The meter runs on cardinality

Reconstructs the telemetry cost-containment funnel from the incident record at GitLab and Datadog, the pricing sheets of three clouds, Prometheus and OpenTelemetry source debates including a rejected pull request, and first-person accounts from Zendesk, eBay, Netflix, Zerodha, Wix and Roblox. A reader leaves able to name the billing atom for each signal, choose between rejection and eviction as a containment philosophy on purpose, and price the funnel before finance does.

What you get out of it

• The billing atom for metrics is the series (one name-and-tag-value combination), on the invoice and on the heap alike: Datadog prices it at $5 per 100 per month, AWS prices the metric-month, and GitLab's ingesters OOM on in-memory series counts.
• Containment splits into two philosophies, rejection (Prometheus fails the whole scrape, Loki 429s the tenant) and eviction (OTel folds excess into an overflow series); the OTel collector rejected a hard cardinality cap that had five approvals because stacking both adds a failure point.
• Every mature pipeline is a three-to-four-order-of-magnitude deletion funnel with named keep-rates per stage; Zendesk's steady state indexes a tenth of what it ingests.
• Cardinality bursts come from churn, not steady state: GitLab's Mimir outage was manufactured by routine pod rotation, and Wix reports 150M new series per day against 50M active.
• Self-hosting moves the meter rather than removing it: GitLab found inter-zone replication traffic, not storage, the dominant surprise cost of running its own metrics platform.

Scope

Why this, now. Observability spend has become a board-level line item while the 2025-2026 incident record (GitLab's Mimir cardinality outages, the OTel collector's rejected cardinality cap) shows the containment mechanisms themselves are still being argued out in public.

What it does not cover. Alert and dashboard design, compliance-driven retention for security logs, continuous profiling economics, and vendor feature comparisons; also, deliberately, every source the build environment's network allowlist could not fetch, including the widely reported but here-unverifiable $65M Datadog customer story.