workload characteristics.

The practical pattern: run your hot-path synchronous services on containers (EKS, GKE, or ECS) and push asynchronous event processing to Lambda or Cloud Functions. Use SQS or EventBridge as the bridge between the two worlds.

Debugging and Observability

Serverless functions are harder to debug locally. Tools like SAM Local and the Serverless Framework help but do not fully replicate the cloud environment. Distributed tracing with X-Ray or Datadog is essential because you lose the ability to SSH into a box and tail logs.

Containers give you familiar debugging primitives. You can exec into a running container, attach a debugger, capture heap dumps. When something breaks at 3 AM, these capabilities matter more than any architecture diagram.