Five years ago this was a heated debate. Today, the decision framework is clearer and the choice often comes down to workload characteristics rather than ideology.
When serverless wins
Lambda and similar serverless platforms are unbeatable for: event-driven workloads with unpredictable traffic (webhooks, async processors), scheduled tasks, functions that run seconds to minutes per invocation, and teams without platform engineering capacity. You pay per invocation, not per idle capacity. For spiky or low-volume workloads, serverless is dramatically cheaper.
The cold start problem has been significantly mitigated. Lambda SnapStart for Java, provisioned concurrency for latency-sensitive functions, and the general improvement in AWS infrastructure have made cold starts a manageable concern rather than a deal-breaker.
When containers win
Long-running connections (WebSockets, SSE, gRPC streaming), workloads that exceed Lambda's 15-minute timeout, applications requiring specific runtime environments or custom binaries, and high-throughput APIs where per-request pricing exceeds flat container pricing — all belong on containers. Container-based architectures also offer better local development parity.
The hybrid approach
Most production architectures are hybrid: a containerised API for synchronous requests, Lambda for async event processing, and Lambda@Edge or Cloudflare Workers for latency-sensitive edge logic. The platforms are complementary, not competing.
Cost comparison at scale
At low traffic: Lambda is almost always cheaper (pay per request, no idle cost). At medium traffic: roughly equivalent — the crossover point is typically around 5–10 million requests per month. At high traffic: containers win — a flat EC2/Fargate cost is much lower than per-request Lambda pricing at millions of daily requests.
Developer experience
Containers win on developer experience. Local development with Docker Compose mirrors production. Debugging is straightforward. Logs are synchronous and structured. Lambda's local development tooling (SAM, LocalStack) has improved but still introduces friction that containers don't have.