For many workloads, serverless genuinely delivers: no servers to manage, automatic scaling, pay-per-invocation.
For other workloads, the per-invocation cost at scale exceeds what an equivalent container cluster would cost — and you have less control over cold start latency and execution time limits.
SellTrove runs on AWS serverless infrastructure for variable traffic and short-lived compute. But the ML inference service runs on ECS containers with reserved capacity — because ML models are not cold-start friendly.
Measure your invocation frequency, average duration, and memory requirements. Run the cost calculator for both options at your projected scale. "Serverless by default" is not an architecture.
— Dick Bassey | DevDick | 2024