Prometheus vs Datadog: how to choose
This is not really a comparison of two monitoring tools; it is a choice between operating your observability stack and buying it. Prometheus is free to license and costs engineering time, storage, and on-call attention. Datadog costs money and removes most of that operational work.
The decision usually turns on two numbers: what a competent engineer's time is worth to you, and how your host and metric counts will grow. Datadog bills per host with additional charges for custom metrics, and the custom metric line is the one that surprises people — a high-cardinality label on a popular metric can multiply the bill without anyone shipping a new service.
Prometheus has no per-metric billing, but it is a single-node system by design. Long retention and high availability mean adding Thanos, Mimir, or Cortex, and at that point you are operating a distributed storage system rather than a scrape target.
Frequently asked questions
At what size does self-hosting Prometheus stop making sense?
It is less about size than about who is available. A single Prometheus handles a surprisingly large estate comfortably. The inflection point is when you need long retention, high availability, and multi-cluster aggregation at once, because that means Thanos or Mimir — a real distributed system with its own failure modes and an owner. If nobody has capacity to own that, buying is the honest answer.
What actually drives the Datadog bill?
Hosts, then custom metrics, then logs. Custom metrics are billed per unique combination of metric name and tag values, so one high-cardinality tag — a user identifier, a request identifier, a raw URL path — can multiply a single metric into hundreds of thousands of billable series. Most Datadog bill shocks are a cardinality accident rather than growth.
Can I use both?
Yes, and it is a common arrangement. Prometheus handles high-cardinality infrastructure metrics and short-retention debugging where volume would be expensive, while Datadog carries business metrics, APM, and the dashboards leadership looks at. Datadog can scrape Prometheus endpoints directly, so the instrumentation does not need to be duplicated.
Does OpenTelemetry change this decision?
It changes the cost of reversing it. Instrumenting with OpenTelemetry rather than a vendor SDK means the exporter is a configuration change, so moving between backends does not mean reinstrumenting every service. It does not reduce the operational burden of self-hosting or the licence cost of a vendor, but it removes most of the lock-in from the choice.
What is the hidden cost of self-hosting?
On-call for the monitoring system itself, which is the component you most need working during an incident. Storage growth is predictable; the unplanned work is cardinality explosions filling the disk, alerting rules that silently stopped evaluating, and upgrades to a stateful component you cannot take offline. Budget for the ownership, not the servers.
Need this managed for you, not just automated?
We're also a hands-on DevOps consultancy — Kubernetes, CI/CD, and cloud infrastructure.
