Your point about translating all cost drivers into a single unit is crucial. We found that even "cost per container-hour" can be misleading if you don't account for the base platform fee. Some vendors have a high fixed cost for the control plane, which makes the per-unit cost look good only at massive scale.
Regarding container churn, we measured it in creations per minute during our CI/CD runs. The per-container-hour model did indeed create a financial disincentive for rapid, ephemeral testing workloads. We had to factor in the cost of a parallel, "cheaper" monitoring stack for our pre-production pipelines, which added operational complexity.
The rate of change for your peak is a sharp addition. If your infrastructure is growing 50% year over year, a three-year contract locks you into either massive over-provisioning at the start or painful renegotiations later. We started building a small buffer into our model, then treating any consumption beyond that as a separate, pre-negotiated burst tier.
That blended rate issue for Datadog is something I'm trying to understand. You mentioned turning on security modules increases APM and log volume automatically.
How do you even forecast that? Are you supposed to run a scaled-down proof of concept with all modules on, just to measure the side effects on your other telemetry before signing a contract? Or is this just a known cost of entry that gets factored in later?
Still learning.