- Production traffic mix can shift between features, making independent resource allocation useful. Autoscaled containers or VMs can add capacity; the presenter also describes Amazon deployments using API Gateway, Lambda, and DynamoDB.
8:55 - Microservices do not require serverless infrastructure, and monoliths can run on serverless infrastructure too.
8:55
Estimated compute for 15,000 requests/second
The presenter multiplies measured application CPU time per request by a target of 15,000 public requests/second, near the monolith's winning throughput. These are extrapolations, not serverless benchmark results. 9:17 9:41
| Configuration | Measured CPU/request | Estimated continuously busy cores | Provisioned cores at 80% utilization | Source |
|---|---|---|---|---|
| Monolith | 0.47 ms | About 7 | About 9 | 9:17 9:41 |
| Rebalanced gRPC | 0.91 ms | About 14 | About 17 | 8:21 9:17 9:41 |
| Rebalanced HTTP | 1.27 ms | About 19 | About 24 | 8:21 9:17 9:41 |
- Calculation: busy cores ≈ requests/second × CPU milliseconds/request ÷ 1,000. Dividing by 0.8 estimates provisioned capacity with 20% utilization headroom.
9:179:41 - The estimates assume similar CPU cost at the target load and effective load distribution. More compute can handle extra work; serverless does not automatically remove that work.
9:41 - Scaling down during low traffic could change total cost. Actual platform tests are needed to determine whether it is cheaper; this video does not establish serverless pricing or capacity.
10:05