AWS Lambda on-demand versus provisioned concurrency for low-traffic HTTP endpoints

Goal

Select the billing model for an idle-sensitive AWS Lambda HTTP endpoint that receives only a few requests per day. The goal is to minimize monthly cost while keeping cold-start latency within an acceptable bound.

Constraint

On-demand invocation billing charges only for requests and duration, with no charge while idle, but the first request after inactivity can initialize slowly. Provisioned concurrency keeps execution environments warmed and reduces that latency, yet charges an hourly rate per environment even when there are no invocations. At low traffic, that fixed cost floor can exceed on-demand per-request charges unless the latency SLA justifies it.

Uncertainty

The break-even point depends on memory configuration, average duration, monthly invocation count, and tolerated cold-start latency. Regional pricing and free-tier limits must be verified before budgeting.

Which model is more cost-effective for this workload, and which metrics should be collected to confirm the break-even point?