Cloud monitoring cost for 10 hosts

A 10-host deployment is the canonical small-startup or small-engineering-team scale. Typical shapes include a 5-engineer Series A startup running Kubernetes on EKS or GKE, a side-project launch with modest production traffic, an internal tooling cluster for a 50-engineer engineering organisation, or the staging environment for a larger production deployment. The common feature is that observability is needed, the infrastructure footprint is small, and the team does not have a dedicated platform engineer to manage observability infrastructure.

At this scale, the optimisation question is not how to cut costs (the bills are already small or zero) but how to avoid premature spending on tooling that does not serve the team's current needs. The temptation is to adopt the same observability stack the team used at a previous larger company, which often means Datadog at $250 per month or more for a setup that Grafana Cloud or New Relic would handle for free. The discipline at 10 hosts is to use free tiers until they break, not to pay for capability the team does not need yet.

The free-tier mechanics differ meaningfully between vendors. Grafana Cloud's free tier is structured around capacity (10K active series, 50 GB of each telemetry type, 3 users) and is well-suited to disciplined Kubernetes deployments where label cardinality is controlled. New Relic's free tier is structured around volume (100 GB of unified ingest covering all telemetry) and is well-suited to APM-heavy workloads where the per-host telemetry volume is modest. Datadog's free tier is structured around host count (5 hosts of infrastructure monitoring only) and is intentionally narrow to encourage conversion to paid plans as the team grows.

For a 10-host team that is genuinely uncertain which vendor will fit best at scale, the right move is to deploy free-tier instrumentation across two vendors in parallel for a quarter, evaluate operational fit and team comfort, then standardise on one. Both Grafana Cloud and New Relic free tiers are generous enough to support this evaluation without engineering compromise. Datadog's free tier is too narrow to serve as a parallel evaluation target.

The single most important cost-management practice at 10 hosts is to stay inside free-tier limits as long as operationally feasible. This means understanding the meter for each vendor and instrumenting accordingly. On Grafana Cloud, the active-series meter dominates. A team that adds Kubernetes pod_name as a metric label can cross the 10K free-tier limit overnight and start paying $8 per 1,000 series above that. The fix is to drop high-cardinality labels at the Grafana Agent before they reach the cloud, using Prometheus relabel_configs.

On New Relic, the unified ingest meter dominates. A team that turns on DEBUG-level logging or adds chatty Custom Events API calls can cross the 100 GB free-tier limit from a single misconfiguration. The fix is to filter logs at the source (drop DEBUG and INFO at the application or log shipper before they enter New Relic) and to sample Custom Events sensibly. A 10-host team with disciplined log volume rarely crosses 100 GB per month organically.

For self-hosted Prometheus at this scale, the lever is operational discipline rather than pricing optimisation. A single Prometheus VM at 10 hosts works fine until something fails. The team needs basic backups (weekly snapshot of the Prometheus data directory to S3), an upgrade plan (test new Prometheus releases monthly in a staging environment before promoting to production), and a documented runbook for the most common failure modes (disk full, scrape target unreachable, alert manager misconfigured). With these in place, self-hosted at 10 hosts is genuinely cheap and operationally manageable.